Peter Hizalev

Tenstorrent can run AI magically fast. Try it yourself. console.tenstorrent.com - Super fast large LLMs. DS R1 685B model flying at 350t/s/u on 16 Blackhole galaxies - Super fast video generation with Wan 2.2 Lighting (Prodia) finishing 5 seconds clip within 3 seconds on 4 galaxies. I know you want more. You’ll get it fast. The team is cooking more.

Tune in tomorrow! Run fast video, speech, code all on Tenstorrent Galaxy Blackhole. Powered by our Networked AI architecture with native scale-out. Hear from our partners and customers deploying at scale. Watch the livestream on May 1st @ 1:30 PM PDT: tenstorrent.com/deploy

At TT-Deploy we will be introducing TT-Lang—the new Python DSL to supercharge creation of new operations and frontier models. Join us to learn more!

News is out @tenstorrent + @prodialabs made the worlds fastest video generation

FlashAttention on Tenstorrent, a technical report

@jimkxa In hardware they already moved to tiled and async programming models, which is why Triton existed and now is extended by OpenAI and Meta to support explicit async threads. Once there is critical mass of these kernels in GenAI training datasets they become AI-portable.

Curious. Did Nvidia end the Cuda "moat" ? If they move to tiles like most other hardware, the AI kernels will be easier to port. x.com/nvidiaaidev/st…

@JLarky Who is Joe Rogan?

Dad, we have been through this, stop telling people you take Ketamine like Elon Musk, you take Creatine like Joe Rogan!

Part 3 of "Matrix Multiplication on Blackwell" is here! It continues our epic journey of describing how Modular implemented the fastest B200 matmul in the industry, revealing the techniques to go from 16% to 85% of SOTA modular.com/blog/matrix-mu…

@JLarky Can’t AI do what he offers?

that's like the most messages I get in DMs nowadays :(

@JLarky Fine, it won’t replace jobs.. in Oregon

Do you think people who are saying that AI will replace jobs ever heard about pumping gas in Oregon?

@clattner_llvm I would be super interested in your take on performance. Is there a true need of automatic exploration of tiling space given today’s hardware and workloads? What’s the place for runtime auto tuning (TVM, Triton)? And for compile-time exploration (Part-IR, XLA Shardy)?

Let me know if you have any specific suggestions or request on these topics.

I got busy with other things, so the next Democratizing AI Compute post is aiming for next week. Stay tuned to dive into C++’y OpenCL/SyCL/OneAPI and compiler’y XLA/MLIR tech. We will look at what worked and what struggled, with a goal of understanding the past.

@antirez lobste.rs is not it?

Not the right moment, give me some time, but soon or later I'll ask for the help of fellow hackers (something like 20 folks), create a first "open technology news group", and create a competitor of Hacker News that has transparent rules and lacks any personal goal.

@kurt Gitlab

Does anyone have an example of a company that's built a piece of SaaS that works great for startups and ALSO works well for enterprise companies?

@burkov is there a place to submit errors in your LM book?

@JLarky I am using UML and mouse to visually design my software. It automatically generates all necessary code. (Circa 2002)

- oh, so you are a developer? What language do you use? - mostly English and a lot of pressing of the [tab] key

@antirez Let’s keep open mind: Stochastic parrots are the new frontier for AI reasoning

You may think contortionists are able, but then give a look at AI researches that need to move from "LLMs are just stochastic parrots" to "LLMs are the new frontier for AI reasoning".

@eevblog What’s on the logo? Is it “staring into the abyss” or “going down the drain”?

The economy is going to crap, so I'm branching out into a smart business to capitalise.

@JLarky Who knew that communism only needed few more yottaflops to work!

@petrohi Or go straight to communism

Why does it have to be coding? Like if AI is so smart, just ask it to make you money. Skip the middle man. Just ask it to create you a business, why is your first thought to ask it to build an app or a website?

@burkov Same way we can argue that a typical RGB input image for Conv2d has naturally a NHWC layout. It beats me why PyTorch insists on putting channels after batch, but it does for some reason.

@petrohi 1D kernels are used in text processing where we have: batch dimension, sequence length, and embedding. So, to use Conv1d we have to x.permute(0, 2, 1).

I love PyTorch but what the hell: why nn.Conv1d expects the input of the shape (batch_size, channels, length) and not (batch_size, length, channels)?