Cerebras CS-4: 750 petaflops, and a sparsity footnote

Cerebras launched the CS-4 on August 19, its first multi-wafer rack system, with a headline of 750 petaflops of AI compute and a claim of up to 30x faster inference than GPU-based systems. One of those numbers holds up under inspection and one needs a footnote. If you're buying AI throughput — and if you run an agent a customer waits on, you are — knowing which is which is worth more than the press release.

What actually happened

The CS-4 packs up to three WSE-3T accelerators, a Turbo variant of the existing WSE-3, into a single rack on Cerebras' new Nexus platform architecture. Per-unit, The Register reports each WSE-3T doubles its predecessor across the board: 250 PFLOPS sparse FP16, 43.2 PB/s of memory bandwidth, 2.4 Tbps of I/O. Three of them gets you the 750 PFLOPS / 129.6 PB/s / 7.2 Tbps in the headline.

Here's the footnote. That 250 PFLOPS per wafer is sparse FP16, assuming roughly 10x sparsity. Dense — which is what LLM inference actually runs — is closer to 25 PFLOPS per wafer, so 75 across the rack. The Register also flags peak memory bandwidth as likely theoretical under real inference load. The number isn't fabricated; it's just measuring a workload most people aren't running.

The benchmark is the part that's genuinely interesting. On gpt-oss-120b, a single CS-4 hit up to 4,400 tokens per second per user, against roughly 350 tok/s on the fastest GPU-based inference service. That's ~12.5x on a specific open-weight model — short of the marketing 30x, and still an enormous number. The other structural change: Direct Wafer Links connect wafers within and across racks without a switch, cutting wafer-to-wafer latency from five microseconds to two, which is what makes the multi-wafer topology worth building. First systems come online later this quarter.

Why inference speed matters for your business

You will never buy one of these. You will buy tokens served on one, and the reason to track it is that inference speed has decoupled from the model. Same weights, wildly different wall-clock, depending entirely on the silicon underneath.

That changes how you spec work. For a batch job — reconciling last month's transactions overnight — tokens per second is noise. For anything a human sits and waits on, it's the whole product: a support agent that answers in under a second feels built, and one that spins for twelve feels broken. Latency isn't a technical detail, it's a conversion metric.

So do what the vendors won't. Benchmark your own workload, on your own prompt lengths, measuring wall-clock from request to last token — not the vendor's tok/s on a 120B open model with a short prompt. Time-to-first-token and end-to-end latency are what your customer experiences; peak throughput on a synthetic benchmark is what a slide deck experiences. And keep the routing layer swappable, because the fastest option in Q4 will not be the fastest option next year.

Sources: The Register, HPCwire.