Skip to content

NVIDIA DGX Spark AI benchmarks

Explore LLM and image benchmark results for NVIDIA DGX Spark, with hardware specifications and the software settings used in each test.

Hardware specifications

CPU
10x Cortex-X925 + 10x Cortex-A725
GPU / SoC
NVIDIA GB10 Grace Blackwell
Memory
128 GB unified
INT8 compute
250 TOPS
Memory bandwidth
273 GB/s
Storage
4 TB SSD
OS
Ubuntu 24.04 (DGX OS kernel)
Driver
NVIDIA 580.159.03, CUDA 13.0

Benchmark results

ModelBenchmarkConfigurationDGX SparkOlares One
Qwen3.8-27BIFBench
Evaluation time
llama.cpp, NVFP4 LOW18.2 h8.7 h
Qwen3.8-27BInference speed
Token generation, single request
llama.cpp, NVFP4 LOW45 tok/s93 tok/s
Qwen3.8-Flash-NextInference speed
Token generation, single request
Strata, Q2_080 tok/s166 tok/s
Qwen-Image-2.1 BaseText to image
Per image
ComfyUI, NVFP427 s12 s
Qwen-Image-2.1 BaseImage edit
Per image
ComfyUI, NVFP437 s17 s
Qwen-Image-2.1 TurboImage edit
Per image
ComfyUI, int813 s5.2 s

Test configurations

Qwen3.8-27B

IFBench

llama.cpp, NVFP4 LOW

Engine and build
llama.cpp b11718 (547208dca), self-built from the b11718 source (scripts/build.sh, CMAKE_CUDA_ARCHITECTURES=121a-real), CUDA 13.0, sm_121a (GB10)
Launch arguments
--mmproj mmproj-F16.gguf -c 104448 -ngl all -fa on -ctk q8_0 -ctv q8_0 --jinja -np 1 --spec-type draft-dflash -md Qwen3.8-27B-DFlash2-Q8_0.gguf -ngld all --spec-draft-n-max 7

Inference speed

llama.cpp, NVFP4 LOW

Engine and build
llama.cpp b11718 (547208dca), self-built from the b11718 source (scripts/build.sh, CMAKE_CUDA_ARCHITECTURES=121a-real), CUDA 13.0, sm_121a (GB10)
Launch arguments
--mmproj mmproj-F16.gguf -c 104448 -ngl all -fa on -ctk q8_0 -ctv q8_0 --jinja -np 4 -kvu --spec-type draft-dflash -md Qwen3.8-27B-DFlash2-Q8_0.gguf -ngld all --spec-draft-n-max 7

Qwen3.8-Flash-Next

Inference speed

Strata, Q2_0

Engine and build
Strata strata 0.2.10 (engine 0.1.56)
Launch arguments
--pack q2_0 --native Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00001-of-00002.gguf --ple-gguf Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00002-of-00002.gguf --expert-profile expert-profile.bin --expert-cache auto --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp rt --max-context 131072 --kv int8 --mmap-experts --mtp-q4 all --mtp-hnorm stream --ple-io ram --kv-resident 32768 --adapt-every 100000

Inference speed

Strata, IQ3_XXS

Engine and build
Strata strata 0.2.10 (engine 0.1.56)
Launch arguments
--pack iq3_xxs --native Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf --expert-profile expert-profile.bin --expert-cache auto --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp rt --max-context 131072 --kv int8 --mmap-experts --mtp-q4 all --mtp-hnorm stream --ple-io ram --kv-resident 32768 --adapt-every 100000

Inference speed

Strata, IQ3_S

Engine and build
Strata strata 0.2.10 (engine 0.1.56)
Launch arguments
--pack iq3_s --native Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf --ple-gguf Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf --expert-profile expert-profile.bin --expert-cache auto --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp rt --max-context 131072 --kv int8 --mmap-experts --mtp-q4 all --mtp-hnorm stream --ple-io ram --kv-resident 32768 --adapt-every 100000

Inference speed

Strata, UD-IQ4_XS

Engine and build
Strata strata 0.2.10 (engine 0.1.56)
Launch arguments
--pack unsloth-ud-iq4_xs --native Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf --expert-profile expert-profile.bin --expert-cache 16405 --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp rt --max-context 131072 --kv int8 --mmap-experts --mtp-q4 all --mtp-hnorm stream --ple-io ram --kv-resident 32768 --resident-budget-gib 55 --adapt-every 100000

Qwen-Image-2.1

ComfyUI, NVFP4

Qwen-Image-2.1 Base: Text to image, Image edit

Qwen-Image-2.1 Turbo: Image edit

Engine and build
ComfyUI v0.44.0-23-g63e5bf59, git checkout in comfyui, venv/comfyui (uv pip, torch 2.15.0+cu132 on ARM64, driver CUDA 13.0 via minor-version compatibility), CUDA, torch 2.15.0+cu132, Python 3.12.3
Launch arguments
--listen 127.0.0.1 --port 8288 --extra-model-paths-config extra_model_paths.yaml --output-directory images --input-directory inputs

ComfyUI, int8

Qwen-Image-2.1 Base: Text to image, Image edit

Qwen-Image-2.1 Turbo: Image edit

Engine and build
ComfyUI v0.44.0-23-g63e5bf59, git checkout in comfyui, venv/comfyui (uv pip, torch 2.15.0+cu132 on ARM64, driver CUDA 13.0 via minor-version compatibility), CUDA, torch 2.15.0+cu132, Python 3.12.3
Launch arguments
--listen 127.0.0.1 --port 8288 --extra-model-paths-config extra_model_paths.yaml --output-directory images --input-directory inputs