Talk:Tensor Processing Unit

adding Nvidia specific solutions?

Nvidia has stated that their DLA unit (maybe part of the Volta accelerator; going to be a part of e.g. the Xavier SoC) is a TPU that Nvidia explicitely open-sourced. See the source links here: Tegra#Xavier. Do you think such a realisation of the concept by a third-party (non-Google) should be added to this article - or shall it receive it's own article? As of now i would prefer the first option. The topic is still compact and seeing the TPU concept evolve over time in a single article might rather benefit the reader - especially because then it is capable of covering all active vendors for this topic. --Alexander.stohr (talk) 15:24, 22 August 2017 (UTC)

A single source for discussing TPU devices seems like a good idea as this market grows. However, I am not sure if this is the place. Right now this is focused on the Google TPU products, and would require a distinct re-write to make it generic. I am not opposed to a doing so, but I am not sure if there is a consensus that TPU has been genericized. (In another context Nvidia has called the same/similar co-processor as a "tensor core" rather than TPU.) Another idea might be a hardware accelerator section within the TensorFlow article, with links to Nvidia and Google product pages. Thoughts? Dbsseven (talk) 15:38, 22 August 2017 (UTC)

I would think if independent reliable sources call it a TPU then it would be okay to incorporate the information into this article. News articles may just repeat Nvidia's talking points so it would be best to rely on technical/academic sources. Sizeofint (talk) 18:33, 22 August 2017 (UTC)

References

Rename to "Google TPU"

The following is a closed discussion of a requested move. Please do not modify it. Subsequent comments should be made in a new section on the talk page. Editors desiring to contest the closing decision should consider a move review. No further edits should be made to this section.

The result of the move request was: not moved. (non-admin closure) Steel1943 (talk) 20:34, 28 December 2017 (UTC)

---

Tensor processing unit → Google TPU – The page describes mainly Google TPUs. I think that's why the company name should be put into a title. 2.92.113.239 (talk) 17:31, 17 December 2017 (UTC)

• Oppose – Does not look like a commercial branded product, rather some R&D chip type; other companies make similar chips, even if under different names. Better expand the article to define the generic structure and include all relevant chips. — JFG ^talk 08:33, 18 December 2017 (UTC)
• Oppose I agree with JFG. Look for other AI FPU's that work on the high speed, low precision fpu principle and expand the article accordingly. scope_creep (talk) 10:49, 18 December 2017 (UTC)

scope creep and @JFG: FYI, there is already a highly related general article: AI accelerator. Therefore I'm not sure this is the place for expansion/generalization, unless there is a lot of Tensor specific content. Dbsseven (talk) 17:25, 18 December 2017 (UTC)

• Oppose - I'm not sure adding "Google" aids the article in any way. Right now there isn't a need for further specificity in the title (IMO). Dbsseven (talk) 17:25, 18 December 2017 (UTC)

• Favour - Google does not have the right to appropriate the word "Tensor", a term specific to mathematics and physics, just as no vendor has the right to appropriate the term "CPU" or "GPU" for one of its products. TPUs are simply a new type of processor specializing in implementing tensor mathematics. Also, other vendors are already manufacturing TPUs, such as nVidia's Volta microarchitecture chip, a combination GPU and TPU, 27,648 of which are currently powering Summit, the World's most powerful supercomputer at Oak Ridge National Laboratory. See: Tensor Cores in NVIDIA's Tesla V100 GPU, a direct competitor of Google's TPU. This media article offers an example of the TPU acronym used generically. Therefore this page ought to instead refer to generic TPUs and onclude references to other pages describing specific instances of TPUs offered by different vendors. (NOTE: The request to close this discussion is extemporaneous given important arguments such as this one have not been made.) 181.88.207.203 (talk) 05:06, 27 June 2018 (UTC)

• Comment I added a WP:PAID disclosure request to the user, User talk:2.92.113.239, because I think it is push to get some free advertising on the part of Google, re: this article. I'm not saying generalize it, I'm looking to add as much detail as possible. This is one processor of class of processors, that should be described by their architecture and api, not by name. I don't like the AI accelerator article, it is essentially a pamphlet offering free advertising. Dbsseven, Kudos to yourself, I see you have tried to smarten it up a bit, but the article is a first cut, and I think it is rank. Statement like this, (in this article) which are almost fancruft: Google stated the first generation TPU design was memory bandwidth limited, instead of the first generation TPU design was memory bandwidth limited.The first violates WP:NOTAVERTISING. The second doesn't. The point i'm trying is make is. It is new field, and lots of new disruptive designs are coming out, and everybody is trying to find what works, but processor design follows an ethos, a design language, they come of the uni's after decades of research, so I know for a fact there is other ultra high speed, low precision FPU processors out there. This article is about the TPU, but it should have been a architecture articles describing ultra high speed, high bandwidth, low precision FPU processors, with the TPU an example, amongst several others, including a good descritpion of the architecture. scope_creep (talk) 18:21, 18 December 2017 (UTC)

---

The above discussion is preserved as an archive of a requested move. Please do not modify it. Subsequent comments should be made in a new section on this talk page or in a move review. No further edits should be made to this section.

Should this remain only dedicated to Google's TPU?

I'm not proposing we move or rename the article, but other companies are starting to sell chips they call "Tensor Processing Units" as well, and the term's been used in research for quite some time. As of October 2018, is it better to incorporate encyclopedic coverage of other TPU content besides Google's? --Daviddwd (talk) 22:55, 4 October 2018 (UTC)

TPU - Tensor Processing Unit

Again, TPU's are a type of processor dedicated to computing Tensors and are manufactured by multiple vendors, as is the case with CPUs and GPUs. TPUs are NOT specific to Google, thus this article needs to me made generic to incorporate info on TPUs from other vendors as well. Thanks. 181.12.115.177 (talk) 04:22, 4 April 2019 (UTC)

The product name "Tensor Processing Unit" is specific to Google. However products that provide tensor processing are not specific to Google; see AI accelerator. A possible cause of confusion is the "tensor cores" used by nVidia, including Volta (microarchitecture). This confusing, but to my knowlegde Google is the only company that uses the specific wording "Tensor processing unit ". --Jules (Mrjulesd) 10:47, 4 April 2019 (UTC)

The article says "Google's TPUs are proprietary and are not commercially available". But they are commercially available. https://coral.withgoogle.com/docs/accelerator/get-started/ . I've updated the article. Sayitclearly (talk) 13:43, 3 September 2019 (UTC)

Again, Google is appropriating the word "Tensor" for its own use. This is inappropriate: Words are for everyone to use and "tensor", like "vector", is no exception. For the uninitiated, tensor describes a field of vectors and is not at all bound to the AI. To the point, tensors were invented by italian mathematician Ricci-Curbastro to aid Einstein express his famous General Relativity in a single equation in an epoch were the computer hadn't even been invented, much less AI. The other words are "Processor" and "Unit", also, I believe, of general use. So should companies also have rights to CPU - Code Processing Unit, FPU - Floating Point Unit and GPU - Graphics Processing Unit? Of course not! So why would TPU - Tensor Processing Unit be the exception? Please, let's recover TPU for general use. 181.12.78.206 (talk) 19:34, 14 December 2020 (UTC)

August 2021: Google Nest cameras will include TPUs

Worth adding that the 2021 refresh of Google Nest Cams (including the Nest Doorbell, no longer called Nest Hello) will include TPUs? the TPUs will allow immediate processing and alerting. (Also offline recording and storing, due to battery inclusion). This section by DAVIDBSPALDING, but using work computer, so can't sign in to Wiki. -- 208.242.14.199 (talk) 15:34, 12 August 2021 (UTC)

Products

The table is cryptic. Perhaps "Mib" and "W" could be use full names. Bill W102102 (talk) 09:57, 13 August 2022 (UTC)