RequantTESTNET
Requant research

Consumer Blackwell Research Programme

The question

What does a consumer GPU deliver once a comparison charges preparation, movement, validation and the application's accuracy target? The RTX 5060 Ti programme followed that question from FP4 products to complete solvers, visibility queries, streaming transfer and request scheduling. All fifteen studies completed their bounded protocols. The results describe the measured hardware and workloads.

How to interpret the evidence

The report set separates qualified optimizations, negative results and unavailable configurations. CPU replays check retained outputs and records. GPU measurements use the hardware and execution boundaries declared in each report. The shared host conditions, retained errors and input cohorts belong to the result. Read research methodology for timing, memory and counter interpretation.

Studies and findings

StudyInvestigationFinding
1FP4 Efficiency on Consumer BlackwellPreparation and error can erase the FP4 lead.
2FP4 Outlier Correction at Fixed Error and TimeCorrection improves some errors at a measurable cost.
3GPU Quantization and Activation FusionResource costs change when quantization and activation fuse.
4Blackwell GPU Pipelines, Traversal and BufferingResource fit and input reuse govern configuration selection.
5Mixed-Precision Attention on Consumer BlackwellArithmetic error and complete model quality need separate checks.
6High-Precision Products from Low-Precision GPU ArithmeticReconstruction performance depends on the workload and reference.
7Adaptive-Precision Dense Linear Solvers on GPUsA small residual can coexist with unacceptable forward error.
8Hybrid Sparse–Dense Matrix Multiplication on GPUsThe active hybrid won zero of 52 tested online groups.
9GPU FFT and Complete Spectral FilteringThe strongest fused FP32 reference provided the positive latency result.
10Quantized GPU Vector Search and Indexed ControlsThe tested FP4 path missed shifted-input recall targets.
11Streaming Covariance and PCA on GPUsFP4 was infeasible at the stricter tested joint error budget.
12Matrix-Free Poisson and Heat Operators on GPUsComplete solves expose constraints beyond operator throughput.
13Vulkan Ray Query and CUDA BVH VisibilityExact regular scenes, reproducible boundary differences.
14Lossless CPU-to-GPU Streaming Beyond VRAMPlain pinned transfer was fastest on the tested 20 GiB stream.
15GPU Board Energy, Batching and Tail LatencyBatching lowered measured board energy in all 12 final pairs.

Reproduction and citation

Requant. “Consumer Blackwell Research Programme.” Technical research overview, 19 November 2026.