DiffusionGemma hits 22 structured generations/sec on a DGX Spark at concurrency 32
bodonoghue85 · x · 2026-09-17
mmastrac reports progress making DiffusionGemma act like Jev: with basically zero tuning on a DGX Spark, a Jev-like patch runs 0.21s p50 at concurrency 1, and pushes 22 structured generations per second at concurrency 32 — a strong showing for diffusion decoding on compact hardware.
More from Infra
- GLM-5.3-powered Infra Agent boosts own inference stack 3x on 100k Chinese accelerators — Dr_Singularity · 2026-09-17
- One CPU Core Inspects 28.8M Packets/Sec: How XDP Rewrites Network Processing — blaizedsouza · 2026-09-17
- Dev's Rust/Wasm AAC Encoder Beats FFMPEG by 4-6x in Benchmarks — wavefnx · 2026-09-17
- OpenRouter weekly tokens surge 25,000% to 126.2 trillion — the AI bubble debate in one chart — The Decoder · 2026-09-17
- How do you verify an untrusted GPU host actually ran the model? Gonka's design notes on 3 cheating vectors — autoimago · 2026-09-17
- Lenovo launches ThinkAgile VX850 V4 servers that let apps stay on VMware or Hyper-V — shashib · 2026-09-17