Posted by hexagr 2 days ago
> Unless you've actually watched tokens stream at those rates, the numbers are hard to internalize. This is the rendering.
I built something similar recently, for the same reason: https://modal.com/llm-almanac/token-timing-simulator.
I like that the output rendering is closer to typical UIs -- syntax highlighting in code mode, tool calls, dim-italic reasoning.
One feature mine has that the author, or anyone else who vibe codes their own version after seeing this, might like to steal is modeling the distribution of output latencies. My implementation is hacky (log-normal roughyl estimated from p50, p90, and p99 values), but still, when you set those to realistic values, it recreates the "jitter" you see in many LLM UIs.
antirez is right that generation tok/s isn't flat as a function of context length, which is a weakness of both simulators.
On the other hand, I've been using Mimo and Minimax a lot recently. They routinely reach 100-150 tokens per second and that feels too fast, to the point where it's hard to keep up with what it's actually doing. Great for subagents though.
There is no way you can follow what is going on even at 30 tokens per second. Maybe you can maintain a rough idea of what is going on for some tens of seconds but that is probably about it. Follow it in any detail, no chance. Reason about what you read, absolutely no chance.
800 tok/s — Cerebras-class, where the bottleneck is your eyeballs
I do not understand why they say this. I am not sure if it is even true. 800 tokens sounds like a page of text and I would assume you can look at one page per second without hitting any limitation of your eyes. Or is the resolution of the human not good enough to see an entire page at once and you have to scan it with the fovea? Scrolling text might of course hit the temporal resolution limit. But why does this even matter, your brain can not process anything close to the amount of information your eyes can take in.
Click on 800.
Try to read the text.
You'll understand.
EDIT: As others have pointed out and I now did some reading on, it is an illusion that you can see all the text on a page at once, that is beyond the resolution limit of the human eye. To actually see all the words, you have to scan the page and that takes several seconds. From the numbers I have seen, it seems that the ultimate limit is probably below 30 tokens per second, no matter what, even using rapid serial visual presentation to cut out eye movements. Even 10 to 20 tokens per second is probably pushing it and unsustainable for many, if not most, people.
For non-trivial work I go through hundreds of thousands of tokens (combined prefill + tg of course) before even getting to some useful text output.
I mostly use LLMs for exploration and studies, rarely code generation. Prefill matters heavily for this. Even in the high hundreds or low thousands prefill rate I spend a lot of time waiting on the LLM (doing other things, not twiddling thumbs)
1. decoding t/s, that is, when the model is generating text in the autoregressive fashion.
2. prefill t/s, that is, prompt processing speed.
3. What is the slope of those two numbers as the context size increases. An implementation that decodes at 50t/s with 2k context but decodes at 7t/s at 100k context is going to be a lot less useful that it seems at a first glance for a big number of real world use cases.
4. What's your use case? Reading a huge text and then having a small output like, fraud probability=12%? Or Reading a small question and generating a lot of text? This changes substantially if a model is usable based on its prefill/decoding speed.
For instance my DS4F inference on the DGX Spark does prefill at 350 t/s and at 200 t/s on already large contexts. But decodes at 13 t/s.
On the Mac Ultra the prefill is like 400 t/s and decoding 35 t/s.
The two systems can perform dramatically differently or almost the same based on the use case. In general for local inference to be acceptable, even if slow, you want at least 100 t/s prefill, at least 10 t/s generation. To be ok-ish from 200 to 400 t/s prefill, 15-25 t/s generation. To be a wonderful experience thousands of t/s prefill, 100 t/s generation.
You should run a multi-session batched decode on that DGX unless your 13 t/s decode is already running into thermal or power limits, which I don't believe it is. (To be clear, this is a real issue on Apple Silicon machines: batched decode does not seem to unlock higher aggregate tok/s unless you're specifically trying to mitigate the drawbacks of slow streamed inference. Especially on the M5 laptops, thermal/power throttling places an early limit on your total compute.
The jury is still out on Strix Halo, but I think batched decode may turn out to be quite useful there since the bandwidth bottleneck is even more constraining there.)