Cerebras deep dive: why CPUs, GPUs, TPUs and wafer-scale chips suit LLM workloads differently
sarahookr · x · 2026-11-21
Cerebras released a hardware deep-dive video with alyciazcary (who broke down Sarah Hooker's Hardware Lottery), walking through semiconductor fundamentals, SRAM vs DRAM, why CPUs struggle, how graphics gave GPUs their parallelism edge, training vs inference workloads, TPU systolic arrays, why wafer scale was hard to build and how Cerebras solved yield—keeping weights in SRAM beside compute to cut data movement during token generation. Full chapter timestamps included.
More from Infra
- Broadcom plunges 9% as Google reportedly cuts 1M AI chips from 2027 plans — PAstynome · 2026-11-21
- Commerce Dept approved 800,000+ advanced AI chip sales to Chinese firms, some military-linked — ylecun · 2026-11-21
- Codex raises WebSocket request limit 8x to 128MB, making large turns 5x faster — charliermarsh · 2026-11-21
- Neon boosts Antigravity cache hit rate ~4%, raising quota effectiveness by at least 20% — cgarciae88 · 2026-11-21
- Kapitsa runs Hy4-flash on M5 Max at 36 tok/s decode, 971 tok/s prefill at 32K — m4rkGS · 2026-11-21
- Epoch AI: trade data consistent with $3B+ of AI chips smuggled to China via Malaysia — Jsevillamol · 2026-11-21