CMP 170HX vs RTX 5090

Best sourced output tok/s per model (single-stream generation) — engine and version, quantization, context and a source link on every row. Hover a value for the quant.

CMP 170HX16 sourced runs · best 255.9 tok/s

RTX 509023 verified runs · best 446.7 tok/s

Specs and evidence

	CMP 170HX	RTX 5090
VRAM	64 GB	32 GB
Sourced runs on the board	16	23
Models measured	11	16
Best output tok/s	255.9 tok/s	446.7 tok/s
Highest measured peak VRAM	58.7 GB	31.3 GB
Street price (latest snapshot)	—	—

Sourced rows only. Peak VRAM is recorded by the person who ran the benchmark, so it is a floor for what the card can hold, not a specification. Price snapshots.

Head-to-head

On the 3 models measured on both cards, CMP 170HX leads on 0, RTX 5090 on 3. The widest gap is on Ornith-1.5-35B-A3B: CMP 170HX 87.1 tok/s vs RTX 5090 258.8 tok/s (RTX 5090 197% faster).

The measured matrix

Model	CMP 170HX	RTX 5090
NVIDIA-Nemotron-3.5-Lightning-30B-A3B	165.1 Q8_0	446.7 NVFP4
Qwen3.8-27B	255.9 W4A16	435.1 IQ4_XS
Ornith-1.0-35B-PrismaAURA-4.75bit-vllm-MTP	—	283.1 NVFP4
Kimi-Linear-48B-A3B-Instruct	—	280.8 IQ4_XS
Ornith-1.5-35B-A3B	87.1 Q8_0	258.8 NVFP4
North-Mini-Code-1.0	—	258.4 UD-Q6_K_XL
Qwen3.6-35B-A3B-MTP	—	222.9 Unsloth-Dynamic-IQ4_XS
Laguna-XS-2.1	—	216.7 NVFP4
KAT-Coder-V2.5-Dev-AWQ-W4A16-ASYM	—	201.2 W4A16
Ornith-1.0-35B	—	196.6 AWQ
Muse-Glimmer-30B	—	124.8 NVFP4
Qwen3.6-27B	—	120.9 NVFP4
Huihui-Qwen3.8-27B-abliterated	—	117.7 NVFP4
Qwen3.8-27B-5090-goldilocks	—	93.5 Q6_K
Qwen3.6-35B-A3B	87.2 UD-Q8_K_XL	—
Ornith-1.5-9B	61.5 BF16	—
Qwen3.8-27B-MTP	46.4 INT8-W8A16	—
Qwen3.8-Flash-Next	46.1 UD-IQ1_S	—
Llama-4-Scout-17B-16E-Instruct	44.5 UD-Q3_K_XL	—
Qwen3.6-27B-MTP	—	28.3 Q5_K_M
gemma-4-31B-it	28.3 NVFP4	—
DeepSeek-V4-Flash-0731	—	26.4 MXFP4
granite-4.2-30b	23.7 Q8_0	—
Llama-3.3-70B-Instruct	20.1 NVFP4	—

Head-to-head (models tested on both): CMP 170HX 0 · RTX 5090 3 · tied 0 · data as of Oct 17, 2026. Full rules: methodology.

Related reading: RTX 5090 vs 3090 · quantization · Value index · Full board