Within a year of announcement. June 25, 2025.
The Sohu chip is an AI ASIC developed by Etched that is claiming 20x faster transformer inference than an Nvidia H100.
https://www.etched.com/
Their announcement only had renders so I don't think they have any chips made yet and I think the performance numbers are theoretical. But they just raised $120M in funding.
I won't bet in this market.
I think it’s quite unlikely that they release a chip that’s dramatically smaller, or where Llama 3 performance is off by more than 2-3x.
(But Llama3 is a best case scenario for the architecture, it definitely won’t be 20x H100 performance when running MoE or long context models because those make batching less effective and Sohu will be memory bandwidth limited. If anyone can run the numbers for how much slower MoE and long context might make it, I’m curious to see it.)
The reported Sohu specs are that it’s on the same process node as B200, but 1/2 the die size and 3/4 the DRAM capacity and bandwidth, and they’re claiming ~10x the performance.
Let me read on this real quick, comb through their claims, and by Saturday I will let you know if I'm down for it, but if this is as simple as it seems I'd be willing to commit 7500 to reasonable limit krder
Not committing yet but if this exchange is appetizing to you I might be willing.
I would also limit order anyone's shares who felt jipped by the change
Per the website: "Etched raises 120M to build Sohu" seems to imply they don't have a functional product yet. If the benchmarks they put out are even close to accurate then it is a complete gamechanger, so I'm going to assume they are very fudged and will probably delay shipment due to inability to get close to them