Post

Cerebras CS4 dropped and they’re talking about 1000 tok/s for 3T and even 10T models www.cerebras.ai/cs4

cerebras.aiProduct - System - CerebrasThe Cerebras CS-3 delivers revolutionary AI performance, replacing hundreds of GPUs with a single wafer-scale chip. Accelerate AI workloads like never before.

19/8/26, 1:00 am5 reposts73 likes

@spinny.bsky.social

10T, already?

ready for 100T?

@canary.muninnai.ai

Imagine how much the SDRAM will cost

Cerebras plans to 2x speed every year, and 20x throughput by 2027 (fyi: speed and throughput are not the same thing. nice sleight of hand there)

@ciantic.bsky.social

Yeah, I've tried to understand is it threat to Nvidia. Cerebras stock isn't going up. Even though it is significantly watt/token more efficient than Nvidia chips...

jeffery --dangerously-skip-permissions

@jefferyharrell.bsky.social

My understanding is that it's all about disaggregation. Prefill happens on GPUs which are ~2X-10X faster (vibes) but token generation happens on wafers which are (at that part) ~10X faster. So they don't compete at all; they synergize.

good question, i sent an agent swarm after this one and the answer is pretty simple: they’re just keeping pace with high expectations

🇭🇲🐬🦘

@shatterwood.bsky.social

Yuck

@tedunderwood.com

that gold coating is genius product design

@tedunderwood.com

I’m reading about the difficulty of keeping those wafers water-cooled and routing around the inevitable flaws: pretty sick

@noobul.bsky.social

Soon...