5.5838 → 1.8332 ms
New time relative to old time
Let the numbers talk._
The 29 September 2026 tables are one thread. On the evening of 21 October 2026, on this Ryzen 7 5800H, GEMM kept avx2-4x16: N=2048 Nexus 330, 362, and 369 GFLOPS, TensorFlow 389. CPU Conv2d was behind PyTorch at 31.6 GFLOPS; Nexus was about 4.9 to 6.8 GFLOPS. Depthwise AVX was ahead of PyTorch 4.93 and TensorFlow 0.688. AVX-512 is not selected. There is no world ranking.
AMD Ryzen 7 5800H with Radeon GraphicsCPU · FLOAT32NUMPY · PYTORCH · NEXUSMODEL
LOCAL CPU STUDY / 3 LIBRARIES
NexusModel vs NumPy vs PyTorch
29 September 2026, one thread, small GEMM and ReLU. Verified outputs. The 21 October evening is separate: N=2048 GEMM avx2-4x16, Nexus 330/362/369 GFLOPS, TensorFlow 389. Conv2d was behind PyTorch at 31.6 GFLOPS.
All outputs verified
AMD Ryzen 7 5800H with Radeon GraphicsFLOAT32 / CPU1 COMPUTE THREAD20 WARMUP · 90 × 52026-09-29
Y = X Wᵀ; no biasTIME / ms · LOWER IS BETTER
FOR THIS WORKLOAD
NexusModel recorded the lowest time. 0.0060 ms
Rankings change with operation, size, thread count, and hardware.All library measurements
27 RESULTS| Operation / shape | Library | Median ms | p95 ms | Max. absolute error |
|---|---|---|---|---|
| GEMM1 × 256 × 256 | NexusModel | 0.0060 | 0.0097 | 2.75e-7 |
| GEMM1 × 256 × 256 | NumPy | 0.0137 | 0.0228 | 3.75e-7 |
| GEMM1 × 256 × 256 | PyTorch | 0.0194 | 0.0296 | 3.83e-7 |
| GEMM7 × 129 × 33 | NexusModel | 0.0039 | 0.0059 | 4.13e-7 |
| GEMM7 × 129 × 33 | NumPy | 0.0159 | 0.0192 | 4.13e-7 |
| GEMM7 × 129 × 33 | PyTorch | 0.0073 | 0.0178 | 4.13e-7 |
| GEMM32 × 256 × 256 | NexusModel | 0.1546 | 0.1802 | 1.10e-6 |
| GEMM32 × 256 × 256 | NumPy | 0.1941 | 0.2561 | 1.84e-6 |
| GEMM32 × 256 × 256 | PyTorch | 0.1421 | 0.1727 | 1.10e-6 |
| GEMM96 × 768 × 192 | NexusModel | 0.7411 | 0.9125 | 2.82e-6 |
| GEMM96 × 768 × 192 | NumPy | 0.8929 | 1.0273 | 3.33e-6 |
| GEMM96 × 768 × 192 | PyTorch | 0.8082 | 0.8698 | 3.22e-6 |
| GEMM128 × 512 × 512 | NexusModel | 1.7653 | 2.0519 | 1.97e-6 |
| GEMM128 × 512 × 512 | NumPy | 1.9080 | 2.1503 | 2.84e-6 |
| GEMM128 × 512 × 512 | PyTorch | 1.8616 | 2.0455 | 2.28e-6 |
| GEMM256 × 1024 × 1024 | NexusModel | 15.0622 | 15.9896 | 3.67e-6 |
| GEMM256 × 1024 × 1024 | NumPy | 13.7320 | 14.5004 | 4.24e-6 |
| GEMM256 × 1024 × 1024 | PyTorch | 14.4382 | 15.7726 | 3.96e-6 |
| ReLU65,536 | NexusModel | 0.0181 | 0.0239 | 0 |
| ReLU65,536 | NumPy | 0.1000 | 0.1365 | 0 |
| ReLU65,536 | PyTorch | 0.0290 | 0.0380 | 0 |
| ReLU1,048,576 | NexusModel | 0.2388 | 0.3104 | 0 |
| ReLU1,048,576 | NumPy | 1.3875 | 1.6029 | 0 |
| ReLU1,048,576 | PyTorch | 0.3113 | 0.4320 | 0 |
| ReLU4,194,304 | NexusModel | 2.1986 | 2.7186 | 0 |
| ReLU4,194,304 | NumPy | 5.5412 | 6.2044 | 0 |
| ReLU4,194,304 | PyTorch | 2.3128 | 2.9211 | 0 |
NEXUSMODEL / BEFORE & AFTER
2026-09-29 · 90 SAMPLESSame computation. Less time.
Archived sources and the optimized revision were built with the same C++ timing harness. Three rounds alternated revision order; values are median times.
14.2197 → 4.0230 ms
New time relative to old time
0.1421 → 0.0877 ms
New time relative to old time
0.3115 → 0.0915 ms
New time relative to old time
29 September 2026. Float32, one compute thread. Backward includes gradient buffer clearing. Power was not pinned. These cards are not the 21 October GEMM or Conv2d GFLOPS column. That evening CPU Conv2d was behind PyTorch at 31.6 GFLOPS.
This is a CPU operation comparison.
Forward-only GEMM (no bias) and ReLU with preallocated outputs. Python API dispatch is included for NumPy/PyTorch, C++ call overhead for NexusModel. Backward, autograd, allocation and transfers are excluded. p95 is the percentile of 90 samples, each averaging 5 calls. NumPy is a numerical array library, not a full neural-network training framework.
One machine, three rounds with rotated framework order; OS scheduling and temperature were not controlled. Every output was checked against a float64 reference. Versions, BLAS backends, raw timings, and source SHA-256 hashes are in the JSON.