RTX 5090 serving explorer
Every point is one measured cell from Scalably's own runs on two GeForce RTX 5090s, read from the engine's counters. Hover or focus a MoE point for its row ID in the dataset. Nothing here is a vendor number.
MoE vs dense, same vLLM flags
Nex-N2.5-mini (sparse MoE, NVFP4) against Qwen3.8-27B (dense, NVFP4), vLLM 0.27.0 with production flags, 256 decode tokens per request. Data: Scalably/moe-vs-dense-vllm-rtx5090 · write-up: scalably.io.
Table view
Qwen Flash at a full 256K window
One request: a ~254K-token prompt and 8,192 output tokens, wall throughput including prefill. Warm is an unchanged-prefix repeat. Data: Scalably/qwen-flash-dual-5090-serving · write-up: scalably.io.
NInfer-ext ran on one GPU. Engine and quantization changed together in several arms, so these are not equal-fidelity comparisons. Warm rows missing for three engines ended before a complete repeat; that is not zero.