Posts
- Inference Is Memory-Bound. That Is What Custom Silicon Buys.
Serving is sold by the token under latency SLOs, so bandwidth and watts beat peak FLOPS—and a decade of production ASICs plus 120 kW racks show the facility changing with the chip.
Serving is sold by the token under latency SLOs, so bandwidth and watts beat peak FLOPS—and a decade of production ASICs plus 120 kW racks show the facility changing with the chip.