GPU Quantization and Activation Fusion
Abstract
Fused bias, ReLU and FP4 preparation are compared with byte-identical separate paths and complete real-classifier references. Thirty CUDA function and block-size queries characterize resources and theoretical occupancy.
Study 3 of the Consumer Blackwell programme. See the research methodology for timing and replay boundaries.
The encoder study compares separate and fused bias/ReLU/FP4 preparation across shapes and profiles. The native hardware study adds hardware packed conversion and byte-identical fused/separate controls. The complete-application study evaluates two classifiers, final labels/task accuracy, complete host requests, native cuBLASLt FP4 and eager/compiled/CUDA-Graph PyTorch references. Those reports retain both useful gains and slower configurations. Compiled-reference CUDA/toolchain differences are recorded rather than hidden.
This checkpoint supplies the remaining resource characterization: thirty function/block-size configurations query actual CUDA attributes and theoretical occupancy. The frozen encoders are recompiled in the traversal-control library; source hashes and exact binary identity are retained in resource data. These are not counter-derived achieved occupancy or spill measurements.
| Encoder variant | Activation fused | Registers/thread | Shared bytes | Local bytes/thread | Theoretical thread occupancy, block128 |
|---|---|---|---|---|---|
| 2 | False | 22 | 0 | 0 | 1.000 |
| 2 | True | 22 | 0 | 0 | 1.000 |
| 10 | False | 34 | 0 | 0 | 1.000 |
| 10 | True | 40 | 0 | 0 | 1.000 |
| 11 | False | 21 | 0 | 0 | 1.000 |
| 11 | True | 21 | 0 | 0 | 1.000 |
| 12 | False | 23 | 0 | 0 | 1.000 |
| 12 | True | 23 | 0 | 0 | 1.000 |
| 13 | False | 25 | 0 | 0 | 1.000 |
| 13 | True | 25 | 0 | 0 | 1.000 |
The native scalar16 implementation uses more registers with activation fused (34 to 40); vector variants retain their measured register counts. A theoretical occupancy of one does not mean the grid fills the device or that memory/compute pipelines are saturated. The full data retain block sizes 128, 256 and 512, while block size alone does not retune register allocation. The pipeline study independently retains large local/shared-storage footprints, unavailable requests and their latency outcomes. Those resource controls do not imply that every slowdown is caused by spilling.
Conclusions remain bounded: removing a pass can help while preparation/work assignment/tile choices still regress; complete classifier quality and request latency must be compared separately from isolated preparation speed. No scientific novelty, architecture ceiling or achieved hardware-utilization claim is made. Counter access remains denied, and the shared-system timing qualification remains in force.
Evidence and source revision
Cite this report
Requant. “GPU Quantization and Activation Fusion.” Technical report, 19 November 2026. Evidence revision d12bcf1c3270f28fe85828725fb2f2c5c790b986. Canonical report.
This is an organizational technical report, not a peer-reviewed publication.