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
| Study | Investigation | Finding |
|---|---|---|
| 1 | FP4 Efficiency on Consumer Blackwell | Preparation and error can erase the FP4 lead. |
| 2 | FP4 Outlier Correction at Fixed Error and Time | Correction improves some errors at a measurable cost. |
| 3 | GPU Quantization and Activation Fusion | Resource costs change when quantization and activation fuse. |
| 4 | Blackwell GPU Pipelines, Traversal and Buffering | Resource fit and input reuse govern configuration selection. |
| 5 | Mixed-Precision Attention on Consumer Blackwell | Arithmetic error and complete model quality need separate checks. |
| 6 | High-Precision Products from Low-Precision GPU Arithmetic | Reconstruction performance depends on the workload and reference. |
| 7 | Adaptive-Precision Dense Linear Solvers on GPUs | A small residual can coexist with unacceptable forward error. |
| 8 | Hybrid Sparse–Dense Matrix Multiplication on GPUs | The active hybrid won zero of 52 tested online groups. |
| 9 | GPU FFT and Complete Spectral Filtering | The strongest fused FP32 reference provided the positive latency result. |
| 10 | Quantized GPU Vector Search and Indexed Controls | The tested FP4 path missed shifted-input recall targets. |
| 11 | Streaming Covariance and PCA on GPUs | FP4 was infeasible at the stricter tested joint error budget. |
| 12 | Matrix-Free Poisson and Heat Operators on GPUs | Complete solves expose constraints beyond operator throughput. |
| 13 | Vulkan Ray Query and CUDA BVH Visibility | Exact regular scenes, reproducible boundary differences. |
| 14 | Lossless CPU-to-GPU Streaming Beyond VRAM | Plain pinned transfer was fastest on the tested 20 GiB stream. |
| 15 | GPU Board Energy, Batching and Tail Latency | Batching lowered measured board energy in all 12 final pairs. |
Reproduction and citation
- Exact Git revision and programme ledger
- CPU/artifact CI for the evidence revision
- Earlier GPU and cryptography reports
Requant. “Consumer Blackwell Research Programme.” Technical research overview, 19 November 2026.