Two 3090s or one 5090? The answer isn't the one everyone gives

The dual-3090 build doesn't give you double the bandwidth for chat — and the reason has nothing to do with the cards. What the measurements actually… Aliteq Ravi Malhotra · Hardware Editor

Two 3090s do not give you 1,872 GB/s for single-stream chat. llama.cpp's default multi-GPU mode is pipeline parallel — the cards take turns, so you get roughly one 3090's decode speed with 48 GB…

The mode that would aggregate bandwidth, -sm tensor, is marked experimental, forbids quantised KV cache, doesn't support MoE models, and is currently filed as broken.

On measured single-card decode the 5090 is 1.83× a 3090 — which is 95.8% of what pure memory bandwidth predicts. Prefill is far more lopsided at 2.89×, because prompt processing is compute-bound…

Buy two 3090s if your job is fit a 70B at all, or if you serve several concurrent users. Buy one 5090 if your job is one conversation, as fast as possible.

Both cards have a real software problem in 2026, in opposite directions: the 3090 is locked out of native FP8, and the 5090 is barely tuned for in vLLM.

Decide what you're actually building before you decide what to buy. One person, one conversation, wanting it fast — the 5090 wins, and the pipeline-parallel behaviour means the dual-3090 alternative…

It depends on batch size, and almost nothing else

Read the full story

Aliteq Two 3090s or one 5090? The answer isn't the one everyone gives Read the full story on Aliteq