Independent Sweep Puts DiffusionGemma Peak Throughput at 3k tok/s, 3x Paper's Claim
bodonoghue85 · x · 2026-09-06
A researcher re-ran throughput benchmarks for DiffusionGemma after noticing the paper omitted batch size details. Using 1xH200, PG19 data, 17.56 toks/forward, 1k input / 8k output with vLLM, he measured peak throughput of 3k tok/s at fp8 and 2.7k at bf16 — roughly 3x higher than reported in the uno paper. The paper's claimed 1k tok/s peak is not accurate and needs updating. He also notes it's unusual to claim inference wins at bf16 rather than fp8, since no one serves at bf16 at scale, and fp8 delivered further gains as expected.
Related event: DiffusionGemma Throughput Beats Paper Claims Threefold in Replication(2 posts)→
More from Infra
- Meta researcher: horrifying infra, not talent, is why OSS can't catch closed frontier — A_K_Nain · 2026-10-23
- Jane Street and NVIDIA speed up Mamba-2 with custom cuTile kernels — blelbach · 2026-10-23
- LatentLoop: asymmetric compute allocation cuts looped-MoE training cost by 19% — pmttyji · 2026-10-23
- Same AI request emits 8.5x more CO2 in Texas than in Sweden — jlundberg · 2026-10-23
- SK Hynix Pulls Plug on Solidigm US IPO as Chip Stocks Slide, Shares Down ~30% This Quarter — rohanpaul_ai · 2026-10-23
- Designer mocks Nvidia: buy idle GPUs or lose access to new ones—definitely not a bubble — michalmalewicz · 2026-10-23