This math is useful. Lots of folks scoffing in the comments below. I have a couple reactions, after chatting with it:

1) 16k tokens / second is really stunningly fast. There’s an old saying about any factor of 10 being a new science / new product category, etc. This is a new product category in my mind, or it could be. It would be incredibly useful for voice agent applications, realtime loops, realtime video generation, .. etc.

2) https://nvidia.github.io/TensorRT-LLM/blogs/H200launch.html Has H200 doing 12k tokens/second on llama 2 12b fb8. Knowing these architectures that’s likely a 100+ ish batched run, meaning time to first token is almost certainly slower than taalas. Probably much slower, since Taalas is like milliseconds.

3) Jensen has these pareto curve graphs — for a certain amount of energy and a certain chip architecture, choose your point on the curve to trade off throughput vs latency. My quick math is that these probably do not shift the curve. The 6nm process vs 4nm process is likely 30-40% bigger, draws that much more power, etc; if we look at the numbers they give and extrapolate to an fp8 model (slower), smaller geometry (30% faster and lower power) and compare 16k tokens/second for taalas to 12k tokens/s for an h200, these chips are in the same ballpark curve.

However, I don’t think the H200 can reach into this part of the curve, and that does make these somewhat interesting. In fact even if you had a full datacenter of H200s already running your model, you’d probably buy a bunch of these to do speculative decoding - it’s an amazing use case for them; speculative decoding relies on smaller distillations or quants to get the first N tokens sorted, only when the big model and small model diverge do you infer on the big model.

Upshot - I think these will sell, even on 6nm process, and the first thing I’d sell them to do is speculative decoding for bread and butter frontier models. The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.

I hope these guys make it! I bet the v3 of these chips will be serving some bread and butter API requests, which will be awesome.

> any factor of 10 being a new science / new product category,

I often remind people two orders of quantitative change is a qualitative change.

> The thing that I’m really very skeptical of is the 2 month turnaround. To get leading edge geometry turned around on arbitrary 2 month schedules is .. ambitious. Hopeful. We could use other words as well.

The real product they have is automation. They figured out a way to compile a large model into a circuit. That's, in itself, pretty impressive. If they can do this, they can also compile models to an HDL and deploy them to large FPGA simulators for quick validation. If we see models maturing at a "good enough" state, even a longer turnaround between model release and silicon makes sense.

While I also see lots of these systems running standalone, I think they'll really shine combined with more flexible inference engines, running the unchanging parts of the model while the coupled inference engine deals with whatever is too new to have been baked into silicon.

I'm concerned with the environmental impact. Chip manufacture is not very clean and these chips will need to be swapped out and replaced at a cadence higher than we currently do with GPUs.

I think the next major innovation is going to be intelligent model routing. I've been exploring OpenClaw and OpenRouter, and there is a real lack of options to select the best model for the job and execute. The providers are trying to do that with their own models, but none of them offer everything to everyone at all times. I see a future with increasingly niche models being offered for all kinds of novel use cases. We need a way to fluidly apply the right model for the job.

For speculative decoding, wouldn’t this be of limited use for frontier models that don’t have the same tokenizer as Llama 3.1? Or would it be so good that retokenization/bridging would be worth it?

The guts of a LLM isn't something I'm well versed in, but

> to get the first N tokens sorted, only when the big model and small model diverge do you infer on the big model

suggests there is something I'm unaware of. If you compare the small and big model, don't you have to wait for the big model anyway and then what's the point? I assume I'm missing some detail here, but what?

Think about this for solving questions in math where you need to explore a search space. You can run 100 of these for the same cost and time of doing one api call to open ai.