FlashAttention co-creator Dan Fu on kernels, inference economics, and owning vs leasing GPUs
realDanFu · x · 2026-11-11
The Compute podcast hosts Dan Fu, SVP of Research at Together AI and co-creator of FlashAttention. Topics covered:
- What CUDA, kernels, and compilers each do
- Why FlashAttention rethought data movement
- How AI is changing kernel engineering
- How new open models get to market and how inference revenue funds them
- Owning versus leasing GPUs
Chapters run from "From AI models to CUDA kernels" (1:04) to "Owning versus leasing compute" (51:17) — a deep systems-level conversation.
More from Infra
- Mintlify open-sources agent-traffic: agents made 489M doc requests, 77% of all traffic — handotdev · 2026-11-11
- Databricks makes Lakebase Search GA, benchmarks show it beating Pinecone at far lower cost — matei_zaharia · 2026-11-11
- xAI dumps JAX, rewrites full RL/inference stack in C/C++ as Grok 4.8 enters RL — ivan_bezdomny · 2026-11-11
- RidgeFill: training-free speculative prefill cuts TTFT up to 1.6x with near-baseline quality — Public_Umpire_1099 · 2026-11-11
- NVIDIA ships NCCL 2.32.6 with Vera Rubin tuning and new device APIs — simonguozirui · 2026-11-11
- Self-hosting agent sandboxes on AWS costs a quarter of Daytona's list price — KlausCodes · 2026-11-11