DeepSeek may unlock consumer GPU matmul performance with memory-bandwidth-first architecture
teortaxesTex · x · 2026-10-25
Commenter teortaxesTex speculates DeepSeek would have little trouble unlocking peak matmuls on RTX 3090/4090s, citing rumors of an internal single-consumer-GPU model, DeepSeek's memory-bandwidth-cheapest architecture, and experiment backbones under 22B. The thread builds on the argument that DeepSeek's core principle — compute is easy to scale, memory is hard — drives its model and infrastructure design, and its reported 80% gross margin economics.
More from Infra
- ~1M unlocked RTX 3090s in China ≈ 140 exaFLOPS BF16, rivaling ~140k H100s for inference — teortaxesTex · 2026-10-25
- RTX 2080 Ti 22GB mod joins Ampere unlock: 1.85x tensor throughput, open-sourced — bdsqlsz · 2026-10-25
- Brex data: agent compute rentals surge, Baseten tops fall 2026 fastest-growing vendors — FinanceYF5 · 2026-10-25
- SpaceX monetizes ~$50B per GW of AI compute with 1-year payback vs CoreWeave's $20B — elonmusk · 2026-10-25
- DeepSeek insider: infra is the big brother and can veto research decisions — teortaxesTex · 2026-10-25
- DeepInfer is deeply tailored to DeepSeek 4.1F: KV cache of just 890 bytes per token — teortaxesTex · 2026-10-25