Graphcore raises $222M at $2.7B valuation
https://techcrunch-com.cdn.ampproject.org/c/s/techcrunch.com/2020/12/28/ai-chipmaker-graphcore-raises-222m-at-a-2-77b-valuation-and-puts-an-ipo-in-its-sights/amp/

https://www.tenstorrent.com/press/

Tenstorrent, a hardware start-up developing next generation computers, announces the addition of industry veteran Jim Keller as President, CTO, and board member.

🚂 FPGA comes back. Titanium FPGAs from EFINIX are focused on the edge application and promises unbeatable per-Watt performance

⚙️[pdf] Hardware Accelerator of CNN by Yann Le Cun, father of Deeo Learning revolution

📺 [youTube] 20min introduction video from Intel about what are the FPGAs and what sort of applications can you use it for

🎱 [plumerAI] Yet another ML Hardware startup
keeps growing and explains why binarized neural networks do the job with less resources

🍰 [youTube] Bonus! The lecture from yesterday by professor Onur Mutlu, ETH, about GPU architecture.

🐻🐻🐻
This channel is back from hibernation and more reviews will come soon :)

🏋🏼Google finally released TPU v4, it will be avaliable for customers later this year.
🥴The previous v3 version was unveiled in 2018 and the v4 is claimed to be twice as fast.
🌽TPU v4 combines in a 4096 chips sumercomputer that reaches 1 exaFLOPs (10**18) of performance

Read more on [hpcwire] and watch the video Google I/O ‘21

🍀 The Hardware Lottery 🎰
by Sarah Hooker, Google Brain [ACM]

- The very first computer hardware was extremely focused on solving one particular problem - numerical differentiation or polynomial models. In the 1960s IBM invented the concept of Instruction Set and made migration between hardware easier for software developers. Till the 2010s we have been living in the world of general-purpose hardware - CPUs.

- Computer Science Ideas win or lose not because one superior one to another, but because some of them did not have the suitable hardware to be implemented in. Back Propagation Algorithm, the key algorithm that made the deep learning revolution possible, was invented independently in 1963, 1976, 1988 and finally applied to CNN in 1989. However, it was only three decades later that deep neural networks were widely accepted as a promising research direction and the significant result was achieved with GPUs, that could run massive parallel computations.

- Today hardware pendulum is swinging back to domain-specific hardware like it was the CPU invention

- Hardware should not remain a limiting factor for the breakthrough ideas in AI research. Hardware and Software should be codesigned for the SOTA algorithms. Algorithm developers need a deeper understanding of the computer platforms.

read also here

Hey folks,

That was a long period of silence here, but I'll try to breathe a new life to this channel  I'll do my best to post more frequently and not only ML hardware, but also about Zero Knowledge Proof accelerators, some areas on computer architecture, trading infrastructure and HPC

Here is  a video for today – what is Zero Knowledge Proof (ZKP)
ZKP is a way for proofer to convince a verifier, who has X and Y, that given for given function F  F(x,w)=y, without revealing w to verifier. 

But I find an explanation with a revealing of puffin absolutely genius in both in simplicity and clarity.

🪨🪨🪨🪨🪨🪨
🪨🪨🪨🐧🪨🪨
🪨🪨🪨🪨🪨🪨

Enjoy the video
Computer Scientist Explains ZKP in 5 Levels of Difficulty | WIRED

Re: Ask HN: How can ChatGPT serve 700M users when I can't run one GPT-4 locally?

I work at Google on these systems everyday (caveat this is my own words not my employers)). So I simultaneously can tell you that its smart people really thinking about every facet of the problem, and I can't tell you much more than that.

However I can share this written by my colleagues! You'll find great explanations about accelerator architectures and the considerations made to make things fast.

https://jax-ml.github.io/scaling-book/

In particular your questions are around inference which is the focus of this chapter
https://jax-ml.github.io/scaling-book/inference/

Edit:
Another great resource to look at is the unsloth guides. These folks are incredibly good at getting deep into various models and finding optimizations, and they're very good at writing it up. Here's the Gemma 3n guide, and you'll find others as well.

https://docs.unsloth.ai/basics/gemma-3n-how-to-run-and-fine-...

canyon289, 10 hours ago

Hey again, folks :)

Two years have passed since my last attempt to revive this channel. In that time, I left my job in prop trading, spent a year doing the MicroMasters in Finance, and in November ’24 joined a skyrocketing (according to Bloomberg) hedge fund. Now we’re driving its HFT strategies to the moon with FPGA platforms and highly optimised software.

The flip side of the coin — I’m still far, far away from machine learning hardware. Not as far as I was when I started this channel half a decade ago, but still not quite there.

So, here’s my proposal to make this group a bit more interactive:
I want to start a learning group on the material mentioned above — How To Scale Your Model.

Plan: 12 sections → 12 seminars. Each seminar is a 1-hour call where we read through the chapter, rephrase it, and discuss.
No fees, no recordings — pure enthusiasm and passion.

Schedule: Tuesdays and Thursdays, 7:00–8:00 BST (UTC+1).

DM me if you’re interested and ready to invest your time.🕰️

@vconst89


P.S. This text was grammar-checked by ChatGPT but originally typed on my phone, so any remaining quirks are 100% mine.