DiffusionGemma's Real-World Performance Shows Massive H100 vs A100 Gap

DiffusionGemma, a diffusion-based variant of Google's Gemma model, is undergoing rigorous internal testing by the qubridInc team. Early results show a dramatic divergence from typical transformer inference behavior when comparing H100 and A100 GPUs. While H100s scaled almost ideally under increasing concurrency, A100s performed well initially but the gap widened far more than expected—beyond the usual 'H100 is faster' difference. This suggests DiffusionGemma's architecture may have unique memory or compute patterns that amplify hardware advantages under load.

The model's sensitivity to workload characteristics is equally striking. On clean, short generations, DiffusionGemma looks 'insane'—almost suspiciously fast. However, introducing longer outputs, mixed request lengths, streaming, multi-user concurrency, or varied temperatures causes rapid behavioral shifts. A single messy, real-world-style batch can bring efficiency down far harder than anticipated. GPU utilization patterns also break from the classic token-by-token decode bottleneck seen in language models. The team is still running more tests and real traffic simulations, and the more they test, the more questions arise about scalability and deployment stability.