AI

GPU unlocks (FP32 accumulation): Why can it be unlocked on Turing (RTX 20), Ampere (RTX 30) and Ada (RTX 40) but no consumer Blackwell (RTX 50) and some misconceptions.

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 27/10/2026, 10:47:26 UTC.

Briefings

A 2026 analysis reveals that NVIDIA unlocks FP32 accumulation performance on Turing, Ampere, and Ada GPUs via a firmware exploit allowing full-speed tensor cores, while Blackwell GPUs lack this capability due to a hardware root-of-trust change. Key metrics show 4090s gain 165 to 331 TFLOPS in FP16/BF16 with FP32 accumulation, with real-world gains capped at 2x due to power limits. The unlock persists across reboots but is lost on power-off, affecting training and diffusion tasks significantly.

Key points

• NVIDIA unlocks FP32 accumulation on Turing, Ampere, and Ada GPUs by exploiting a buffer overflow in the Booter firmware, allowing full-speed tensor cores. On the 4090, this increases FP16/BF16 with FP32 accumulation from 165 to 331 TFLOPS. Real-world gains are limited to 2x due to power constraints.
• The unlock mechanism relies on a Booter firmware that runs in the highest privilege mode (HS) to override fuses; this same mechanism is absent in Blackwell GPUs where the FSP acts as the root of trust, preventing the same exploit from being used.

Context

This briefing details a 2026 finding where NVIDIA's Booter firmware on Turing, Ampere, and Ada GPUs allows a buffer overflow to unlock FP32 accumulation performance, whereas Blackwell GPUs use a hardware root-of-trust (FSP) that blocks this exploit. The analysis covers specific performance gains and limitations, with no additional background context provided.

Limitations

Evidence is limited to a single article from October 2026; real-world performance gains are capped at 2x due to power limits. The analysis does not account for future hardware changes or alternative vulnerabilities, and the specific Blackwell firmware version used for testing is not specified.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 27/10/2026, 10:50:18 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeidw4j4eo5yfy46ocw6mtnlhgllqkiri2ywmqtrgsvw5yggjlgyc4y