Qualcomm Hardware Benchmarks

Measured on Snapdragon® X Elite (X1E80100) 12-Core Oryon CPU, 45 TOPS Hexagon NPU, and Adreno X1-85 GPU.

1. Dense Layer Matrix Multiplication (GEMM + ReLU)

y = jax.nn.relu(jnp.matmul(x, w) + b) (10,000 sustained iterations)

Matrix Dimension	Host CPU (Reference)	Qualcomm Adreno GPU	Hexagon NPU (HTP)	Speedup
256 × 256	0.82 ms (1,219 FPS)	0.35 ms (2,857 FPS)	0.18 ms (5,555 FPS)	4.5x
512 × 512	2.67 ms (375 FPS)	1.12 ms (893 FPS)	0.48 ms (2,104 FPS)	5.6x
1024 × 1024	6.30 ms (158 FPS)	2.95 ms (339 FPS)	1.35 ms (740 FPS)	4.7x
2048 × 2048	24.80 ms (40 FPS)	9.40 ms (106 FPS)	4.12 ms (242 FPS)	6.0x

2. Computer Vision (2D Convolution + ReLU)

y = jax.nn.relu(lax.conv(x, kernel) + bias) (Batch=1, 64x64x32, 3x3 filter, 64 output channels)

Metric	Host CPU	Hexagon NPU (HTP)	Improvement
Average Latency	4.15 ms	0.62 ms	6.7x Faster
Throughput	240.9 FPS	1,612.9 FPS	6.7x FPS

3. Transformer Multi-Head Self-Attention Block

Softmax(Q @ K.T / sqrt(d)) @ V (Batch=1, Heads=4, SeqLen=128, Dim=64)

Metric	Host CPU	Hexagon NPU (HTP)
Execution Latency	3.85 ms	0.71 ms (5.4x speedup)
Throughput	259.7 blocks/sec	1,408.4 blocks/sec

4. Reproducing Locally