Hidden FLOPs on consumer cards, and the GPU-hour getting dearer

Sources:

Zagitta/gpu-unlock (GitHub README, v1.0.0), surfaced on the X Following feed via @lxfater
DeepInfra pricing email, received 2026-10-25 (raw/gmail/2026-10-26-newsletters.md), pricing page
The Information: data center late fees hit $5M+ per day (raw/rss/)
AI Weekly, 2026-10-25 (CNBC figures on Nvidia's neocloud exposure and CoreWeave credit default swaps)
X feed: Nscale filing and lead times, DRAM/substrate lead times, CoreWeave funding gap, Musk on neocloud debt Topic: GPU hardware, compute economics

TL;DR

Two stories about the price of a FLOP. gpu-unlock is a Linux kernel module that runs before the NVIDIA driver binds an Ampere GeForce card (GA102/103/104/106), uses a known firmware-unlock technique to write the protected "speed selector" registers, then hands the card to the unmodified driver. The effect is to double the issue rate of FP16/BF16-input, FP32-accumulate tensor instructions, a rate NVIDIA keeps at half on GeForce versus the professional parts built on the same silicon. On a 3090 Ti at 450 W, dense BF16 GEMM goes from 87.5 to 157.4 TFLOPS and TF32 from 42.6 to 82.6; user reports on 3090, 3080, 3070 and 3060 show +25% to +81%, lowest on power-capped cards. Real workloads gain less: LLM prefill +14-36%, while decode (memory-bandwidth bound) and INT8 do not change. Outputs were bit-identical before and after. Meanwhile, rented compute is getting more expensive. DeepInfra raises GPU-hour rates on November 9: A100 80GB $0.89 to $1.49 (+67%), H100 $2.20 to $2.90, H200 $2.69 to $3.49, B200 $3.69 to $4.89, B300 $4.89 to $5.89, citing industry-wide hardware cost increases, while serverless per-token prices stay flat. The constraint is upstream of the GPU: Nscale's IPO filing shows about 25,000 active GPUs against 461,000 active and contracted, transformer lead times exceed 160 weeks, and data center tenants now demand $5M+ per day in late-delivery penalties.

Same silicon, two prices: a firmware limit and a power limit

One FLOP ceiling is set by product segmentation, the other by the grid.

flowchart LR
  G["GA102 silicon<br/><small>same die, pro and GeForce</small>"] --> L["Firmware rate cap<br/><small>half-rate BF16 into FP32</small>"]
  L -->|gpu-unlock| U["Full issue rate<br/><small>87 to 157 TFLOPS BF16</small>"]
  U --> P["Prefill<br/><small>+14-36%</small>"]
  U --> D["Decode<br/><small>unchanged, bandwidth bound</small>"]
  H["Datacenter GPUs<br/><small>ship in months</small>"] --> W["Power and gear<br/><small>160+ week transformers</small>"]
  W -->|scarce| R["GPU-hour rent<br/><small>A100 +67% at DeepInfra</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class G,H input
  class L,W loop
  class U,P exit
  class D,R err
  linkStyle 1 stroke:#2f9e44,stroke-width:2px
  linkStyle 6 stroke:#e03131,stroke-width:2px

Blue is the hardware, amber is the cap that binds, green is the unlocked throughput, red is what does not improve or gets dearer.

Key findings

The cap is an issue-rate limit, not a clock limit. gpu-unlock writes SS0 = 0x88888888 (full BF16 rate) and the 4-bit SS1 selector, performs a function-level reset, and lets the stock driver bind. It is opt-in via the kernel command line and needs no driver patch per NVIDIA release.
Gains follow the roofline. Compute-bound kernels (GEMM, prefill, diffusion, FlashAttention) speed up; bandwidth-bound decode does not. A 4090 fork claims 165 to 331 TFLOPS of instruction throughput; that one is secondhand and unverified.
Rent is rising at the old end too. The largest DeepInfra increase is on the A100, the oldest card. Older GPUs are not getting cheaper as new ones arrive, because the scarce input is energized rack space.
Neocloud credit is under stress. CNBC (via AI Weekly): Nvidia has $7B invested in AI cloud providers plus $164.5B in guarantees and commitments ($171.5B total); five-year CoreWeave default insurance costs about 1,082 basis points, above the ~1,000 usually read as distress. Musk says debt markets have "almost entirely shut off" for AI-only neoclouds.
Supply lead times. DRAM, SSDs, hard drives and ABF substrate run 2-5x normal lead times (ABF near a year vs three months), so GPUs are the only balanced part of the chain.

Relation to prior wiki pages

Extends the "memory sets the rent" result (10-24), where H200 rent rose 160% vs 40% for the H100 because buyers pay for memory, not FLOPs. gpu-unlock is the mirror case: doubling FLOPs on a 3090 leaves decode untouched because decode is memory-bound. Both say bandwidth, not compute, prices inference.
Confirms the operator gap (10-23, SemiAnalysis): about 50% output differences on identical B200s from stack choices alone. Here, 80% of GEMM throughput on a 3090 was a firmware setting. Hardware headline FLOPs are not delivered FLOPs.
Tension with the token-arbitrage thread (10-25), where inference resellers earned $10-20 per GPU-hour on rented B300s. If rent rises 20-30% while per-token serverless prices stay flat (DeepInfra's own choice), reseller margins compress unless stack efficiency improves.
Updates compute-economics: power and electrical gear, not GPUs, are now the binding constraint, and penalties for missing them are priced at $5M/day.

Research angle

For local inference, the unlock moves the prefill/decode balance: prefill gets cheaper, so prefill-heavy workloads (long prompts, RAG, speculative-decoding verification) benefit most on consumer cards. Expect NVIDIA to close the firmware path in a driver or VBIOS update; treat the user reports as unreproduced until a benchmark suite confirms them.