Google open-sources TorchTPU, a native PyTorch back end for TPUs

Posted by ilreb 1 day ago

Counter378Comment89OpenOriginal

Comments

Comment by pcmoritz 1 day ago

For people interested in this: I've been working on an open-source libtpu implementation for https://tenstorrent.com/ accelerators as a side project, it can already run most of torchtpu, see https://github.com/pcmoritz/libtt (based on the tenstorrent software stack). It shows all the things involved in lowering torch to tile based accelerators.

Comment by asparagui 1 day ago

Very cool! I have been using Lean to drive PJRT like so: https://github.com/brettkoonce/lean4-mlir and in theory your code could unlock another platform!

Comment by pcmoritz 1 day ago

Thanks for sharing, this is super cool! Right now my approach to correctness it to just run as many tests as possible and fix them (torchtpu greatly expands the amount of available tests), but formal verification is also super cool. Having a look at your book https://lean.brettkoonce.com/blueprint.pdf

The Tenstorrent ISA is documented btw https://github.com/tenstorrent/tt-isa-documentation, I wonder if one could do the verification end-to-end with the hardware.

Comment by yhjc2692 1 day ago

This is super cool!

Comment by pcmoritz 1 day ago

Thanks for the encouragement, if you have the hardware and end up trying it out I'd love to get your feedback (feel free to drop an issue in the repo) :)

Comment by yhjc2692 1 day ago

I would love to do that, but unfortunately don’t have one on hand (yet?). In fact, I’m curious - did you buy a TT device to work on this, or are you using a cloud provider like Koyeb?

Comment by pcmoritz 1 day ago

I first built my own to try it out (the exact hardware is described in https://relativity.cc/hardware/ under "one Blackhole card") and when I liked the hardware I got a quietbox 2. It is very high quality but a little expensive (but inexpensive compared to other LLM hardware and it has excellent interconnects which is very rare in the consumer/dev space).

Comment by dalemhurley 1 day ago

Google and Apple are going to move consumers and ordinary business users to edge inference. This is more disruptive to the upcoming IPOs than most investors understand.

Comment by lofties 1 day ago

Google, Apple, and don't forget our friends DeepSeek!

Comment by anuptalwalkar 1 day ago

I would love to see them move off into edge computing. There should be some autonomy and siloed experience in day to day usage.

Comment by tjwebbnorfolk 1 day ago

I would love to be able to buy a TPU for inference

Comment by gnabgib 1 day ago

You can.. they're called Coral. https://en.wikipedia.org/wiki/Tensor_Processing_Unit#Edge_TP... https://developers.google.com/coral https://physicalailab.io/en/google-coral-tpu-edge-tpu-status...

Comment by Dr4kn 1 day ago

They are just wildly outdated for most tasks. Even projects like frigate suggest buying different hardware. Coral was released in 2019 and hasn't been updated once. With a jetson nano or even a raspi ai hat you get way more tops

Comment by tjwebbnorfolk 22 hours ago

Ok I'll clarify: I want to buy a current-ish generation of TPU that can run LLMs. Coral is like 8 years old now

Comment by emmelaich 1 day ago

You can buy a Google Coral TPU, does that count?

Comment by Dr4kn 1 day ago

With them being released in 2019 with zero hardware revisions they are just outdated for almost every workload

Comment by villgax 23 hours ago

Yeah right, try generating an image on Apple Silicon or an Android phone with any of the Qwen or Flux family of models, let alone videos or prefill heavy LLMs.

Comment by npalli 1 day ago

Interesting, Amazon doing the same for AWS Trainium.

https://github.com/aws-neuron/torch-neuronx

Comment by AnotherGoodName 1 day ago

Pytorch conference is happening today for explaining the coincidental timing

Comment by sroussey 1 day ago

That explains the 5x speed increases huggingface mentioned for Mac using metal.

Comment by mathisfun123 1 day ago

Link?

Comment by fancyfredbot 1 day ago

Cool! XLA is pretty good + potentially a better compiler infrastructure than inductor which will lead to improved performance on all hardware.

Also I guess this is the final nail in the coffin of Tensorflow.

Comment by 23agt 1 day ago

Google is selling TPU shovels via Cloud services and also physically to select partners. So this makes sense.

There will be an abundance of shovels. Nebius already crashed by almost half from its hype peak.

Comment by airza 1 day ago

it doesn't seem like there's an abundance of shovels coming anytime soon

Comment by tjwebbnorfolk 1 day ago

Everyone says there's a massive compute shortage. But where is it showing up? For whom is there a shortage?

Every time I ask Claude to do something, it does it just fine. I never have to wait in any sort of queue that would be common in other kinds of shortages.

Comment by eitally 1 day ago

It's showing up for enterprises (and ISVs) who need capacity from cloud vendors and can't get it. Sometimes there are some GPU shapes available, but not always what you either need or want to pay for (e.g. a TPU vs a V100 vs a B300), and also not always in spot instances (frequently not).

The hyperscalers are all so oversubscribed that they require commits from enterprises so if you only have a transient need you might be out of luck [from your vendor of choice]. Additionally, not all GPU service providers are equal. Semi Analysis has some great articles explaining why.

Comment by octoberfranklin 1 day ago

But where is it showing up?

Pricing -- exactly where shortages show up in a free market economy.

Comment by tjwebbnorfolk 22 hours ago

Right, but pricing is for the marginal unit of hardware.

what I'm wondering: is there a shortage of compute in the current installed base? Why isn't the installed base of compute increasing drastically in price, and/or forcing me to wait in a queue to use it?

Comment by octoberfranklin 13 hours ago

Why isn't the installed base of compute increasing drastically in price

Have you looked at RAM prices lately?

Comment by airza 23 hours ago

have you tried to purchase a device with RAM or a GPU in the last two years?

Comment by tjwebbnorfolk 22 hours ago

Yes it's all very expensive. But is there not enough compute in general? Or is there not enough extra compute for me personally where I want it ?

like, I can rent a VM from AWS for the same price as 5 years ago...

Comment by bornfreddy 22 hours ago

I mean there is an abundance of shovels - but everyone wants to dig with excavators, and there's a shortage of those.

Comment by juleiie 1 day ago

Shovels and gold aren’t the problem, humans are the problem

either AI achieves hyper-efficiency and triggers systematic economic demand destruction due to unemployment, or it underperforms and sparks a massive tech-sector asset bubble burst

If AI is as good as valuations - it crashes. If AI isn’t as good, it crashes. all eggs in one basket economy.

Every future leads to a massive financial shitstorm

AI is priced in as a technology to end all markets. The last bull market you will ever see.

It’s a miracle markets didn’t notice this “damned if you succeed damned if you don’t” type of inevitable checkmate

I have been building a bunker

Comment by bsder 1 day ago

> It’s a miracle markets didn’t notice this “damned if you succeed damned if you don’t” type of inevitable checkmate

They learned from 2008 that Too Big Too Fail(tm) means the government bails you out.

This was the problem with not OBLITERATING the banks in 2008. Had the US government taken over the banks and wiped the investors, there would be more pushback right now.

Comment by omgwtfbyobbq 1 day ago

It's possible. If goog's AI overview is accurate, it seems somewhat similar to speculative euphoria seen in the past (eg companies involved in electricity prior to the great depression). It's larger (currently 35% for AI versus 25% for electric/related cos), I'm guessing because of more credit. At the same time, the housing component of the market prior to the GFC was about 23%, and it wasn't as economically damaging as the GD, so hopefully the AI bubble unwinding won't be as bad.

Comment by esseph 1 day ago

> so hopefully the AI bubble unwinding won't be as bad

It's currently larger than the 1880s railroad boom.

"Overbuilding and speculative debt ultimately culminated in the Panic of 1893, plunging the nation into a severe economic depression."

Comment by omgwtfbyobbq 1 day ago

Not according to Goog's. The railroad bubble of the 1880s was supposedly 60+% of the market. It may not be accurate though.

"During the 1880s, private railroad enterprises represented over 60% of the total capitalization of the entire U.S. stock market, peaking at 63% in 1881 and 1885."

Comment by esseph 21 hours ago

AI spending is significant more in inflation adjusted dollars. (Roughly 550 Bn in modern dollars). We're projecting trillions in AI spend.

But as a % of total GDP, you are correct.

Comment by omgwtfbyobbq 16 hours ago

I asked Goog's for enterprise value of the sector as a percentage of total market in all 4 situations, so it's not exact because enterprise value isn't available back to the 1800s, but it should be ballpark assuming the overall picture is accurate.

In terms of infra spending as a percent of GDP, 1880s railroads were at 5-6% at peak while AI is estimated to reach 3+% in a year or two.

Comment by fasterik 1 day ago

[dead]

Comment by ismailmaj 1 day ago

Amazing news, I remember the days when you had to fight XLA to make PyTorch run on TPUs, it seems we finally got first-party support!

Since HW specifics live mostly in the kernel layer (Pallas or CUDA), I imagine the Jax vs PyTorch question is more of a style preference now?

Comment by nico 1 day ago

Pardon my ignorance, but who else besides Google has TPUs? Are they commercially available somewhere? What’s their main application right now outside of Google?

Comment by cavisne 1 day ago

You can rent them on cloud + they sell directly to Anthropic. Its pretty well known that Anthropic is a heavy user of TPU's (along with Trainium & GPU's)

Comment by atty 1 day ago

Google is selling TPUs to enterprise clients for on prem deployments. We’ve evaluated them but decided to stick with Nvidia, at least for one more generation. We’re a relatively large user of GCP.

Comment by 1 day ago

Comment by Lucasoato 1 day ago

Are there TPUs outside Google Cloud?

Comment by boulos 1 day ago

Yes. Google is famously selling TPUs on prem to big customers (e.g., Anthropic) and Crux AI is now a venture to do just that.

Comment by sschueller 1 day ago

Google coral USB and m.2 devices are TPUs.

Comment by Archit3ch 1 day ago

"Have you seen any software run on Coral TPUs?"

"No, but I have seen the house that the project built, and it is terrific." ~ some Google director

Comment by charleslmunger 1 day ago

https://frigate.video

Is a pretty successful open source NVR that used corals when it launched (and many people still use them with it)

Comment by qlte 1 day ago

Yeah I have one for Frigate on a Raspberry Pi. Has worked well but definitely showing its age performance wise. I got a cheap mini PC (right before prices like quadrupled) which has about the same performance on an iGPU). I wouldn't be surprised if Frigate was the funnel for like 90% of Coral purchases or more.

Comment by boulos 1 day ago

Only in marketing name. They're not related otherwise.

Comment by goda90 1 day ago

Maybe just marketing, but they call the NPUs in their Pixel phones "TPUs"

Comment by fulafel 1 day ago

TPUs are really matrix processors (they process 2d tensors right?), it's technically fair to call most matrix processors TPUs.

Comment by westurner 1 day ago

TPU / neural compute accelerator / Math coprocessor;

"OpenTPU: Open-Source Reimplementation of Google Tensor Processing Unit (TPU)" https://news.ycombinator.com/item?id=44111452

From https://news.ycombinator.com/item?id=45848676 :

> How does Cerebras WSE-3 with 44GB of 'L2' on-chip SRAM compare to Google's TPUs, Tesla's TPUs, NorthPole, Groq LPU, Tenstorrent's, and AMD's NPU designs?

There are TPUs made out of silicon carbide with standard fab.

There is in-flight tensor processing with photons, but that's not quite a TPU.

Comment by hackernudes 1 day ago

I think the commenter was asking if Google sells TPU chip hardware to use outside of Google's own infrastructure and cloud.

Comment by tingkuanpei 1 day ago

Very interesting work. Google should have done this a long time ago. PyTorch's ecosystem is just so rich.

Comment by g58892881 1 day ago

great! where can I buy a TPU?

Comment by nelaggy 1 day ago

so if the moat isn't the model and the moat isn't the inference layer, i guess the moat is the hardware now? what does this mean for frontier labs?

Comment by badrequest 1 day ago

hardware has always been the moat in the AI space

Comment by AnotherGoodName 1 day ago

Yeah a lot of people bought into the idea of CUDA being a moat but we’re well past the point where it’s pretty trivial to jump to another equivalent stack on the software side.

Comment by jacquesm 1 day ago

Those are the two main interfaces: the inference hardware and the endpoints. Endpoints are pretty much standardized to the point that you can switch them out in seconds. CUDA still has an edge but every release like this chips away at that. Even so, TPUs are unobtanium so while I appreciate Google's effort here I also recognize that it is not something that will be useful to me.

In that sense AMD is still better positioned to give NVidia a run for their money but given the family connection and AMD's underdog status I am not very hopeful that we'll see meaningful competition in inference hardware any time soon.

NVidia is a much harder nut to crack than OpenAI or Anthropic. They are easily a decade ahead of viable competition. Check out the GB300 if you don't believe that.

Comment by kavalg 1 day ago

While I have the same sentiment, I am also curious why a competitor can't accelerate their developments with AI. In the end, hardware development is to a large extent also software development. Is it the patents that NVIDIA has that are preventing this? Or maybe some trade secrets? AFAIK the NVIDIA hardware is manufactured entirely by TSMC.

Comment by lp92 20 hours ago

Google does provide TPUs to big players like Anthropic and other large enterprise customers for on prem deploymens. That's their target just like those the GB300 is aimed at.

Comment by nixon_why69 1 day ago

A decade is a long time and the highest end isn't the only market segment.

High bandwidth flash will be here soon, I bet someone makes an economical competitor thats not quite as fast but much cheaper.

Comment by coredog64 1 day ago

Even if it's allegedly possible, the devil has always been in the details. It always sucks to find out that there's some quirk of your model architecture that breaks an implicit assumption that underpins the export process.

Comment by yaoke259 1 day ago

[dead]

Comment by AndrewKemendo 1 day ago

Will always be true

Comment by JacobAsmuth 1 day ago

Products can also be a moat.

Comment by yipinwong 1 day ago

Dumb question from non-AI engineer.

Is this somehow Google's answer to today's best DeepSeek open-sourcing inferencing engine?

Comment by bitpush 1 day ago

I dont think big companies like Google react to news like that. Very likely this was meant to go out on Tuesday for weeks now. And it just happened to have coincided with another news.

Comment by JacobAsmuth 1 day ago

Yeah they read the news this morning and whipped up a quick Torch->XLA translation layer and coordinated with four different inference engines (who also all wrote their own blog posts) and shoved this out the door quick to stay on top of the news cycle

Comment by parthdesai 1 day ago

Why be rude on purpose? OP clearly prefaced by saying it was a dumb question

Comment by caaqil 1 day ago

[flagged]

Comment by luckydata 1 day ago

[flagged]

Comment by mathisfun123 1 day ago

they're releasing this today because the pytorch conf is this week...

Comment by 0x00cl 1 day ago

We might never know, but they probably didn't read the news and started working to get this out, they were probably working on it for a while now.

Personally the question is if they are "saving" other projects or news so they can be released in case a competitor makes a move? Who knows if this was ready days or weeks ago and were waiting on someone to drop some news, so they can shadow those news.

Comment by geremiiah 1 day ago

No, these are two very different objectives.

Comment by Tanjreeve 1 day ago

They do different things and sit in different parts of the stack (training vs inference). TorchTPU seems more like a play to get more people familiar with Googles TPU chips as they want to sell them to people outside of google.

Comment by beardedetim 1 day ago

Yeah I'd assume this is to allow us to train on their TPUs with torch instead of needing to rent a GPU

Comment by yipinwong 1 day ago

ty. that helps a lot and know where to research next.

Comment by mathisfun123 1 day ago

lol no this is a replacement for https://github.com/pytorch/xla

Comment by frobisher 1 day ago

Question: does this make JAX obsolete? I was having fun learning it!

Comment by csvance 1 day ago

It makes PyTorch work nicely on TPU, which is impressive mostly because of how unstructured PyTorch is with its conventions. From a compiler standpoint JAX is much easier to work with since its all pure functions. This means that JAX also has access to fairly cutting edge compiler tech like Enzyme which is quite relevant for scientific machine learning work.

Comment by amelius 1 day ago

Soon to be written by AI anyway, so they might as well release it now.

Comment by tuna74 1 day ago

Is the Linux driver available?

Comment by saucesaft 1 day ago

bad news for jax

Comment by jax_bro 1 day ago

not really!

Comment by jart 1 day ago

Let me guess, they gave you something better and it's eating you up inside? That's the way your hard-core Commie works.

Comment by jax_bro 17 hours ago

what does this mean!

Comment by saagarjha 23 hours ago

wat

Comment by jart 21 hours ago

indeed https://www.destroyallsoftware.com/talks/wat

Comment by ranger_danger 1 day ago

How about open-sourcing the proprietary blackbox Titan security chip, that you pinky-promised to open but never did?

Comment by InferenceAcadem 1 day ago

[flagged]

Comment by Androcles 1 day ago

[flagged]

Comment by karansuthar 1 day ago

[dead]