Carmack: Replace HBM with NAND Flash to Cut Inference Costs
ID_AA_Carmack · x · 2026-07-07
John Carmack pointed out that memory cost and capacity are the primary bottlenecks for AI accelerators. Unlike game rendering, model inference features deterministic memory access patterns—model weights don't require true random access; they just need to be read continuously at the required bandwidth, and can even tolerate cold start latencies of several milliseconds. Because NAND flash is over a hundred times cheaper per GB than HBM, there is significant room for cost optimization worth exploring with dedicated pin protocols, even if the flash controller is equipped with a 1024-bit interface to achieve HBM-level bandwidth.
Related event: Carmack Suggests Using NAND Flash Over HBM for AI Inference(2 posts)→
More from Infra
- SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
- SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
- Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11