Glyd

Tested GPUs

The NVIDIA GPUs each Glyd release is checked and timed on, and Glyd against bf16 on each, release to release.

v0.22.0, Sep 28, 2026Hopper · Ampere · Blackwell · Ada

The GPUs

GPUArchitectureMemoryVerified
NVIDIAH100 PCIeHopper · 80 GBHopper80 GBBit for bitlog
NVIDIAA100 SXM4 40 GBAmpere · 40 GBAmpere40 GBBit for bitlog
NVIDIAA10Ampere · 24 GBAmpere24 GBBit for bitlog
NVIDIAA10GAmpere · 24 GBAmpere24 GBBit for bitlog
NVIDIARTX PRO 6000 Blackwell Server EditionBlackwell · 96 GBBlackwell96 GBBit for bitlog
NVIDIARTX 4080 SUPERAda · 16 GBAda16 GBBit for bitlog

Verified: with exact=True, the installed package's output on the GPU is bf16's, bit for bit.

Glyd against bf16, release to release

bf16 and Glyd on the same GPU. A generation step's GPU time after the prompt, and one forward pass over a prompt; lower is better. Change: Glyd against bf16, in percent.

Weights on the GPU, GBGeneration step, 1 sequence, msPrompt of 2,048 tokens, ms
Releasebf16GlydChangebf16GlydChangebf16GlydChangeLog
H100 PCIeQwen3-8B, 12-bit layout
v0.22.016.3812.27−2512.0811.08−8119.1155.2+30log
v0.21.016.3812.27−2512.2911.24−9———log
A100 SXM4 40 GBQwen3-8B, 12-bit layout
v0.22.016.3812.27−2519.0115.29−20176.8197.4+12log
v0.21.016.3812.27−2517.4515.83−9———log
A10Qwen3-8B, 12-bit layout
v0.22.016.3812.27−2535.7027.75−22543.3750.9+38log
A10GQwen3-8B, 12-bit layout
v0.22.016.3812.27−2537.1327.80−25589.9614.3+4log
RTX PRO 6000 Blackwell Server EditionQwen3-8B, 12-bit layout
v0.22.016.3812.27−2513.4011.29−16103.0119.6+16log
v0.21.016.3812.27−2513.4211.22−16103.2130.5+26log
RTX 4080 SUPERQwen3-4B-Instruct-2507, smallest layout
v0.22.08.046.00−2514.5511.75−19197.3198.4+1log
v0.21.08.046.00−25———198.1211.5+7log
Methodology and raw logs