The site confuses the inference engine in the Edge TPU with the datacenter TPU. They are two unrelated projects. Based on the paper they're borrowing from, I think they are trying to go for a much older datacenter inference-only TPU, or only implementing the inference capabilities of the datacenter TPU.
walterbell · 7h ago
Are there recent papers on datacenter TPU?
mdaniel · 9h ago
Yeowzers that FAQ is filled with watch-outs
The /forks contained https://github.com/csirlin/OpenTGPTPU which had a commit 3 hours ago but it seems they have not yet updated the FAQ for their version. Anyway, the fact it has commits greater than 8 years ago makes it seem like a more reasonable submission
> The TPU is Google's custom ASIC for accelerating the inference phase of neural network computations.
this seems hopelessly out of date/confused
walterbell · 5h ago
Additional text from Google's 2017 paper abstract says:
This paper evaluates a custom ASIC---called a Tensor Processing Unit (TPU)---deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory.
The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs (caches, out-of-order execution, multithreading, multiprocessing, prefetching, ...) that help average throughput more than guaranteed latency.
The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters.
TPUs were purpose-built specifically for AI. TPUs are an application-specific integrated circuit (ASIC), a chip designed for a single, specific purpose: running the unique matrix and vector-based mathematics that’s needed for building and running AI models..
TPU v2.. built an interconnected machine — our first TPU pod — with 256 TPU chips connected with a very high-bandwidth, custom interconnect.. liquid cooling was added with TPU v3 to help address efficiency needs, while TPU v4 introduced optical circuit switches to allow the chips in pods to communicate even faster and more reliably.
TPUs also underpin Google DeepMind’s cutting-edge foundation models, including the newly unveiled Gemini 1.5 Flash, Imagen 3, and Gemma 2, propelling advancements in AI.. Forget about a single chip, or a single TPU pod — we’re building a global network of data centers filled with TPUs.
surfmike · 3h ago
How would you describe it instead? Curious and learning
imtringued · 33m ago
Google does everything, both inference and training, on their TPUs.
Inference is easier, since the person deploying a model knows the architecture ahead of time and therefore can write custom code for their particular model.
When training you want to be as flexible as possible. The framework and hardware should not impose any particular architecture. This means lots of kernels and combinations of kernels. Miss one and you're out.
"Google Edge TPU devices", 100 comments (2019), https://news.ycombinator.com/item?id=19130896 & https://news.ycombinator.com/item?id=19313813
"Coral Edge TPU review", 100 comments (2020), https://news.ycombinator.com/item?id=24808755
"TPU transformation: 10 years of our AI-specialized chips", 60 comments (2024), https://news.ycombinator.com/item?id=41148532
The /forks contained https://github.com/csirlin/OpenTGPTPU which had a commit 3 hours ago but it seems they have not yet updated the FAQ for their version. Anyway, the fact it has commits greater than 8 years ago makes it seem like a more reasonable submission
this seems hopelessly out of date/confused
Inference is easier, since the person deploying a model knows the architecture ahead of time and therefore can write custom code for their particular model.
When training you want to be as flexible as possible. The framework and hardware should not impose any particular architecture. This means lots of kernels and combinations of kernels. Miss one and you're out.