Unlocking hidden RTX 4070 Ti SUPER compute: FP16 matmul gets 1.9x faster
lxfater · x · 2026-10-24
A hands-on article tests the open-source gpu-unlock project on Ubuntu with an RTX 4070 Ti SUPER (16GB).
- How it works: the unlock module modifies tensor-core speed-limit settings before the NVIDIA driver loads, then the stock driver takes over; mainly unlocks FP16/BF16-input, FP32-accumulate Tensor Core paths
- Results (8192-dim matmul, 285W cap): FP16→FP32 accumulate throughput rose from 90.47 to 169.36 TFLOPS, per-run time 12.15ms→6.49ms (47% faster); BF16 hit 170.23 TFLOPS; TF32 went 45.53→88.73; strict FP32 control was unchanged (34.13 vs 34.12)
- Correctness: 65536-cell checks passed with failures=0, maxabserror=0
- Compatibility: covers some RTX 30-series (e.g. 3090), RTX 4090, experimental 4070 Ti SUPER; plain 4070/4080 lines and RTX 20/50 not supported; Linux only, match exact device IDs
- Caveat: gains apply to matmul throughput, not full inference/image-gen/game speed
More from Infra
- Agent-assisted self-hosting now takes days, not months, dev argues on X — kuza55 · 2026-10-24
- AI capex debt charts echo 1980s credit boom — when does the party end? — SumitGup · 2026-10-24
- Architect pushes back on 'AWS sold out through 2027' claims as hype — DavidLinthicum · 2026-10-24
- Homemade animation walks through TSMC's CoWoS-S advanced packaging flow, bugs admitted — jwt0625 · 2026-10-24
- Muon optimizer may nearly halve LLM training compute, challenging a decade of AdamW — bycloud · 2026-10-24
- Infographic: why AI is an energy problem, from computation and cooling to the grid — CurieuxExplorer · 2026-10-24