CMP 170HX vs DGX Spark

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

DGX Spark52 verified runs · best 86.8 tok/s

Specs and evidence

	CMP 170HX	DGX Spark
VRAM	64 GB	128 GB
Sourced runs on the board	16	52
Models measured	11	9
Best output tok/s	255.9 tok/s	86.8 tok/s
Highest measured peak VRAM	58.7 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 2 models measured on both cards, CMP 170HX leads on 1, DGX Spark on 1. The widest gap is on Qwen3.8-27B: CMP 170HX 255.9 tok/s vs DGX Spark 56.0 tok/s (CMP 170HX 357% faster).

The measured matrix

Model	CMP 170HX	DGX Spark
Qwen3.8-27B	255.9 W4A16	56.0 NVFP4 W4A4
NVIDIA-Nemotron-3.5-Lightning-30B-A3B	165.1 Q8_0	—
Qwen3.6-35B-A3B	87.2 UD-Q8_K_XL	—
Ornith-1.5-35B-A3B	87.1 Q8_0	—
Qwen 3.8 Flash	—	86.8 EXL3 3.05bpw
Qwen3.8-Flash-Next	46.1 UD-IQ1_S	75.7 4.05bpw_h6_ng6
GLM-5.3-Flash	—	64.1 EXL3 2.05 bpw
Ornith-1.5-9B	61.5 BF16	—
CYBER-FROST-3.8	—	60.7 EXL3
Qwen3.8-Flash-Next-hibrid48	—	60.0 NVFP4 (output head)
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B	—	48.0 NVFP4
DeepSeek-V4.1-Flash	—	47.0 FP4
Qwen3.8-27B-MTP	46.4 INT8-W8A16	—
Llama-4-Scout-17B-16E-Instruct	44.5 UD-Q3_K_XL	—
gemma-4-31B-it	28.3 NVFP4	—
granite-4.2-30b	23.7 Q8_0	—
Underdog-Saluki-27B	—	22.6 IQ2-mix
Llama-3.3-70B-Instruct	20.1 NVFP4	—

Head-to-head (models tested on both): CMP 170HX 1 · DGX Spark 1 · tied 0 · data as of Oct 30, 2026. Full rules: methodology.

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