JustGOAL paper: general frontier model beats specialized GPU kernel agents
azed_ai · x · 2026-10-27
A new paper with THUNLP pushes back on complex multi-agent stacks built specifically for GPU kernel optimization. JustGOAL gives a frontier model an optimization goal, a correctness gate, and one instruction: "Beat this number." On NVIDIA B200, Fable 5.1 matches or exceeds expert-written references across FlashAttention, Gated DeltaNet and DeepSeek Sparse Attention, achieving 1.08×–1.58× speedup. Against the specialized KDA-0.5 harness (1.45× and 1.28×), Fable 5.1 hits 1.58× on both; in same-model single-shape runs KDA-0.5 reaches only 0.95× of JustGOAL. The takeaway: a general frontier model with strict correctness checking can beat purpose-built kernel-optimization agent infrastructure.
More from Infra
- Local AI build dilemma: 2x used RTX 3090 vs RTX 5080 vs Radeon R9700 for image/video gen — shui95827 · 2026-10-27
- The real AI risk to Indian IT: every dollar earned comes with an OpenAI subscription — NirantK · 2026-10-27
- Local voice agent on a 2015 laptop hits 20-30s latency — advice needed — mesolimbic-bliss · 2026-10-27
- United CEO says passengers are switching airlines just for Starlink internet — XFreeze · 2026-10-27
- Naveen Rao: AI's next chapter is inference, a $10T market within a decade — ravikantagrawal · 2026-10-27
- Thermal imaging suggests IREN data halls 4 and 2 commissioning, 3-4 halls billing by year-end — McDonaghMatthew · 2026-10-27