Skip to content

fp16 matvec FMA

Same register-resident 4x8 matvec tile as the fp32 matvec test, but weights, inputs, and dot() accumulation are all f16. Pure ALU: the fp16 win here comes only from the ALU, not from halved memory traffic.

Measured in FLOP/s. Higher values mean better performance. Benchmark your device.

Performance distribution

 
1.65 GFLOP/s271 GFLOP/s44.4 TFLOP/s

Performance thresholds

99% of results achieve at least 5.83 GFLOP/s

95% of results achieve at least 97.5 GFLOP/s

90% of results achieve at least 245 GFLOP/s

Median (50th percentile) 2.14 TFLOP/s

Best performer

Highest recorded performance: 44.4 TFLOP/s. View benchmark profile.

Recorded 9/14/2026, 2:52:24 PM

WebGPU device
Vendornvidia
Device—
Architectureblackwell
Description—
WebGL device
VendorGoogle Inc. (NVIDIA)
RendererANGLE (NVIDIA, NVIDIA GeForce RTX 5090 (0x00002B85) Direct3D11 vs_5_0 ps_5_0, D3D11)
Browser
NameChrome
Version153.0.0.0
Operating system
NameWindows
Version10