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.
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
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) :)
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?
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).
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.
They are just wildly outdated for most tasks.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
I've been really hoping to get Whisper running on a coral, but it can't even do that. There was recently some progress getting it to export to tflite (and even an official example added) but the coral is missing some instructions.
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.
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.
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.
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?
No, rationing is what you get in command/centralized economies like the USSR. In free market economies you get price explosion, like we're seeing right now with DRAM wafers.
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
> 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.
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.
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."
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.
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?
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)
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.
> 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.
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.
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.
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.
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.
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.
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.
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.
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
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.
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.
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.
