No fp8 on Ampere: maker ships int8 Krea 2 Turbo build for 20/30-series GPUs
in_habitants · reddit · 2026-11-13
After merging the Stillroom style into Krea 2 Turbo, the author shipped fp8 first—but a 3070 Ti mobile user reported 2x slowdown, since Ampere lacks fp8 math and ComfyUI converts every layer on the fly. int8 is hardware-accelerated from sm 7.5, so the author rebuilt an int8 checkpoint from bf16 (224 linear layers in int8, per-row scale, txtfusion kept in bf16).
Key finding: the official int8 doesn't use plain absmax for per-row scales (only 43% byte match when tried)—it clips scales up to 6% below absmax. A per-row error search landed at 0.00871 vs the official 0.00874.
Rule of thumb: fp8 for 40-series and up, int8 for 20/30-series, bf16 if VRAM allows. All three builds are on Civitai under the Krea 2 license.
More from Infra
- Cloudflare hosts Thinking Machines' inkling 2 small on Workers AI at day-0 speed — michellechen · 2026-11-14
- Liquid-Codec decoder ported to Apple Neural Engine, 6x faster at 316x real time — JosephJacks_ · 2026-11-14
- AMD MI355X with FlyDSL kernels beats Blackwell running MiniMax M3.1 in vLLM — AnushElangovan · 2026-11-14
- Tencent paper: pruning neurons by routed contribution keeps 91-98% MoE performance — rohanpaul_ai · 2026-11-14
- AMD's hrx-loom stack: agent-written kernels beat llama.cpp across 51 LLMs on Strix Halo — jfowers_amd · 2026-11-14
- Microsoft Cloud tops $63B as AI buildout keeps scaling, says Nadella — thedealdirector · 2026-11-14