The quote is:
“As TPU demand grows from AI labs, capital markets firms, and high-performance computing applications, we'll begin to deliver TPUs to a select group of customers in their own data centers in a hardware configuration to expand our addressable market opportunity.”
It’s a competitive advantage which allows them to deliver lower cost per inference. If quality continues to asymptote, best value becomes most interesting to consumers.
Then again, if they can provide competition to Nvidia, it may force Nvidia to lower prices across the board.
Every TPU they keep for Google Cloud reduces the need to buy GPUs from Nvidia. Or really, in this market, means they have additional ability to support AI compute beyond what they can buy from Nvidia.
Because it has been historically strongly coupled to the Google infra, both hardware and software. Now the situation has improved and now they are selling this to external clients, though it is still in a limited fashion.
They are finally starting to. Oracle just announced a week or so ago that they're going to be introducing TPUs to their cloud next year for customers to use.
Interesting times ahead with a number of chip makers now vying for this market. TPUs are a mature chip now which gives some strong advantage. At the same time as the TPU announcement, they also mentioned they'll be deploying Cerebras systems as well.
This is a marketing announcement, it's got the right level of info for the target audience. If you want more technical detail there's loads available elsewhere online.
There does come a point where if you're in the follow up article discussing what's new in 2 new variants of the 8th generation of a niche product for that vendor it stops being reasonable from them explaining what everything means each time and starts being reasonable to either click the links to the earlier introduction or search around for it to get a proper understanding as if you didn't know "TPU" was "Tensor Processing Unit" then there is a lot more you want to know before reading this article.
I disagree. I think you're saying that because this is your niche and finding the term isn't difficult for you.
However, consider the Tolkien community. I enjoy lord of the rings and want to know more. But there's an absurd acronym for another book (home) that no one ever expands in that community. And it's not in the wiki's and it's not in Google (bc what do you get if you google "tolkien home"?) it was intentionally frustrating for me to figure out what that acronym was. The people in the tolkien community also adopted your pov and just wouldn't help, expecting new people to bootstrap themselves. It was not a fun experience. These styling conventions aren't for the 99% of the time they're for the 1% of the time.
The experience can't have been that bad, you just used "bc" and "pov" without batting an eye. I'm sure it's not a one off from you either. Reasonably so, because it'd be crazy to expand and explain what everything means all the time when it takes so little effort for anyone who wants to to familiarize themselves yet is so convenient for everyone who already knows in a place we'd expect it to be extremely common to understand such things.
And what I'm saying here is different from "you should not ever care to expand acronyms/abbreviations/initialisms", but that does not necessarily mean your goal should be that it's "fun" for every possible person (you'll never achieve that goal, under or overexplaining is never fun). It's about what's reasonably efficient for everyone likely to read the article. If 50%, 10%, 1%, and 0.1% of your audience will understand an acronym, abbreviation, or piece of jargon then it has different implications for what the right balance of reasoning is. Not every piece need target a generic audience, not every acronym need assume the reader can't be bothered to learn anything not spelled out for them in the page, and not everything is equally hard to figure out when it's not spoon-fed to the reader.
And, again, what is "Tensor Processing Unit" going to gain you if you don't already know that's what TPU stands for in the context of this article? I.e. who knows what a tensor is and why youd have a processor for them yet gains something by the acronym being expanded? People who don't know what a TPU is should start with a different article completely, probably several, and then read this one. Expecting every piece of writing should assume no knowledge blocks more knowledge than it helps share.
One thing I think the "lucky 10,000" XKCD gets bang on is that it doesn't frame sharing something with the 9,999 in mind as error any more than it blames the 1 for not knowing - instead the problem is dogging people for saying they've never heard of something. That's truly a golden rule, regardless the audience and so on, and doesn't require changing the way humans write to avoid problems like the one you ran into.
I see your point, but it's good practice to introduce abbreviations at least once before using them. I even still see GPU defined at least once in most professionally written articles.
Also, it doesn't just come out and say what the peak throughput of these chips are, just 2.7x the prior generation. This is one of those things that Apple does every Mac release--it's always some big dumb bar that is bigger or some line up and to the right that has absolutely no explanation.
It is good writing practice, even in technical documents, to spell out an acronym the first time it is used in your document.
> We first designed Tensor Processing Units (TPUs) more than a decade ago motivated by Google’s...
Literally all it takes. Basically free and it helps both broaden your audience and avoid mistakes. Also, high key, it is a great opportunity for a back link to https://cloud.google.com/tpu, which you'd think Google would care about.
To be fair, does it matter? It seems pretty clear what they are to be used for. Does Nvidia define what CUDA is anymore? It doesn't really matter what the actual acronym is.
Full disclosure: I work in this space at Google, but my opinions here are my own.
You benefit indirectly because it gives Nvidia a competitor in the datacenter space and takes some pressure off the GPU market. You can also rent them on GCP.
Plenty of people and firms use TPUs, both directly and indirectly. You can rent them via Google Cloud. Apple Intelligence and Anthropic both use them. And of course your Gemini traffic is using the TPUs as well.
They were originally a separate company (and I think still based in the UK) so it's a bit more separate than probably random other internal departments
There are various bits of this blog post that detail how they’re making it easier for devs to use TPUs, particularly the discussion of TorchTPU and the attendant link. That being said if you’re looking for a link to a tutorial I think it’s easily findable on Google. “Tpu tutorial Google cloud” or “Tpu developer hub” would be good queries. For example [this](https://docs.cloud.google.com/compute/docs/tpus/quickstart-c...) is at least somewhat responsive.
I don’t know why everyone keeps complaining! It’s just a simple twenty-seven step process neatly spread across three portals! (Documentation coming soon.)
Title: Our newest TPUs: 8tx and 8i Max, built for the agentic era Small discussion previously (16 points, 4 comments) https://news.ycombinator.com/item?id=50425018
Their TPUSync post is highly recommended as well. This is basically how many of the open-source repos like Strata for Qwen series models are working.
https://cloud.google.com/blog/topics/ai-infrastructure/intro...
https://github.com/Niko1221/Strata/blob/main/docs/HOW_IT_WOR...
Wow ... Cloud Managed Lustre. That has to be a nightmare to operate.
I wish Google Colab start to provide new TPU model, the most powerful we can get on Colab is still v6e-1 which is too small for serious work.
Why doesn’t Google sell these?
They are now https://au.finance.yahoo.com/news/google-to-sell-tpu-chips-t...
The quote is: “As TPU demand grows from AI labs, capital markets firms, and high-performance computing applications, we'll begin to deliver TPUs to a select group of customers in their own data centers in a hardware configuration to expand our addressable market opportunity.”
They could still be renting them out.
It’s a competitive advantage which allows them to deliver lower cost per inference. If quality continues to asymptote, best value becomes most interesting to consumers.
Then again, if they can provide competition to Nvidia, it may force Nvidia to lower prices across the board.
The huge capex of designing and taping out a chip generally means that you wanna make and sell as many of said chip as possible.
Every TPU they keep for Google Cloud reduces the need to buy GPUs from Nvidia. Or really, in this market, means they have additional ability to support AI compute beyond what they can buy from Nvidia.
Google wants to rent as many TPUs as possible.
Only if you think the external market is relevant in comparison to Google itself.
Because it has been historically strongly coupled to the Google infra, both hardware and software. Now the situation has improved and now they are selling this to external clients, though it is still in a limited fashion.
They are finally starting to. Oracle just announced a week or so ago that they're going to be introducing TPUs to their cloud next year for customers to use.
https://www.msn.com/en-us/technology/tech-companies/oracle-e...
Interesting times ahead with a number of chip makers now vying for this market. TPUs are a mature chip now which gives some strong advantage. At the same time as the TPU announcement, they also mentioned they'll be deploying Cerebras systems as well.
They do. In particular they sold to Anthropic.
But they don't sell to you or I, which is what matters. Why would people care about Anthropic or Oracle?
And even if Google sold them, they probably won't sell any affordable ones. Nvidia sells DGX Spark, but who can afford to buy them?
Enterprise support is far easier than random customer support.
They don't care about you. Anthropic/Oracle pays a sh*t ton of $$$ for it. How much can you pay? lol
Where's the watt/token or $/token figures for each of these?
And how does each compare to their NVDA price equivalent?
Is it just me or the write up have very little information? I can’t grasp how this compare to other tech.
This is a marketing announcement, it's got the right level of info for the target audience. If you want more technical detail there's loads available elsewhere online.
The article mentions "TPU" 79 times, but never explains what the term means.
(It's "Tensor Processing Unit".)
There does come a point where if you're in the follow up article discussing what's new in 2 new variants of the 8th generation of a niche product for that vendor it stops being reasonable from them explaining what everything means each time and starts being reasonable to either click the links to the earlier introduction or search around for it to get a proper understanding as if you didn't know "TPU" was "Tensor Processing Unit" then there is a lot more you want to know before reading this article.
I disagree. I think you're saying that because this is your niche and finding the term isn't difficult for you.
However, consider the Tolkien community. I enjoy lord of the rings and want to know more. But there's an absurd acronym for another book (home) that no one ever expands in that community. And it's not in the wiki's and it's not in Google (bc what do you get if you google "tolkien home"?) it was intentionally frustrating for me to figure out what that acronym was. The people in the tolkien community also adopted your pov and just wouldn't help, expecting new people to bootstrap themselves. It was not a fun experience. These styling conventions aren't for the 99% of the time they're for the 1% of the time.
It's not an intuitive leap to google "tolkien home acronym" of which both the AI summary and the first result give the answer.
It's also quite similar to CPU
But yeah in general it's nice to at least explain the acronym once
The experience can't have been that bad, you just used "bc" and "pov" without batting an eye. I'm sure it's not a one off from you either. Reasonably so, because it'd be crazy to expand and explain what everything means all the time when it takes so little effort for anyone who wants to to familiarize themselves yet is so convenient for everyone who already knows in a place we'd expect it to be extremely common to understand such things.
And what I'm saying here is different from "you should not ever care to expand acronyms/abbreviations/initialisms", but that does not necessarily mean your goal should be that it's "fun" for every possible person (you'll never achieve that goal, under or overexplaining is never fun). It's about what's reasonably efficient for everyone likely to read the article. If 50%, 10%, 1%, and 0.1% of your audience will understand an acronym, abbreviation, or piece of jargon then it has different implications for what the right balance of reasoning is. Not every piece need target a generic audience, not every acronym need assume the reader can't be bothered to learn anything not spelled out for them in the page, and not everything is equally hard to figure out when it's not spoon-fed to the reader.
And, again, what is "Tensor Processing Unit" going to gain you if you don't already know that's what TPU stands for in the context of this article? I.e. who knows what a tensor is and why youd have a processor for them yet gains something by the acronym being expanded? People who don't know what a TPU is should start with a different article completely, probably several, and then read this one. Expecting every piece of writing should assume no knowledge blocks more knowledge than it helps share.
One thing I think the "lucky 10,000" XKCD gets bang on is that it doesn't frame sharing something with the 9,999 in mind as error any more than it blames the 1 for not knowing - instead the problem is dogging people for saying they've never heard of something. That's truly a golden rule, regardless the audience and so on, and doesn't require changing the way humans write to avoid problems like the one you ran into.
I see your point, but it's good practice to introduce abbreviations at least once before using them. I even still see GPU defined at least once in most professionally written articles.
On professionally written articles on nvidia.com?
Is it going to kill them to spell it out once?
Also, it doesn't just come out and say what the peak throughput of these chips are, just 2.7x the prior generation. This is one of those things that Apple does every Mac release--it's always some big dumb bar that is bigger or some line up and to the right that has absolutely no explanation.
It also mentions GPU and CPU, should they spell those out too?
Not necessarily. Those terms are, at least in my experience, much better known that "TPU".
> Is it going to kill them to spell it out once?
would it kill you to use an internet search engine (such as google)?
I'm not the one marketing my product.
You're not the target.
If you don't know what TPU means in that context, I don't think they are marketing to you. The first TPU was released 11 years ago!
There are always new customers.
It is good writing practice, even in technical documents, to spell out an acronym the first time it is used in your document.
> We first designed Tensor Processing Units (TPUs) more than a decade ago motivated by Google’s...
Literally all it takes. Basically free and it helps both broaden your audience and avoid mistakes. Also, high key, it is a great opportunity for a back link to https://cloud.google.com/tpu, which you'd think Google would care about.
To be fair, does it matter? It seems pretty clear what they are to be used for. Does Nvidia define what CUDA is anymore? It doesn't really matter what the actual acronym is.
Maybe they assumed you would just Google it ;)
New gpu chips are like the old IBM Watson. I'll never get one. "That's great, but it does nothing for me."
Full disclosure: I work in this space at Google, but my opinions here are my own.
You benefit indirectly because it gives Nvidia a competitor in the datacenter space and takes some pressure off the GPU market. You can also rent them on GCP.
Good to know, thanks, and all the best.
TFA makes some claims but offers zero data on specs or anything else. Is this a press release?
> 8t for large-scale training and 8i for real-time reasoning
> TPU 8i Max for massive context windows and the highest compute density per rack, and TPU 8tx for secure and confidential inference
You had one job, Google
And again no one can use it.
Plenty of people and firms use TPUs, both directly and indirectly. You can rent them via Google Cloud. Apple Intelligence and Anthropic both use them. And of course your Gemini traffic is using the TPUs as well.
I know right?
> TPUs are now a rapidly growing business with nearly 100 customers
‘Nearly 100 customers’ in 10 years isn’t very impressive
Each customer spending a billion dollars is impressive
They'll get less with a billion customers, unfortunately.
4 customers make up a majority of nvidias revenue
Lots of impressive companies have one customer (the government).
It is if each customer is paying for a million of them.
> And again no one can use it.
So basically Google has both SOTA models (Argon and Carbon and top of the line TPUs, but...
Google is in trouble?
(reading HN search is dead and hence Google is done)
> And again no one can use it.
Did you by chance fail to read who builds these things?
> While we’ve always co-designed our TPUs in partnership with Google DeepMind
I love it when companies expose their org charts publicly like this. As if anybody cares that they think about Google and DeepMind as different.
They were originally a separate company (and I think still based in the UK) so it's a bit more separate than probably random other internal departments
I’m guessing their skipping the T8 Pro model for the T9
/sarcasm at the unoriginal naming scheme
I'm hearing good things about the T1000. it's an ideal platform for highly survivable autonomous agents.
Typical Google. A literal headline “Opening up TPUs to every developer” but zero information or links on how I, a developer, can use said TPUs.
There are various bits of this blog post that detail how they’re making it easier for devs to use TPUs, particularly the discussion of TorchTPU and the attendant link. That being said if you’re looking for a link to a tutorial I think it’s easily findable on Google. “Tpu tutorial Google cloud” or “Tpu developer hub” would be good queries. For example [this](https://docs.cloud.google.com/compute/docs/tpus/quickstart-c...) is at least somewhat responsive.
I don’t know why everyone keeps complaining! It’s just a simple twenty-seven step process neatly spread across three portals! (Documentation coming soon.)
You’ll know sign up is complete when you started getting charged.
Well... Hopefully...
Or you'll get charged, but some obscure internal flag will be out of sync and you can't get ahold of anyone to reset it.
I am ready for the post-agentic era now... The agentic era turned out to be the most proclaimed era in the history... It's time for a change.
I can't take anyone who uses "agentic" seriously.
Agentic (seriously)
A product announcement page with 0 links to purchase or ways to purchase is useless.
So an announcement about a hypothetical highly effective new chemotherapy for glioblastoma, with no links for you to buy it yourself is useless ?
You are not the target market.
Then why is the page available for me to view?
You don't have to view it.
Continue to follow the analogy. When drug companies announce new chemotherapies or antibiotics, they still do press releases!