10 interesting stories served every morning and every evening.
Due to aggressive automated scraping of FederalRegister.gov and eCFR.gov, programmatic access to these sites is limited to access to our extensive developer APIs.
If you are human user receiving this message, we can add your IP address to a set of IPs that can access FederalRegister.gov & eCFR.gov; complete the CAPTCHA (bot test) below and click “Request Access”. This process will be necessary for each IP address you wish to access the site from, requests are valid for approximately one quarter (three months) after which the process may need to be repeated.
An official website of the United States government.
If you want to request a wider IP range, first request access for your current IP, and then use the “Site Feedback” button found in the lower left-hand side to make the request.
...
Read the original on www.federalregister.gov »
See the canonical version of this blog post at the Microsoft Open Source Blog!
Ten years ago, Microsoft released the source for MS-DOS 1.25 and 2.0 to the Computer History Museum, and then later republished them for reference purposes. This code holds an important place in history and is a fascinating read of an operating system that was written entirely in 8086 assembly code nearly 45 years ago.
Today, in partnership with IBM and in the spirit of open innovation, we’re releasing the source code to MS-DOS 4.00 under the MIT license. There’s a somewhat complex and fascinating history behind the 4.0 versions of DOS, as Microsoft partnered with IBM for portions of the code but also created a branch of DOS called Multitasking DOS that did not see a wide release.
A young English researcher named Connor “Starfrost” Hyde recently corresponded with former Microsoft Chief Technical Officer Ray Ozzie about some of the software in his collection. Amongst the floppies, Ray found unreleased beta binaries of DOS 4.0 that he was sent while he was at Lotus. Starfrost reached out to the Microsoft Open Source Programs Office (OSPO) to explore releasing DOS 4 source, as he is working on documenting the relationship between DOS 4, MT-DOS, and what would eventually become OS/2. Some later versions of these Multitasking DOS binaries can be found around the internet, but these new Ozzie beta binaries appear to be much earlier, unreleased, and also include the ibmbio.com source.
Scott Hanselman, with the help of internet archivist and enthusiast Jeff Sponaugle, has imaged these original disks and carefully scanned the original printed documents from this “Ozzie Drop”. Microsoft, along with our friends at IBM, think this is a fascinating piece of operating system history worth sharing.
Jeff Wilcox and OSPO went to the Microsoft Archives, and while they were unable to find the full source code for MT-DOS, they did find MS DOS 4.00, which we’re releasing today, alongside these additional beta binaries, PDFs of the documentation, and disk images. We will continue to explore the archives and may update this release if more is discovered.
Thank you to Ray Ozzie, Starfrost, Jeff Sponaugle, Larry Osterman, our friends at the IBM OSPO, as well as the makers of such digital archeology software including, but not limited to Greaseweazle, Fluxengine, Aaru Data Preservation Suite, and the HxC Floppy Emulator. Above all, thank you to the original authors of this code, some of whom still work at Microsoft and IBM today!
If you’d like to run this software yourself and explore, we have successfully run it directly on an original IBM PC XT, a newer Pentium, and within the open source PCem and 86box emulators.
...
Read the original on www.hanselman.com »
You can install this module as a component from NPM:
npm install –save canvas-confetti
You can then require(‘canvas-confetti’); to use it in your project build. Note: this is a client component, and will not run in Node. You will need to build your project with something like webpack in order to use this.
You can also include this library in your HTML page directly from a CDN:
Note: you should use the latest version at the time that you include your project. You can see all versions on the releases page.
Thank you for joining me in this very important message about motion on your website. See, not everyone likes it, and some actually prefer no motion. They have ways to tell us about it and we should listen. While I don’t want to go as far as tell you not to have confetti on your page just yet, I do want to make it easy for you to respect what your users want. There is a disableForReducedMotion option you can use so that users that have trouble with chaotic animations don’t need to struggle on your website. This is disabled by default, but I am considering changing that in a future major release. If you have strong feelings about this, please let me know. For now, please confetti responsibly.
When installed from npm, this library can be required as a client component in your project build. When using the CDN version, it is exposed as a confetti function on window.
confetti takes a single optional object. When window. Promise is available, it will return a Promise to let you know when it is done. When promises are not available (like in IE), it will return null. You can polyfill promises using any of the popular polyfills. You can also provide a promise implementation to confetti through:
const MyPromise = require(‘some-promise-lib’);
const confetti = require(‘canvas-confetti’);
confetti.Promise = MyPromise;
If you call confetti multiple times before it is done, it will return the same promise every time. Internally, the same canvas element will be reused, continuing the existing animation with the new confetti added. The promise returned by each call to confetti will resolve once all animations are done.
The confetti parameter is a single optional options object, which has the following properties:
* particleCount Integer (default: 50): The number of confetti to launch. More is always fun… but be cool, there’s a lot of math involved.
* angle Number (default: 90): The angle in which to launch the confetti, in degrees. 90 is straight up.
* spread Number (default: 45): How far off center the confetti can go, in degrees. 45 means the confetti will launch at the defined angle plus or minus 22.5 degrees.
* startVelocity Number (default: 45): How fast the confetti will start going, in pixels.
* decay Number (default: 0.9): How quickly the confetti will lose speed. Keep this number between 0 and 1, otherwise the confetti will gain speed. Better yet, just never change it.
* gravity Number (default: 1): How quickly the particles are pulled down. 1 is full gravity, 0.5 is half gravity, etc., but there are no limits. You can even make particles go up if you’d like.
* drift Number (default: 0): How much to the side the confetti will drift. The default is 0, meaning that they will fall straight down. Use a negative number for left and positive number for right.
* flat Boolean (default: false): Optionally turns off the tilt and wobble that three dimensional confetti would have in the real world. Yeah, they look a little sad, but y’all asked for them, so don’t blame me.
* ticks Number (default: 200): How many times the confetti will move. This is abstract… but play with it if the confetti disappear too quickly for you.
* origin Object: Where to start firing confetti from. Feel free to launch off-screen if you’d like.
origin.x Number (default: 0.5): The x position on the page, with 0 being the left edge and 1 being the right edge.
origin.y Number (default: 0.5): The y position on the page, with 0 being the top edge and 1 being the bottom edge.
* origin.x Number (default: 0.5): The x position on the page, with 0 being the left edge and 1 being the right edge.
* origin.y Number (default: 0.5): The y position on the page, with 0 being the top edge and 1 being the bottom edge.
* colors Array<String>: An array of color strings, in the HEX format… you know, like #bada55.
* shapes Array<String|Shape>: An array of shapes for the confetti. There are 3 built-in values of square, circle, and star. The default is to use both squares and circles in an even mix. To use a single shape, you can provide just one shape in the array, such as [‘star’]. You can also change the mix by providing a value such as [‘circle’, ‘circle’, ‘square’] to use two third circles and one third squares. You can also create your own shapes using the confetti.shapeFromPath or confetti.shapeFromText helper methods.
* scalar Number (default: 1): Scale factor for each confetti particle. Use decimals to make the confetti smaller. Go on, try teeny tiny confetti, they are adorable!
* zIndex Integer (default: 100): The confetti should be on top, after all. But if you have a crazy high page, you can set it even higher.
* disableForReducedMotion Boolean (default: false): Disables confetti entirely for users that prefer reduced motion. The confetti() promise will resolve immediately in this case.
This helper method lets you create a custom confetti shape using an SVG Path string. Any valid path should work, though there are a few caveats:
* All paths will be filed. If you were hoping to have a stroke path, that is not implemented.
* Paths are limited to a single color, so keep that in mind.
* All paths need a valid transform matrix. You can pass one in, or you can leave it out and use this helper to calculate the matrix for you. Do note that calculating the matrix is a bit expensive, so it is best to calculate it once for each path in development and cache that value, so that production confetti remain fast. The matrix is deterministic and will always be the same given the same path value.
* For best forward compatibility, it is best to re-generate and re-cache the matrix if you update the canvas-confetti library.
* Support for path-based confetti is limited to browsers which support Path2D, which should really be all major browser at this point.
This method will return a Shape — it’s really just a plain object with some properties, but shhh… we’ll pretend it’s a shape. Pass this Shape object into the shapes array directly.
As an example, here’s how you might do a triangle confetti:
var triangle = confetti.shapeFromPath({ path: ‘M0 10 L5 0 L10 10z’ });
confetti({
shapes: [triangle]
This is the highly anticipated feature to render emoji confetti! Use any standard unicode emoji. Or other text, but… maybe don’t use other text.
While any text should work, there are some caveats:
* For flailing confetti, something that is mostly square works best. That is, a single character, especially an emoji.
* Rather than rendering text every time a confetti is drawn, this helper actually rasterizes the text. Therefore, it does not scale well after it is created. If you plan to use the scalar value to scale your confetti, use the same scalar value here when creating the shape. This will make sure the confetti are not blurry.
The options for this method are:
* options Object:
text String: the text to be rendered as a confetti. If you can’t make up your mind, I suggest “🐈”.
scalar Number, optional, default: 1: a scale value relative to the default size. It matches the scalar value in the confetti options.
color String, optional, default: #000000: the color used to render the text.
fontFamily String, optional, default: native emoji: the font family name to use when rendering the text. The default follows best practices for rendring the native OS emoji of the device, falling back to sans-serif. If using a web font, make sure this font is loaded before rendering your confetti.
* text String: the text to be rendered as a confetti. If you can’t make up your mind, I suggest “🐈”.
* scalar Number, optional, default: 1: a scale value relative to the default size. It matches the scalar value in the confetti options.
* color String, optional, default: #000000: the color used to render the text.
* fontFamily String, optional, default: native emoji: the font family name to use when rendering the text. The default follows best practices for rendring the native OS emoji of the device, falling back to sans-serif. If using a web font, make sure this font is loaded before rendering your confetti.
var scalar = 2;
var pineapple = confetti.shapeFromText({ text: ‘🍍’, scalar });
confetti({
shapes: [pineapple],
scalar
This method creates an instance of the confetti function that uses a custom canvas. This is useful if you want to limit the area on your page in which confetti appear. By default, this method will not modify the canvas in any way (other than drawing to it).
Canvas can be misunderstood a bit though, so let me explain why you might want to let the module modify the canvas just a bit. By default, a canvas is a relatively small image — somewhere around 300x150, depending on the browser. When you resize it using CSS, this sets the display size of the canvas, but not the image being represented on that canvas. Think of it as loading a 300x150 jpeg image in an img tag and then setting the CSS for that tag to 1500x600 — your image will end up stretched and blurry. In the case of a canvas, you need to also set the width and height of the canvas image itself. If you don’t want to do that, you can allow confetti to set it for you.
Note also that you should persist the custom instance and avoid initializing an instance of confetti with the same canvas element more than once.
The following global options are available:
* resize Boolean (default: false): Whether to allow setting the canvas image size, as well as keep it correctly sized if the window changes size (e.g. resizing the window, rotating a mobile device, etc.). By default, the canvas size will not be modified.
* useWorker Boolean (default: false): Whether to use an asynchronous web worker to render the confetti animation, whenever possible. This is turned off by default, meaning that the animation will always execute on the main thread. If turned on and the browser supports it, the animation will execute off of the main thread so that it is not blocking any other work your page needs to do. Using this option will also modify the canvas, but more on that directly below — do read it. If it is not supported by the browser, this value will be ignored.
* disableForReducedMotion Boolean (default: false): Disables confetti entirely for users that prefer reduced motion. When set to true, use of this confetti instance will always respect a user’s request for reduced motion and disable confetti for them.
Important: If you use useWorker: true, I own your canvas now. It’s mine now and I can do whatever I want with it (don’t worry… I’ll just put confetti inside it, I promise). You must not try to use the canvas in any way (other than I guess removing it from the DOM), as it will throw an error. When using workers for rendering, control of the canvas must be transferred to the web worker, preventing any usage of that canvas on the main thread. If you must manipulate the canvas in any way, do not use this option.
var myCanvas = document.createElement(‘canvas’);
document.body.appendChild(myCanvas);
var myConfetti = confetti.create(myCanvas, {
resize: true,
useWorker: true
myConfetti({
particleCount: 100,
spread: 160
// any other options from the global
// confetti function
Stops the animation and clears all confetti, as well as immediately resolves any outstanding promises. In the case of a separate confetti instance created with confetti.create, that instance will have its own reset method.
confetti();
setTimeout(() => {
confetti.reset();
}, 100);
var myCanvas = document.createElement(‘canvas’);
document.body.appendChild(myCanvas);
var myConfetti = confetti.create(myCanvas, { resize: true });
myConfetti();
setTimeout(() => {
myConfetti.reset();
}, 100);
Launch some confetti the default way:
confetti();
confetti({
particleCount: 150
confetti({
spread: 180
Get creative. Launch a small poof of confetti from a random part of the page:
confetti({
...
Read the original on github.com »
TSMC announced its leading-edge 1.6nm-class process technology today, a new A16 manufacturing process that will be the company’s first Angstrom-class production node and promises to outperform its predecessor, N2P, by a significant margin. The technology’s most important innovation will be its backside power delivery network (BSPDN).
Just like TSMC’s 2nm-class nodes (N2, N2P, and N2X), the company’s 1.6nm-class fabrication process will rely on gate-all-around (GAA) nanosheet transistors, but unlike the current and next-generation nodes, this one uses backside power delivery dubbed Super Power Rail. Transistor and BSPDN innovations enable tangible performance and efficiency improvements compared to TSMC’s N2P: the new node promises an up to 10% higher clock rate at the same voltage and a 15%–20% lower power consumption at the same frequency and complexity. In addition, the new technology could enable 7%–10% higher transistor density, depending on the actual design.
The most important innovation of TSMC’s A16 process, which was unveiled at the company’s North American Technology Symposium 2024, is the introduction of the Super Power Rail (SPR), a sophisticated backside power delivery network (BSPDN). This technology is tailored specifically for AI and HPC processors that tend to have both complex signal wiring and dense power delivery networks.
Backside power delivery will be implemented into many upcoming process technologies as it allows for an increase in transistor density and improved power delivery, which affects performance. Meanwhile, there are several ways to implement a BSPDN. TSMC’s Super Power Rail plugs the backside power delivery network to each transistor’s source and drain using a special contact that also reduces resistance to get the maximum performance and power efficiency possible. From a production perspective, this is one of the most complex BSPDN implementations and is more complex than Intel’s Power Via.
The choice of backside power rail implementation is perhaps why TSMC decided not to add this feature to its N2P and N2X process technologies, as it would make using the production nodes considerably more expensive. Meanwhile, by offering a 1.6nm-class node with GAA nanosheet transistors and SPR as well as 2nm-class nodes with GAAFETs only, the company will now have two distinct nodes that will not compete with each other directly but offer distinctive advantages for different customers.
The production timeline for A16 indicates that volume production of A16 will commence in the second half of 2026. Therefore, actual A16-made products will likely debut in 2027. This timeline positions A16 to potentially compete with Intel’s 14A node, which will be the Intel’s most advanced node at the time.
...
Read the original on www.tomshardware.com »
is a Turing equivalent 16-bit computer made entirely from a clock and NAND gates emulated on the web. NAND features its own CPU, machine code language, assembly language, assembler, virtual machine language, virtual machine translator, programming language, compiler, IDE, and user interface. NAND is based on the Jack-VM-Hack platform specified in the Nand to Tetris course and its associated book.
A simple program that inputs some numbers and computes their average, showing off control flow, arithmetic operations, I/O, and dynamic memory allocation.
This program was supplied by the Nand to Tetris software suite.
The game of Pong, showing off the language’s object-oriented model. Use the arrow keys to move the paddle left and right to bounce a ball. Every bounce, the paddle becomes smaller, and the game ends when the ball hits the bottom of the screen.
This program was supplied by the Nand to Tetris software suite.
The game of 2048, showing off recursion and complex application logic. Use the arrow keys to move the numbers around the 4x4 grid. The same numbers combine into their sum when moved into each other. Once the 2048 tile is reached, you win the game, though you can keep playing on until you lose. You lose the game when the board is full and you can’t make any more moves.
A program that deliberately causes a stack overflow via infinite recursion to perform a virtual machine escape. It leverages the fact that there are no runtime checks to prevent a stack overflow. No other modern platform will let you do this :-)
Upon running, the program will constantly print the stack pointer to the screen. Once this displayed value exceeds 2048, the stack will have reached the end of its intended memory space and spill onto the heap memory space, causing the print statement to malfunction in explosive fashion:
Two things of noteworthy interest are worth pointing out.
If you reload the page and run this program on an empty RAM (a RAM full of zeroes), you will notice that the program resets itself halfway through its execution despite not pressing the “Reset” button. Why this happens is simple: the jailbroken runtime executes an instruction that sets the program counter’s value to 0, effectively telling the program to jump to the first instruction and start over.
If you run the GeneticAlgorithm example program and then run this immediately afterwards, the program in its rampage reads old RAM memory that was simply never overwritten.
A program that exploits the fact that the runtime doesn’t prevent stack smashing to call a function that would otherwise be inaccessible. In order to understand how this works, let’s examine this illustration of NAND’s stack frame layout.
taken from the Nand to Tetris book.
If you’re unfamiliar with stack layouts, here’s the main idea behind the exploit. Whenever a function returns, it needs to know where (which machine code instruction memory address) it should go to proceed with execution flow. So, when the function is first called, this memory address, along with some other unimportant data, is temporarily stored on the stack in a memory region referred to as the stack frame as a reference for where to return. The illustration describes the exact position of this return address relative to the function call, a position that can be reverse engineered.
The program enables the user to overwrite a single memory address in the RAM to any value. Putting two and two together, if the user were to overwrite the return address of a stack frame with the address of another function, they essentially gain the ability to execute arbitrary code included in the program.
Indeed, if you enter 267 as the memory location and 1715 as the value to overwrite, two numbers reverse engineered by manually inspecting the stack memory space and the assembler, you’ll see this idea in working action.
This isn’t a vulnerability unique to NAND. It exists in C as well! How cool!
Believe it or not, out of the many, many different components of NAND, this single-handedly took the longest to develop!
This program is a creature simulation that utilizes simple machine learning. It follows the artificial intelligence coded series (parts one and two) from Code Bullet. Make sure to check out his channel, he makes some really cool stuff!
Every dot has its own “brain” of acceleration vectors, and they evolve to reach a goal through natural selection. Every generation, dots that “die” closer to the goal are more likely to be selected as the parents for the next generation. Reproduction inherently causes some of the brain to mutate, wholly effectively simulating natural evolution.
Nevertheless, there is much to be desired. Due to performance, the only factor dots use to evolve is their closeness to the goal upon death, endowing the natural selection algorithm with low entropy. Due to memory usage, there are smaller than satisfactory limits on the number of dots and the sizes of their brains. Lastly, due to technical complexity, re-placing obstacles during the simulation does not guarantee that the dots will have large enough brains to reach the goal. Brain sizes are only determined at the beginning of the program.
I’ve utilized a myriad of optimization techniques to snake around the following hardware restrictions and make this possible:
* NAND has a limited ROM memory space, meaning the program won’t compile if there’s too much code. In fact, the final version of this program uses 99.2% of the instruction memory space.
* NAND has a limited RAM memory space, meaning the program has to be careful to optimize heap memory usage. In fact, the reason why the screen fills with static between generations is to use the screen memory space as temporary swap memory for the next generation — the RAM is already completely full!
* NAND has no floating point type (decimal numbers) and can only represent the integers between -32768 and 32767, making calculating fitness less precise and more challenging to implement. Integer overflows must also be accounted for.
To avoid beating around the bush, I’ve stuck to documenting these techniques and additional insights in this program’s codebase for those interested.
Before we start, the most important detail to remember about writing programs in Jack is that there is no operator priority; this is probably why your program isn’t working.
For example, you should change:
4 * 2 + 3 to (4 * 2) + 3
if (~x & y) to if ((~x) & y)
but you can keep if (y & ~x) the same as there is no operator ambiguity.
Without parenthesis, the evaluation value of an ambiguous expression is undefined.
NAND boasts its own complete tech stack. As a consequence, NAND can only be programmed in Jack, its weakly typed object-oriented programming language. In layman’s terms, Jack is C with Java’s syntax.
Let’s take the approach of example-based learning and dive right in.
* This program prompts the user to enter a phrase
* and an energy level. Program output:
* Whats on your mind? Superman
* Whats your energy level? 3
* Superman!
* Superman!
* Superman!
class Main {
function void main() {
var String s;
var int energy, i;
let s = Keyboard.readLine(“Whats on your mind? “);
let energy = Keyboard.readInt(“Whats your energy level? “);
let i = 0;
let s = s.appendChar(33); // Appends the character ‘!’
while (i < energy) {
do Output.printString(s);
do Output.println();
let i = i + 1;
taken from the Nand to Tetris lecture slides.
If you’ve already had some experience with programming, this should look very familiar; it is clear that Jack was heavily inspired by Java. Main.main is the entry point to the program. The code segment demonstrates basic usage of variables as well as the while loop for control flow.
Additionally, it uses Keyboard.readLine and Keyboard.readInt to read input from the user, and Output.printString and Output.println to print output to the screen — all of which are defined in detail in the Jack OS Reference. By default, the Jack OS is bundled with your program during compilation to enable interfacing with strings, memory, hardware, and more.
Every programming language has a fixed set of primitive data types. Jack supports three: int, char, and boolean. You can extend this basic repertoire with your own abstract data types as needed. Prior knowledge about object-oriented programming directly carries over to this section.
/** Represents a point in 2D plane. */
class Point {
// The coordinates of the current point instance:
field int x, y;
// The number of point objects constructed so far:
static int pointCount;
/** Constructs a point and initializes
it with the given coordinates */
constructor Point new(int ax, int ay) {
let x = ax;
let y = ay;
let pointCount = pointCount + 1;
return this;
/** Returns the x coordinate of the current point instance */
method int getx() { return x; }
/** Returns the y coordinate of the current point instance */
method int gety() { return y; }
/** Returns the number of Points constructed so far */
function int getPointCount() {
return pointCount;
/** Returns a point which is this
point plus the other point */
method Point plus(Point other) {
return Point.new(x + other.getx(),
y + other.gety());
/** Returns the Euclidean distance between the
current point instance and the other point */
method int distance(Point other) {
var int dx, dy;
let dx = x - other.getx();
let dy = y - other.gety();
return Math.sqrt((dx * dx) + (dy * dy));
/** Prints the current point instance, as “(x, y)” */
method void print() {
var String tmp;
let tmp = “(”;
do Output.printString(tmp);
do tmp.dispose();
do Output.printInt(x);
let tmp = ”, ”;
do Output.printString(tmp);
do tmp.dispose();
do Output.printInt(y);
...
Read the original on github.com »
A minimal GPU implementation in Verilog optimized for learning about how GPUs work from the ground up.
If you want to learn how a CPU works all the way from architecture to control signals, there are many resources online to help you.
GPUs are not the same.
Because the GPU market is so competitive, low-level technical details for all modern architectures remain proprietary.
While there are lots of resources to learn about GPU programming, there’s almost nothing available to learn about how GPU’s work at a hardware level.
The best option is to go through open-source GPU implementations like Miaow and VeriGPU and try to figure out what’s going on. This is challenging since these projects aim at being feature complete and functional, so they’re quite complex.
This is why I built tiny-gpu!
With this motivation in mind, we can simplify GPUs by cutting out the majority of complexity involved with building a production-grade graphics card, and focus on the core elements that are critical to all of these modern hardwareaccelerators.
This project is primarily focused on exploring:
Architecture - What does the architecture of a GPU look like? What are the most important elements?
Parallelization - How is the SIMD progamming model implemented in hardware?
Memory - How does a GPU work around the constraints of limited memory bandwidth?
After understanding the fundamentals laid out in this project, you can checkout the advanced functionality section to understand some of the most important optimizations made in production grade GPUs (that are more challenging to implement) which improve performance.
tiny-gpu is built to execute a single kernel at a time.
In order to launch a kernel, we need to do the following:
Load data memory with the necessary data
Specify the number of threads to launch in the device control register
Launch the kernel by setting the start signal to high.
The GPU itself consists of the following units:
The device control register usually stores metadata specifying how kernels should be executed on the GPU.
In this case, the device control register just stores the thread_count - the total number of threads to launch for the active kernel.
Once a kernel is launched, the dispatcher is the unit that actually manages the distribution of threads to different compute cores.
The dispatcher organizes threads into groups that can be executed in parallel on a single core called blocks and sends these blocks off to be processed by available cores.
Once all blocks have been processed, the dispatcher reports back that the kernel execution is done.
The GPU is built to interface with an external global memory. Here, data memory and program memory are separated out for simplicity.
tiny-gpu data memory has the following specifications:
* 8 bit data (stores values of <256 for each row)
tiny-gpu program memory has the following specifications:
* 16 bit data (each instruction is 16 bits as specified by the ISA)
Global memory has fixed read/write bandwidth, but there may be far more incoming requests across all cores to access data from memory than the external memory is actually able to handle.
The memory controllers keep track of all the outgoing requests to memory from the compute cores, throttle requests based on actual external memory bandwidth, and relay responses from external memory back to the proper resources.
Each memory controller has a fixed number of channels based on the bandwidth of global memory.
The same data is often requested from global memory by multiple cores. Constantly access global memory repeatedly is expensive, and since the data has already been fetched once, it would be more efficient to store it on device in SRAM to be retrieved much quicker on later requests.
This is exactly what the cache is used for. Data retrieved from external memory is stored in cache and can be retrieved from there on later requests, freeing up memory bandwidth to be used for new data.
Each core has a number of compute resources, often built around a certain number of threads it can support. In order to maximize parallelization, these resources need to be managed optimally to maximize resource utilization.
In this simplified GPU, each core processed one block at a time, and for each thread in a block, the core has a dedicated ALU, LSU, PC, and register file. Managing the execution of thread instructions on these resources is one of the most challening problems in GPUs.
Each core has a single scheduler that manages the execution of threads.
The tiny-gpu scheduler executes instructions for a single block to completion before picking up a new block, and it executes instructions for all threads in-sync and sequentially.
In more advanced schedulers, techniques like pipelining are used to stream the execution of multiple instructions subsequent instructions to maximize resource utilization before previous instructions are fully complete. Additionally, warp scheduling can be use to execute multiple batches of threads within a block in parallel.
The main constraint the scheduler has to work around is the latency associated with loading & storing data from global memory. While most instructions can be executed synchronously, these load-store operations are asynchronous, meaning the rest of the instruction execution has to be built around these long wait times.
Asynchronously fetches the instruction at the current program counter from program memory (most should actually be fetching from cache after a single block is executed).
Decodes the fetched instruction into control signals for thread execution.
Each thread has it’s own dedicated set of register files. The register files hold the data that each thread is performing computations on, which enables the same-instruction multiple-data (SIMD) pattern.
Importantly, each register file contains a few read-only registers holding data about the current block & thread being executed locally, enabling kernels to be executed with different data based on the local thread id.
Dedicated arithmetic-logic unit for each thread to perform computations. Handles the ADD, SUB, MUL, DIV arithmetic instructions.
Also handles the CMP comparison instruction which actually outputs whether the result of the difference between two registers is negative, zero or positive - and stores the result in the NZP register in the PC unit.
Dedicated load-store unit for each thread to access global data memory.
Handles the LDR & STR instructions - and handles async wait times for memory requests to be processed and relayed by the memory controller.
Dedicated program-counter for each unit to determine the next instructions to execute on each thread.
By default, the PC increments by 1 after every instruction.
With the BRnzp instruction, the NZP register checks to see if the NZP register (set by a previous CMP instruction) matches some case - and if it does, it will branch to a specific line of program memory. This is how loops and conditionals are implemented.
Since threads are processed in parallel, tiny-gpu assumes that all threads “converge” to the same program counter after each instruction - which is a naive assumption for the sake of simplicity.
In real GPUs, individual threads can branch to different PCs, causing branch divergence where a group of threads threads initially being processed together has to split out into separate execution.
tiny-gpu implements a simple 11 instruction ISA built to enable simple kernels for proof-of-concept like matrix addition & matrix multiplication (implementation further down on this page).
For these purposes, it supports the following instructions:
* BRnzp - Branch instruction to jump to another line of program memory if the NZP register matches the nzp condition in the instruction.
* CMP - Compare the value of two registers and store the result in the NZP register to use for a later BRnzp instruction.
* RET - Signal that the current thread has reached the end of execution.
Each register is specified by 4 bits, meaning that there are 16 total registers. The first 13 register R0 - R12 are free registers that support read/write. The last 3 registers are special read-only registers used to supply the %blockIdx, %blockDim, and %threadIdx critical to SIMD.
Each core follows the following control flow going through different stages to execute each instruction:
FETCH - Fetch the next instruction at current program counter from program memory.
REQUEST - Request data from global memory if necessary (if LDR or STR instruction).
WAIT - Wait for data from global memory if applicable.
The control flow is laid out like this for the sake of simplicity and understandability.
In practice, several of these steps could be compressed to be optimize processing times, and the GPU could also use pipelining to stream and coordinate the execution of many instructions on a cores resources without waiting for previous instructions to finish.
Each thread within each core follows the above execution path to perform computations on the data in it’s dedicated register file.
This resembles a standard CPU diagram, and is quite similar in functionality as well. The main difference is that the %blockIdx, %blockDim, and %threadIdx values lie in the read-only registers for each thread, enabling SIMD functionality.
I wrote a matrix addition and matrix multiplication kernel using my ISA as a proof of concept to demonstrate SIMD programming and execution with my GPU. The test files in this repository are capable of fully simulating the execution of these kernels on the GPU, producing data memory states and a complete execution trace.
This matrix addition kernel adds two 1 x 8 matrices by performing 8 element wise additions in separate threads.
This demonstration makes use of the %blockIdx, %blockDim, and %threadIdx registers to show SIMD programming on this GPU. It also uses the LDR and STR instructions which require async memory management.
.threads 8
.data 0 1 2 3 4 5 6 7 ; matrix A (1 x 8)
.data 0 1 2 3 4 5 6 7 ; matrix B (1 x 8)
MUL R0, %blockIdx, %blockDim
ADD R0, R0, %threadIdx ; i = blockIdx * blockDim + threadIdx
CONST R1, #0 ; baseA (matrix A base address)
CONST R2, #8 ; baseB (matrix B base address)
CONST R3, #16 ; baseC (matrix C base address)
ADD R4, R1, R0 ; addr(A[i]) = baseA + i
LDR R4, R4 ; load A[i] from global memory
ADD R5, R2, R0 ; addr(B[i]) = baseB + i
LDR R5, R5 ; load B[i] from global memory
ADD R6, R4, R5 ; C[i] = A[i] + B[i]
ADD R7, R3, R0 ; addr(C[i]) = baseC + i
STR R7, R6 ; store C[i] in global memory
RET ; end of kernel
The matrix multiplication kernel multiplies two 2x2 matrices. It performs element wise calculation of the dot product of the relevant row and column and uses the CMP and BRnzp instructions to demonstrate branching within the threads (notably, all branches converge so this kernel works on the current tiny-gpu implementation).
.threads 4
.data 1 2 3 4 ; matrix A (2 x 2)
.data 1 2 3 4 ; matrix B (2 x 2)
MUL R0, %blockIdx, %blockDim
ADD R0, R0, %threadIdx ; i = blockIdx * blockDim + threadIdx
CONST R1, #1 ; increment
CONST R2, #2 ; N (matrix inner dimension)
CONST R3, #0 ; baseA (matrix A base address)
CONST R4, #4 ; baseB (matrix B base address)
CONST R5, #8 ; baseC (matrix C base address)
DIV R6, R0, R2 ; row = i // N
MUL R7, R6, R2
SUB R7, R0, R7 ; col = i % N
...
Read the original on github.com »
Tribler is a Bittorrent-compatible alternative to Youtube. It is designed to protect your privacy, build a web-of-trust, be attack-resilient, and reward content creators directly. We are building a micro-economy without banks, without advertisers, and without any government. Together with Harvard University, the Tribler team deployed one of the first fully distributed ledgers in August 2007, see BBC News coverge and a New Scientist article. In coming years we will further expand our micro-economy based on bandwidth tokens. We aim to become the key place where audiences find their torrents, creative talents get discovered, and artists get financial rewards from their fans. Tribler is the place where 100 percent of the money goes to artists and the people that run the infrastructure.
Over 2 million people have used Tribler over the years. The Tribler project was started in 2005 at Delft University of Technology and over 100+ developers contributed code to it. We are continuously improving it and further expanding the scientific developers team.
Technical foundations of Tribler are the Bittorrent protocol, an overlay for P2P communication across NAT/firewalls, gradual building of trust in public keys with Bittorrent seeding, and our token economy with incentives for Tor-like relaying and hidden seeding. For 12 years we have been building a very robust self-organising Peer-to-Peer system. Today Tribler is robust: “the only way to take Tribler down is to take The Internet down” (but a single software bug could end everything).
This wiki page contains our main technical documentation, highlights:
* Trustchain: our 10.000 transactions per second ledger
Open projects for new TUDelft master thesis students: Tor-like streaming, self-sovereign identity and authentication on Android, relevance ranking of search results (+swarm popularity), perfect metadata through distributed crowdsourcing, self-reinforcing trust, and perfect network connectivity using NAT/Firewall traversal. Speculative projects with long-term focus: prediction market for climate change. A market designed against frontrunners and high-frequency trading abusers in general.
Social media today is obsessed with profit, filled with advertisements, overflowing with falsehoods, and infested with fake news. We’re trying to fix these hard problems in a unique way: by building trust. Our audacious ambition is a clean-slate re-creation of The Internet itself with foundations of trust. Craiglist and eBay showed us in 1995 that trustworthy trade was possible online. Uber, Etsy, and AirBnB show that entire industries can be disrupted by a single platform with a natural monopoly.
For the past 18 years we have build and deployed platforms to create trust. Before Wikipedia and Youtube existed we studied the mechanisms behind trust and user-generated content on a small scale. Several years before Wikipedia emerged we deployed a music encyclopedia with unconstrained write access, it never became popular because we focused too much on software, instead community growth.
Today we keep a narrow focus and continuously expand Tribler with trustworthy decentralized technology. We launched sub-second keyword search for Bittorrent swarms without any server back in 2010 (see our old Google Tech Talk on this topic). One of our operational trust browsing prototypes:
* Our work from 2004, 2-year in-depth measurement and analysis of Bittorrent (.pdf 25 pages), largest measurement to date. Covers eight months of the BitTorrent/Suprnova.org file sharing ecosystem. In particular, we show measurement results of the popularity and the availability of BitTorrent, of its download performance, of the content lifetime, and of the structure of the community responsible for verifying uploaded content.
Tribler supports torrent search without websites, anonymous downloading, torrent streaming, channels of torrents, and sharing content for tokens. Overview of Tribler (.html 5 pages). All Tribler features are implemented in a completely distributed manner, not relying on any centralized component. Still, Tribler manages to remain fully backwards compatible with BitTorrent. The 2006 overview of Tribler (.pdf 6 pages) featuring taste groups, friends, friends-of-friends and faster downloads by donating bandwidth to friends (protocol spec of friend boosting). Note that the 2006-2009 Tribler protocol specification (.pdf 47 pages) is now mostly outdated, as we switched to our new synchronization protocol called Dispersy (see below).
Trust in social media content is essential for a sustainable ecosystem. We introduced channels of Bittorrent swarms in 2009 with the Tribler 4.x release. Each user can vote on channels to increase their visibility and tell everybody the channel owner is not a spammer and not spreading fake items. The reputation of both the voters and channel owner are important.
Tribler protects your privacy by not storing anything on any server. To protect your privacy even more, we have prototyped search algorithms based on homomorphic cryptography. We presented a new algorithm system for privacy-respecting
scalable Gnutella-like search in 2014. Our approach to scalability is a similarity function in the encrypted domain (homomorphic), enabling semantic clustering with privacy.
Back in 2006 we introduced long-lived identities to separate trustworthy peers from freeriders and spammers (PermID). To protect your privacy further we also devised an alternative to onion routing which potentially could have stronger security guarantees (correlation attack). See the details in this thesis on
Multi-core architecture for anonymous Internet streaming
which includes a performance analysis of running code.
We deployed one of the worlds first fully distributed ledgers in August of 2007. For over a decade we meticulously measured, analysed, improved, and enhanced this live system. Today it defines the state-of-the-art in blockchain research, but in the early days it barely functioned at all. A total of five Ph. D. students of Delft contributed key parts and upgrades.
At launch we called our initiative “bandwidth-as-a-currency”. Today we have specific terminology for what we did: a token economy. We are making Internet bandwidth a tradable commodity without any middleman or need for any centralised governance. Our efforts span over a decade, making us the veterans in the field. Our ledger provides an incentive for Bittorrent seeding and Tor-like relaying. For numerous years the tit-for-tat algorithm provided the only incentive for contributions in Bittorrent. No incentive for seeding existed, except when central servers kept track of your uploads and downloads. We measured closed invite-only communities for numerous years and mathematically showed their rich-get-richer properties. For details see Fast download but eternal seeding: the reward and punishment of sharing ratio enforcement and our measurement paper understanding bandwidth economics and ratio enforcement (.pdf 5 pages). We measured 508,269 peers in 444 swarms within five BitTorrent communities, ranging from public to highly elite. We observe download performance, connectability, seeder/leecher ratios, seeding duration, and statistics regarding the resource supply.
We got inspiration for a novel blockchain design based on operating our own ledger and studying token economies. Our current work is called Trustchain, a unique design from 2012 where all participants have their own personal blockchain and create their own genesis block. Our older work used a graph-based approach and graph-based reputation algorithms. Trustchain records transactions in a tamper-proof and scalable manner. It does not require mining and does not try to solve the double spending problem. Our primitive 2007 ledger pre-dates Bitcoin, additionally our 2012 DAG-based approach pre-dates IOTA and the Texas DAG patents.
We are fans of Bitcoin, but also showed in an early analysis the flaws in this concept. Our approach to digital signatures is the essential difference which sets us apart from others. Mono-signatures form the foundation of all other projects we have seen in the past decade. Meaning, in systems such as Bitcoin a transaction is already valid with a single signature. Our Trustchain design does not permit transactions with merely a single signature. Trustchain only supports multi-party agreement recording, others are not valid. We believe that we created a more powerful system by removing single-signature transactions. Only time can tell the usefullness of this academically-pure and minimal design.
The foundation of our approach is making repeated successful interactions between actors explicit and durable. Cryptographically signed records of successful encounters serve as proof-of-work certificates. The validity and value of these certificates is determined by a trust and reputation system. Relaying for anonymity and seeding in Tribler constitutes work which is rewarded with a signed certificate. Helping others and uploading in Bittorrent swarms is rewarded with bandwidth tokens (e.g. signed certificates). Mining in our system becomes download parts of a swarm and uploading them to multiple interested parties. In 2013 we got the credit mining part of our system operational in early Beta. The screenshot below from November 2013 shows the boosting of various swarms. Note the investment yields of “struck gold” and “poor” in the right column.
For our narrow focus of a Bittorrent client we are exploring the fundamentals of identity, trust, and trade. With over 1 billion users of Youtube and Bittorrent we know there is a mass audience ready for something better.
Our approach has very boring foundations, when compared to newer and more sexy work, like IPFS, FileCoin, or Storj. We first measured Bittorrent in 2002, it is a flourishing mature ecosystem and ready for an upgrade. Bootstrapping an ecosystem is hard, we designed and deployed a superior alternative to Bittorrent. It became an official IETF Internet Standard, but completely flopped. This formed our preference for simplicity, elegance and our allergy for bloatware, clean-slate work, and over-engineering. Numerous other projects try to create a generic approach using an ICO for funding and promising the early adopters a dazzling return-on-investment. Tribler is different. rant warning. We are non-profit academics. We do not want to replace the old elite with a new crypto-currency elite. What is changed if we replace backroom deals, lobbyists, middleman, and legal monopolies with the tools of the new elite: algorithms, early investor rewards, proof-of-dominating-stake, and smart contracts? Replacing the analog world and breading digital-native inequality does not make the world a better place. We are creating a micro-economy based on fairness, trust, equality, and self-governance. By design we banish rent-seeking. Critical infrastructure rarely makes profit. We are trying to build critical infrastructure.
As of December 2014 Tribler has a build-in version of a Tor-like anonymity system. This is completely disconnected from ‘The’ Tor network. It is still ongoing work. It gives you probably superior protection than a VPN, but no protection against resourceful spying agencies.
We have implemented the main parts of the Tor wire protocol within Tribler. Instead of the TCP protocol that ‘the’ Tor network uses, we use UDP. The enables us to do NAT puncturing and traversal. We have created our own network using this Tor variant, our code is not compatible with normal Tor. Work started as a small trial in December 2013 with anonymous Bittorrent downloading. Essential part of our work is that everybody who downloads anonymously also becomes a relay. This brings the Bittorrent tit-for-tat idea to darknets. With this ongoing work we aim to offer in 2018 with Tribler V7.0 proxied downloading for any Bittorrent swarm.
Lengthy documentation in the form of two master thesis documents is available. First is a general documentation of the tunnel and relay mechanism, Anonymous HD video streaming, .pdf 68 pages. Second is focused on encryption part, called Anonymous Internet: Anonymizing peer-to-peer traffic using applied cryptography, .pdf 85 pages. In addition, there are the specifications for the protocols for anonymous downloading and hidden seeding on this wiki.
The current foundation of Tribler is the Dispersy overlay. Dispersy functionality includes: making connections, sending messages, puncturing NAT boxes, and distributed database synchronization. Every 5 seconds Dispersy sends out a message to establish a new connection or re-connect to a known peer. Note that we are transitioning to a new overlay for the durations of 2018.
Overlay communication, peer discovery and content discovery (keyword search) are essential building blocks of a peer-to-peer system. Tribler preserves the content and peers it discovered in the past. Every Tribler client runs a full SQL database engine. Several times per second each Tribler peer sends and receives updates for this database. Our protocol for distributed database synchronization is called Dispersy. See a simple messaging client written with just a few lines of code as a simple tutorial example; outdated broken tutorial.
Dispersy is a fully decentralized system for synchronization (.pdf), capable of running in challenged network environments. Key features of Dispersy are stateless synchronization using Bloomfilters, decentralized NAT traversal, and data bundle selection algorithms that allow the system to scale over 100,000 bundles in the presence of high churn and high-load scenario’s.
Dispersy uses a simple database schema, with the sync table containing the data bundles to synchronise across peers in the packet field.
Android porting teams are working on the downloading and Tor-like protocol part of Tribler and the overlay, channels and search portions. As of June 2014 there is initial running code. The focus is on stability and creating a mature build environment using Jenkins. See below two actual screenshot of current running code. Download the alpha .APK here: https://jenkins.tribler.org/job/Build-Tribler_Android-Python/lastBuild/
The following work is ongoing. We have an operational Android app that can spread itself via NFC. The app can spread viral via friends, even if it is blocked from a central app store.
Original student assignment: The aim is to create an Open Source Android smartphone app to help bypass restrictions by non-democratic governments. The Arab Spring showed the importance of video recording of mass protests. However, possession of a video recording on your phone of human rights violations and mass uprisings brings grave danger. The idea is to make this app “check-point-proof”, meaning that a somewhat knowledgeable person will not detect the presence of the app and will not discover any video content. The app itself should be hidden, you can make a “stealth” app by somehow removing the app icon from your app list (sadly it simply still shows up in the uninstall app list). The app is activated simply by “dialing” a secret telephone number or other method your deem secure. Starting point for your work can be found here: http://stackoverflow.com/questions/5921071/how-to-create-a-stealth-like-android-app. Your Stealth app need to be able to virally spread and be able to bypass an government restrictions on the official app store. Include the feature for NFC and direct-wifi transfer of the .apk with an easy on-screen manual and steps. Thus users can pass your app along to their friends.
Peer-to-Peer (P2P) networks work on the presumption that all nodes in the network are connectable. However, NAT boxes and firewalls prevent connections to many nodes on the Internet. We created a method to puncture NATs which does not require a server. Our method is therefore a simple no-server-needed alternative to the complex STUN, TURN and ICE approaches.
We conducted one of the largest measurements of NAT/Firewall behavior and puncture efficiency in the wild. Our method is a UDP hole-punching technique. We measured the success rate using volunteers running Tribler. Number of users in our trials are 907 and 1531 people. Our results show that UDP hole punching is an effective method to increase the connectability of peers on the Internet: approximately 64% of all peers are behind a NAT box or firewall. More than 80% of hole punching attempts between these peers succeed.
Brief description of our UDP puncture method in IETF draft
As Tribler scientists and engineer we are actively trying to make a better world. Our micro-economy is our living lab for experimenting with alternative models for capitalism. We aim to re-invent money by creating the first sustainable economy without any moral hazards from bankers, politicians, and megacorporation. Citizens and only the citizens are in control with self-governance.
Our grand vision in a 1+ hour lecture given at Stanford University, via their Youtube channel. We want to do more then be a Youtube alternative. Our grand vision is liberating both media and money. See the talk Abstract and slides (.pdf 78 pages). Keywords: transform money, “Bank-of-Bits”, global financial meltdown isolation. Use cooperation&stability, not volatility&greed. Alter the essence of capitalism (rich get richer) by abolishing compound interest rate and facilitation of safe zero-cost money transfers & lending. We aim for a direct assault on the essence of capitalism, aiming even further then the Bitcoin accomplishment (bypassing the central bank).
* 2014: Test network goes live for anonymous Tor-like downloading (not connected in any with with ‘the’ Tor project)
* 2010: Wikipedia.org uses our technology for live trial
* 2007: Our distributed ledger launched in the wild
* 2004: Slashdot for first time with largest Bittorrent study
...
Read the original on github.com »
are the nouns of your API. Deciding how to name and model these nouns is arguably the hardest and most important part of designing an API. The resources you expose organize your users’ mental model of how your product works and what it can do. At Increase, our team has used a principle called “no abstractions” to help. What do we mean by this?
Much of our team came from Stripe, and when designing our API we considered the same values that have been successful there. Stripe excels at designing in their API — extracting the essential features of a complex domain into something their users can easily understand and work with. In their case this most notably means modeling payments across many different networks into an API resource called a PaymentIntent . For example, Visa and Mastercard have subtly different reason codes for why a chargeback can be initiated, but Stripe combines those codes into a single enum so that their users don’t need to consider the two networks separately.
This makes sense because many of Stripe’s users are early startups working on products totally unrelated to payments. They don’t necessarily know, or need to know, about the nuances of credit cards. They want to integrate Stripe quickly, get back to building their product, and stop thinking about payments.“For Increase users, trying to hide the underlying complexity of these networks would irritate them, not simplify their lives.”Increase’s users are not like this. They often have deep existing knowledge of payment networks, think about financial technology all the time, and come to us because of our direct network connections and the depth of integration that lets them build. They want to know exactly when the FedACH window closes and when transfers will land. They understand that setting a different Standard Entry Class code on an ACH transfer can result in different return timing. Trying to hide the underlying complexity of these networks (by, for example, modeling ACH transfers and wire transfers with a single API resource) would irritate them, not simplify their lives.
Early conversations with these users helped us articulate what we dubbed the “no abstractions” principle as we built the first version of our API. Some examples of the way this mindset has subsequently affected its design:Instead of inventing our own names for API resources and their attributes, we tend to use the vocabulary of the underlying networks. For example, the parameters we expose when making an ACH transfer via our API are named after fields in the Nacha specification. Similar to how we use network nomenclature, we try to model our resources after real-world events like an action taken or a message sent. This results in more of our API resources being immutable. An approach that’s worked well for our API is to take a cluster of these immutable resources (all of the network messages that can be sent as part of the ACH transfer lifecycle, for example) and group them together under a state machine “lifecycle object”. For example, the ach_transfer object in our API has a field called status that changes over time, and several immutable sub-objects that are created as the transfer moves through its lifecycle. A newly-minted ach_transfer object looks like:{
“id”: “ach_transfer_abc123″,
“created_at”: “2024-04-24T00:00:00+00:00″,
“amount”: 1000,
“status”: “pending_approval”,
“approval”: null,
“submission”: null,
“acknowledgement”: null
// other fields omitted here for clarity
}After that same transfer has moved through our pipeline and we’ve submitted it to FedACH, it looks like:{
“id”: “ach_transfer_abc123″,
“created_at”: “2024-04-24T00:00:00+00:00″,
“amount”: 1000,
“status”: “submitted”,
// immutable, populated when the transfer is approved
“approval”: {
“approved_by”: “administrator@yourcompany.com”,
“approved_at”: “2024-04-24T01:00:00+00:00″
// immutable, populated when the transfer is submitted
“submission”: {
“trace_number”: “058349238292834″,
“submitted_at”: “2024-04-24T02:00:00+00:00″
// immutable, populated when the transfer is acknowledged
“acknowledgement”: {
“acknowledged_at”: “2024-04-24T03:00:00+00:00″
// other fields omitted for clarity
}If, for a given API resource, the set of actions a user can take on different instances of the resource varies a lot, we tend to split it into multiple resources. For example, the set of actions you can take on an originated ACH transfer is different (the complete opposite, really) than the actions you can take on a received ACH transfer, so we separate these into ach_transfer and inbound_ach_transfer resources.
This approach can make our API more verbose and intimidating at first glance — there are a lot of resources on the left-hand side of our documentation page! We think it makes things more predictable over the long-term, though.
Importantly, our engineering team has committed to this approach. When you design a complex API over several years, you make small incremental decisions all the time. Committing to foundational principles upfront has reduced the cognitive load for these decisions. For example, when sending a wire transfer to the Federal Reserve, there’s a required field called Input Message Accountability Data which serves as a globally-unique ID for that transfer. When building support for wire transfers, an engineer in an abstraction-heavy API might have to deliberate how to name this field in a “user-friendly” way - trace_number? reference_number? id? At Increase that hypothetical engineer names the field input_message_accountability_data and moves on. When an Increase user encounters this field for the first time, while it might not be the most immediately recognizable name at first, it helps them understand immediately how this maps to the underlying system.
No Abstractions isn’t right for every API, but considering the level of abstraction that’s appropriate for the developers integrating against it is a valuable exercise. This will depend on their level of experience working with your product domain and the amount of energy they’ll be committing to the integration, among other things. If you’re building an abstraction-heavy API, be prepared to think hard before adding new features. If you’re building an abstraction-light API, commit to it and resist the temptation to add abstractions when it comes along.Email jobs@increase.com to learn more about our open roles.Banking services provided by First Internet Bank of Indiana, Member FDIC. Increase is a financial technology company, not a bank. Cards Issued by First Internet Bank of Indiana, pursuant to a license from Visa Inc. Deposits are insured by the FDIC up to the maximum allowed by law through First Internet Bank of Indiana, Member FDIC.
...
Read the original on increase.com »
That’s four months after HashiCorp rugpulled their entire development community and ditched open source for the ‘Business Source License.’
As someone on Hacker News pointed out so eloquently:
IBM is like a juicer that takes all the delicious flavor out of a fruit
“HashiCorp already did a great job pre-draining all their flavor.”
Some people wonder if HashiCorp’s decision to drop open source was because they wanted to juice the books for a higher price. I mean, six billion dollars? And they’re not even a pointless AI company!
This blog post is a transcript of the video I posted today, Corporate Open Source is Dead. You can watch it on YouTube.
Meanwhile, Redis dropped the open BSD license and invented their own ‘Source Available’ license.
And last year, I covered how Red Hat found a way to just barely comply with the open source GPL license for their Enterprise Linux distro.
Other companies like MongoDB, Cockroach Labs, Confluent, Elasticsearch, and Sentry also went ‘Source Available’. It started with some of the smaller players, but as rot sets in at even the biggest ‘open source’ companies, open source devs are choosing the nuclear option.
Terraform, HashiCorp’s bread and butter, was forked into OpenTofu, and adopted by the Linux Foundation. Companies who built their businesses on top of Terraform quickly switched over. Even juicier, OpenBao—a fork of HashiCorp’s other big project Vault—is backed by IBM! What’s going to happen with that fork now?
At least forks seem pretty straightforward in Hashi-land. In the wake of Redis’ wanton destruction, it seems like there’s a new fork every week!
And some developers are even exploring ditching the Redis code entirely, like redka’s an API-compatible wrapper on top of SQLite!
After Red Hat closed its door—most of the way, at least they didn’t try pulling a switcheroo on the license itself! Oracle, SUSE, and CIQ scrapped together the OpenELA alliance to maintain forks of Enterprise Linux. And CentOS users who’ll be left in a lurch as June marks the end of CentOS 7 support have to decide whether to use AlmaLinux or one of the ELA projects now.
All these moves shattered the playbook startups and megacorps used—and now we’re seeing, abused—to build up billions in revenue over the past decade.
It was all in the name of ‘open source’.
As free money dries up and profits slow, companies slash headcount almost as fast as community trust.
2024 is the year Corporate Open Source—or at least any remaining illusions about it—finally died.
It’s one thing to build a product with a proprietary codebase, and charge for licenses. You can still build communities around that model, and it’s worked for decades.
But it’s totally different when you build your product under an open source license, foster a community of users who then build their own businesses on top of that software, then yoink the license when your revenue is affected.
Bryan Cantrill’s been sounding the alarm for years—yes, that Bryan Cantrill, the one who posted this gem:
Brian’s presentation from 12 years ago is worth a watch, and the bottom line is summed up by Drew DeVault:
[Contributor License Agreements are] a strategy employed by commercial companies with one purpose only: to place a rug under the project, so that they can pull at the first sign of a bad quarter. This strategy exists to subvert the open source social contract.
By working on a project with a CLA, where you sign away your code, you’re giving carte blanche for the company to take away your freedom to use their software.
From a company’s perspective, if they want CLAs or if they want to use an anti-open-source license, they do not care about your freedoms. They’re protecting revenue streams. They’ll often talk about freeloaders, whether it’s Amazon building a competing hosted solution, or some startup that found a way to monetize support.
But in the end, even if you have GPL code and you charge people to get it, it’s not truly free as in freedom, if the company restricts how you can use, modify, and share the code.
But there’s a distinction here, and I know a few people watching this are already yelling at me. There’s “free” software, and there’s “open source.”
People in the free software community correctly identified the danger of calling free software ‘open source.’
I don’t think we have to be so dogmatic about it, but there is a fundamental philosophical difference between the free software community, with organizations like the Free Software Foundation and Software Freedom Conservancy behind it, and the more business-oriented ‘open source’ culture.
Open source culture relies on trust. Trust that companies you and I helped build (even without being on the payroll) wouldn’t rugpull.
But time and time again, that trust is shattered.
Is this slow death of corporate open source bad? Well, it’s certainly been annoying, especially for devs like me who felt connected to these communities in the past. But it’s not all bad.
In fact, this could be a huge opportunity; what happened to the spunky startups like Ansible, HashiCorp, Elasticsearch, or Redis? They were lighting their industries on fire with great new software.
What happened to building up communities of developers, crossing cultural and economic barriers to make software that changed the world?
There are still projects doing that, but so many succumb to enterprise money, where eye-watering amounts of revenue puts profit over philosophy.
But as money dries up, as more developers get laid off after the insane hiring trends of the past five years, maybe small dev teams can move the needle.
The AI bubble hasn’t popped yet, so some great people are getting sucked into that vortex.
But someone else could be on the cusp of the next great open source project. Just… don’t add a CLA, okay?
And it’s not just devs; big companies can join in. Historically bad players like Microsoft and maybe even Oracle—man, it pains me to say that. They’ve even made strides in the past decade!
IBM could even mend some wounds, like they could reunite OpenTofu and Terraform. There’s precedent, like when IO.js merged back into Node.js after a fork in 2015.
People asked what Red Hat could do to get me interested in Enterprise Linux again. It’s simple: stop treating people who don’t bring revenue to the table like garbage. Freeloaders are part of open source—whether they’re running homelab or a competing business.
Companies who want to befriend open source devs need to show they care about more than just money. Unfortunately, the trend right now is to rugpull to juice the quarterlies, because money line always goes up!
But you know what? I’d just prefer honesty. If revenue is so dependent on selling software, just… make the software proprietary. Don’t be so coy!
But to anyone who’s not a multi-billion dollar corporation, don’t be a victim of the next rugpull. The warning signs are clear: Don’t sign a CLA. Stay away from projects that require them.
Stick to open source licenses that respect your freedom, not licenses written to juice revenue and prep a company for a billion-dollar-buyout.
Maybe it’s time for a new open source rebellion. Maybe this time, money won’t change company culture as new projects arise from the ash heap. Maybe not, but at least we can try.
...
Read the original on www.jeffgeerling.com »
To see all available qualifiers, see our documentation.
We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
We read every piece of feedback, and take your input very seriously.
You signed in with another tab or window. Reload to refresh your session.
You signed out in another tab or window. Reload to refresh your session.
You switched accounts on another tab or window. Reload to refresh your session.
...
Read the original on github.com »
To add this web app to your iOS home screen tap the share button and select "Add to the Home Screen".
10HN is also available as an iOS App
If you visit 10HN only rarely, check out the the best articles from the past week.
If you like 10HN please leave feedback and share
Visit pancik.com for more.