Comment by canyon289
7 days ago
I work at Google on these systems everyday (caveat this is my own words not my employers)). So I simultaneously can tell you that its smart people really thinking about every facet of the problem, and I can't tell you much more than that.
However I can share this written by my colleagues! You'll find great explanations about accelerator architectures and the considerations made to make things fast.
https://jax-ml.github.io/scaling-book/
In particular your questions are around inference which is the focus of this chapter https://jax-ml.github.io/scaling-book/inference/
Edit: Another great resource to look at is the unsloth guides. These folks are incredibly good at getting deep into various models and finding optimizations, and they're very good at writing it up. Here's the Gemma 3n guide, and you'll find others as well.
https://docs.unsloth.ai/basics/gemma-3n-how-to-run-and-fine-...
Same explanation but with less mysticism:
Inference is (mostly) stateless. So unlike training where you need to have memory coherence over something like 100k machines and somehow avoid the certainty of machine failure, you just need to route mostly small amounts of data to a bunch of big machines.
I don't know what the specs of their inference machines are, but where I worked the machines research used were all 8gpu monsters. so long as your model fitted in (combined) vram, you could job was a goodun.
To scale the secret ingredient was industrial amounts of cash. Sure we had DGXs (fun fact, nvidia sent literal gold plated DGX machines) but they wernt dense, and were very expensive.
Most large companies have robust RPC, and orchestration, which means the hard part isn't routing the message, its making the model fit in the boxes you have. (thats not my area of expertise though)
> Inference is (mostly) stateless. ... you just need to route mostly small amounts of data to a bunch of big machines.
I think this might just be the key insight. The key advantage of doing batched inference at a huge scale is that once you maximize parallelism and sharding, your model parameters and the memory bandwidth associated with them are essentially free (since at any given moment they're being shared among a huge amount of requests!), you "only" pay for the request-specific raw compute and the memory storage+bandwidth for the activations. And the proprietary models are now huge, highly-quantized extreme-MoE models where the former factor (model size) is huge and the latter (request-specific compute) has been correspondingly minimized - and where it hasn't, you're definitely paying "pro" pricing for it. I think this goes a long way towards explaining how inference at scale can work better than locally.
(There are "tricks" you could do locally to try and compete with this setup, such as storing model parameters on disk and accessing them via mmap, at least when doing token gen on CPU. But of course you're paying for that with increased latency, which you may or may not be okay with in that context.)
> The key advantage of doing batched inference at a huge scale is that once you maximize parallelism and sharding, your model parameters and the memory bandwidth associated with them are essentially free (since at any given moment they're being shared among a huge amount of requests!)
Kind of unrelated, but this comment made me wonder when we will start seeing side channel attacks that force queries to leak into each other.
1 reply →
mmap is not free. It just moves bandwidth around.
1 reply →
> Inference is (mostly) stateless
Quite the opposite. Context caching requires state (K/V cache) close to the VRAM. Streaming requires state. Constrained decoding (known as Structured Outputs) also requires state.
> Quite the opposite.
Unless something has dramatically changed, the model is stateless. The context cache needs to be injected before the new prompt, but for what I understand (and please do correct me if I'm wrong) the the context cache isn't that big, like in the order of a few tens of kilobytes. Plus the cache saves seconds of GPU time, so having an extra 100ms of latency is nothing compare to a cache miss. so a broad cache is much much better than a narrow local cache.
But! even if its larger, Your bottleneck isn't the network, its waiting on the GPUs to be free[1]. So whilst having the cache really close ie in the same rack, or same machine, will give the best performance, it will limit your scale (because the cache is only effective for a small number of users)
[1] a 100megs of data shared over the same datacentre network every 2-3 seconds per node isn't that much, especially if you have a partitioned network (ie like AWS where you have a block network and a "network" network)
1 reply →
[dead]
> So I simultaneously can tell you that its smart people really thinking about every facet of the problem, and I can't tell you much more than that.
"we do 1970s mainframe style timesharing"
there, that was easy
For real. Say it takes 1 machine 5 seconds to reply, and that a machine can only possibly form 1 reply at a time (which I doubt, but for argument).
If the requests were regularly spaced, and they certainly won’t be, but for the sake of argument, then 1 machine could serve 17,000 requests per day, or 120,000 per week. At that rate, you’d need about 5,600 machines to serve 700M requests. That’s a lot to me, but not to someone who owns a data center.
Yes, those 700M users will issue more than 1 query per week and they won’t be evenly spaced. However, I’d bet most of those queries will take well under 1 second to answer, and I’d also bet each machine can handle more than one at a time.
It’s a large problem, to be sure, but that seems tractable.
Yes. And batched inference is a thing, where intelligent grouping/bin packing and routing of requests happens. I expect a good amount of "secret sauce" is at this layer.
Here's an entry-level link I found quickly on Google, OP: https://medium.com/@wearegap/a-brief-introduction-to-optimiz...
But that’s not accurate. There are all sorts of tricks around KV cache where different users will have the same first X bytes because they share system prompts, caching entire inputs / outputs when the context and user data is identical, and more.
Not sure if you were just joking or really believe that, but for other peoples’ sake, it’s wildly wrong.
Really? So the system recognises someone asked the same question and serves the same answer? And who on earth shares the exact same context?
I mean i get the idea but sounds so incredibly rare it would mean absolutely nothing optimisation wise.
3 replies →
I'm pretty sure that's not right.
They're definitely running cluster knoppix.
:-)
Makes perfect sense, completely understand now!
I don't think it's either useful or particularly accurate to characterize modern disagg racks of inference gear, well-understood RDMA and other low-overhead networking techniques, aggressive MLA and related cache optimizations that are in the literature, and all the other stuff that goes into a system like this as being some kind of mystical thing attended to by a priesthood of people from a different tier of hacker.
This stuff is well understood in public, and where a big name has something highly custom going on? Often as not it's a liability around attachment to some legacy thing. You run this stuff at scale by having the correct institutions and processes in place that it takes to run any big non-trivial system: that's everything from procurement and SRE training to the RTL on the new TPU, and all of the stuff is interesting, but if anyone was 10x out in front of everyone else? You'd be able to tell.
Signed, Someone Who Also Did Megascale Inference for a TOP-5 For a Decade.
Doesn't google have TPU's that makes inference of their own models much more profitable than say having to rent out NVDIA cards?
Doesn't OpenAI depend mostly on its relationship/partnership with Microsoft to get GPUs to inference on?
Thanks for the links, interesting book!
Yes. Google is probably gonna win the LLM game tbh. They had a massive head start with TPUs which are very energy efficient compared to Nvidia Cards.
The only one who can stop Google is Google.
They’ll definitely have the best model, but there is a chance they will f*up the product / integration into their products.
14 replies →
Google will win the LLM game if the LLM game is about compute, which is the common wisdom and maybe true, but not foreordained by God. There's an argument that if compute was the dominant term that Google would never have been anything but leading by a lot.
Personally right now I see one clear leader and one group going 0-99 like a five sigma cosmic ray: Anthropic and the PRC. But this is because I believe/know that all the benchmarks are gamed as hell, its like asking if a movie star had cosmetic surgery. On quality, Opus 4 is 15x the cost and sold out / backordered. Qwen 3 is arguably in next place.
In both of those cases, extreme quality expert labeling at scale (assisted by the tool) seems to be the secret sauce.
Which is how it would play out if history is any guide: when compute as a scaling lever starts to flatten, you expert label like its 1987 and claim its compute and algorithms until the government wises up and stops treating your success persobally as a national security priority. It's the easiest trillion Xi Xianping ever made: pretending to think LLMs are AGI too, fast following for pennies on the dollar, and propping up a stock market bubble to go with the fentanyl crisis? 9-D chess. It's what I would do about AI if I were China.
Time will tell.
3 replies →
Yeah honestly. They could just try selling solutions and SLAs combining their TPU hardware with on-prem SOTA models and practically dominate enterprise. From what I understand, that's GCP's gameplay too for most regulated enterprise clients.
9 replies →
Hasn’t the Inferentia chip been around long enough to make the same argument? AWS and Google probably have the same order of magnitude of their own custom chips
1 reply →
But they’re ASICs so any big architecture changes will be painful for them right?
4 replies →
Im a research person building models so I can't answer your questions well (save for one part)
That is, as a research person using our GPUs and TPUs I see first hand how choices from the high level python level, through Jax, down to the TPU architecture all work together to make training and inference efficient. You can see a bit of that in the gif on the front page of the book. https://jax-ml.github.io/scaling-book/
I also see how sometimes bad choices by me can make things inefficient. Luckily for me if my code/models are running slow I can ping colleagues who are able to debug at both a depth and speed that is quite incredible.
And because were on HN I want to preemptively call out my positive bias for Google! It's a privilege to be able to see all this technology first hand, work with great people, and do my best to ship this at scale across the globe.
> Another great resource to look at is the unsloth guides.
And folks at LMSys: https://lmsys.org/blog/
This caught my attention "But today even “small” models run so close to hardware limits".
Sounds analogous to the 60's and 70's i.e "even small programs run so close to hardware limits". If optimization and efficiency is dead in software engineering, it's certainly alive and well in LLM development.
Why does the unsloth guide for gemma 3n say:
> llama.cpp an other inference engines auto add a <bos> - DO NOT add TWO <bos> tokens! You should ignore the <bos> when prompting the model!
That makes the want to try exactly that? Weird
Nothing smart about making something that is not useful for humans.
No, you just over complicate things.
If people at google are so smart why can't google.com get a 100% lighthouse score?
I have met a lot of people at Google, they have some really good engineers and mediocre ones. But mostl importantly they are just normal engineers dealing normal office politics.
I don't like how the grand parent mystifies this. This problem is just normal engineering. Any good engineer could learn how to do it.
Because most smart people are not generalists. My first boss was really smart and managed to found a university institute in computer science. The 3 other professors he hired were, ahem, strange choices. We 28 year old assistents could only shake our heads. After fighting a couple of years with his own hires the founder left in frustration to found another institution.
One of my colleagues was only 25, really smart in his field and became a professor less than 10 years later. But he was incredibly naive in everyday chores. Buying groceries or filing taxes resulted in major screw-ups regularly
I have met those supersmart specialists but in my experience there are also a lot of smart people who are more generalists.
The real answer is likely internal company politics and priorities. Google certainly has people with the technical skills to solve it but do they care and if they care can they allocate those skilled people to the task?
2 replies →
Pro-tip they're just not. A lot of tech nerds really like to think they're a genius with all the answers ("why don't they just do XX"), but some eventually learn that the world is not so black and white.
The Dunning-Kruger effect also applies to smart people. You don't stop when you are estimating your ability correctly. As you learn more, you gain more awareness of your ignorance and continue being conservative with your self estimates.
A lot of really smart people working on problems that don't even really need to be solved is an interesting aspect of market allocation.
Can you explain what you mean about 'not needing to be solved'? There are versions of that kind of critique that would seem, at least on the surface, to better apply to finance or flash trading.
I ask because scaling an system that a substantially chunk of the population finds incredibly useful, including for the more efficient production of public goods (scientific research, for example) does seem like a problem that a) needs to be solved from a business point of view, and b) should be solved from a civic-minded point of view.
I think the problem I see with this type of response is that it doesn't take into context the waste of resources involved. If the 700M users per week is legitimate then my question to you is: how many of those invocations are worth the cost of resources that are spent, in the name of things that are truly productive?
And if AI was truly the holy grail that it's being sold as then there wouldn't be 700M users per week wasting all of these resources as heavily as we are because generative AI would have already solved for something better. It really does seem like these platforms are, and won't be, anywhere as useful as they're continuously claimed to be.
Just like Tesla FSD, we keep hearing about a "breakaway" model and the broken record of AGI. Instead of getting anything exceptionally better we seem to be getting models tuned for benchmarks and only marginal improvements.
I really try to limit what I'm using an LLM for these days. And not simply because of the resource pigs they are, but because it's also often a time sink. I spent an hour today testing out GPT-5 and asking it about a specific problem I was solving for using only 2 well documented technologies. After that hour it had hallucinated about a half dozen assumptions that were completely incorrect. One so obvious that I couldn't understand how it had gotten it so wrong. This particular technology, by default, consumes raw SSE. But GPT-5, even after telling it that it was wrong, continued to give me examples that were in a lot of ways worse and kept resorting to telling me to validate my server responses were JSON formatted in a particularly odd way.
Instead of continuing to waste my time correcting the model I just went back to reading the docs and GitHub issues to figure out the problem I was solving for. And that led me down a dark chain of thought: so what happens when the "teaching" mode rethinks history, or math fundamentals?
I'm sure a lot of people think ChatGPT is incredibly useful. And a lot of people are bought into not wanting to miss the boat, especially those who don't have any clue to how it works and what it takes to execute any given prompt. I actually think LLMs have a trajectory that will be similar to social media. The curve is different and I, hopefully, don't think we've seen the most useful aspects of it come to fruition as of yet. But I do think that if OpenAI is serving 700M users per week then, once again, we are the product. Because if AI could actually displace workers en masse today you wouldn't have access to it for $20/month. And they wouldn't offer it to you at 50% off for the next 3 months when you go to hit the cancel button. In fact, if it could do most of the things executives are claiming then you wouldn't have access to it at all. But, again, the users are the product - in very much the same way social media played into.
Finally, I'd surmise that of those 700M weekly users less than 10% of those sessions are being used for anything productive that you've mentioned and I'd place a high wager that the 10% is wildly conservative. I could be wrong, but again - we'd know about that if it were the actual truth.
10 replies →
[flagged]
11 replies →
They won’t be honest and explain it to you but I will. Takes like the one you’re responding to are from loathsome pessimistic anti-llm people that are so far detached from reality they can just confidently assert things that have no bearing on truth or evidence. It’s a coping mechanism and it’s basically a prolific mental illness at this point
2 replies →
> working on problems that don't even really need to be solved
Very, very few problems _need_ to be solved. Feeding yourself is a problem that needs to be solved in order for you to continue living. People solve problems for different reasons. If you don't think LLMs are valuable, you can just say that.
The few problems humanity has that need to be solved:
1. How to identify humanity's needs on all levels, including cosmic ones...(we're in the Space Age so we need to prepare ourselves for meeting beings from other places)
2. How to meet all of humanity's needs
Pointing this out regularly is probably necessary because the issue isn't why people are choosing what they're doing...it's that our systems actively disincentivize collectibely addressing these two problems in a way that doesn't sacrifice people's wellbeing/lives... and most people don't even think about it like this.
The notion that simply pretending to not understand that I was making a value judgment about worth is an argument is tiring.
Well, we all thought advertising was the worst thing to come out of the tech industry, someone had to prove us wrong!
Just wait until the two combine.