RequantTESTNET
Requant research

FP4 Efficiency on Consumer Blackwell

Abstract

The pilot compares five formats across six shapes, two input profiles and three preparation boundaries. The held-out selector retains a budget violation, and complete classifier measurements test whether the kernel gains survive application costs.

Study 1 of the Consumer Blackwell programme. See the research methodology for timing and replay boundaries.

What the pilot answers

The complete retained grid distinguishes three different workloads: a prepared matrix product, changing A with reusable B, and preparation of both operands. It also separates errors introduced by quantization from errors in the GPU matrix arithmetic. See the results for cohort ranges and all formats.

In the tested baseline, preparation reduces or removes FP4's observed latency advantage in several shapes. Small matrix products also do not benefit simply because FP4 has high nominal Tensor Core throughput. Larger tested products can still favor FP4 with preparation included, but with materially higher error.

These observations support investigating preparation and shape-dependent selection. They do not establish that FP4 is useful for a particular model or scientific application. Relative Frobenius error is not an application quality metric.

What remains unresolved

  • Equal-error comparisons and task-level quality.
  • Optimized and alternative FP4 encoders, tile shapes and schedules.
  • More shapes, real activations, fresh-data streaming and uncaptured request latency.
  • Host transfer, allocation and operand-reuse amortization costs.
  • Exclusive-device replication and statistical stability across separate sessions.
  • Generalization of a format/configuration selector to held-out workloads.

Existing GPU contexts remained present. The two cohorts use independent inputs and reverse ordering, but are not independent machines or separate-day trials. Frequency, temperature and cache effects are not eliminated by this protocol.

Next bounded experiment

This encoder/fusion/tile follow-up has now been executed; its results and retained regressions are in ENCODER_REPORT.md. The text below preserves the question that motivated it. Complete application and format-selection gates remain open in the program ledger.

Optimize the online FP4 encoder while preserving its exact quantization contract. Keep the same retained inputs, scale rule, packed-byte oracle and GEMM configuration. Compare the original encoder with a vectorized or warp-cooperative implementation. Measure preparation alone, online-A plus GEMM and online-A+B plus GEMM.

Require exact agreement of every encoded byte and scale. A speed improvement that changes the quantization recipe belongs in a separate error/quality experiment. Include small shapes where extra synchronization can make cooperation slower. Retain unsuccessful configurations rather than reporting only the fastest one.

Research status

The bounded first pilot is complete. The research program remains in progress under the gates in RESEARCH.md. This is a starting evidence record, not a final publication or an exhausted search for the best low-precision method.

The completed efficiency investigation

The pilot is one part of Study 1. The encoder and tile experiments extend the distribution and shape grid. The application experiments separate repeated graph arithmetic from fresh requests, model accuracy and weight reuse. The held-out selector retains one budget violation across 192 held-out decisions. It also measures the selection overhead.

Evidence and source revision

Cite this report

Requant. “FP4 Efficiency on Consumer Blackwell.” Technical report, 19 November 2026. Evidence revision d12bcf1c3270f28fe85828725fb2f2c5c790b986. Canonical report.

This is an organizational technical report, not a peer-reviewed publication.