Hardware › Strix Halo vs DGX Spark: measured LLM tok/s on 10 shared models
Strix Halo vs DGX Spark: measured LLM tok/s on 10 shared models
Where both run llama.cpp: Strix Halo ahead on 3, DGX Spark on 1, within 5% on 3. 3 rows pair different engines. Same model at the same bit-width, best known number per box with source links; not a controlled test.
Snapshot 2026-10-29 · 413 configs on AI boxes and 288 on GPU cards · 121 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.
Strix Halo vs DGX Spark, measured on 10 shared model and bit-width pairs. Where both run llama.cpp: Strix Halo ahead on 3, DGX Spark on 1, within 5% on 3. 3 rows pair different engines. Median ratio Strix Halo to DGX Spark, best known number per box: 1x. Snapshot 2026-10-29.
Same model, same bit-width, both boxes
| Model | Bits | Strix Halo | Strix Halo source | DGX Spark | DGX Spark source | Ratio Strix Halo to DGX Spark |
|---|---|---|---|---|---|---|
| Gemma 4 12B IT QAT | 4-bit | 29.3 tok/s (UD-Q4_K_XL, llama.cpp/vulkan) | source | 28.4 tok/s (Q4_K_M, llama.cpp) | source | 1x |
| Gemma 4 31B IT QAT | 4-bit | 11.4 tok/s (Q4_0, llama.cpp/vulkan) | source | 11 tok/s (Q4_K_M, llama.cpp/cuda) | source | 1x |
| gemma-4-26B-A4B-it | 4-bit | 78 tok/s (UD-Q4_K_XL, llama.cpp) | source | 69.9 tok/s (Q4_K_M, llama.cpp/cuda) | source | 1.1x |
| gemma-4-26B-A4B-it | 8-bit | 41.8 tok/s (UD-Q8_K_XL, llama.cpp/rocm-7.2.4) | source | 47.5 tok/s (Q8_0, llama.cpp/cuda) | source | 0.9x |
| Nemotron-3-Nano-Omni-30B-A3B | 4-bit | 53.2 tok/s (NVFP4, llama.cpp/vulkan) | source | 41.7 tok/s (NVFP4, vllm) | source | 1.3x |
| Nemotron-3-Super-120B-A12B | 4-bit | 18.9 tok/s (IQ4_XS, llama.cpp/vulkan) | source | 24.5 tok/s (NVFP4, vllm) | source | 0.8x |
| Qwen3-Coder-Next-80B-A3B | 4-bit | 61.9 tok/s (IQ4_XS, llama.cpp/vulkan) | source | 50.1 tok/s (Q4_K_XL, llama.cpp) | source | 1.2x |
| Qwen3.6-35B-A3B | 4-bit | 81.3 tok/s (Q4_0, llama.cpp/vulkan) | source | 77.4 tok/s (NVFP4, vllm) | source | 1.1x |
| Qwen3.6-35B-A3B | 8-bit | 49.2 tok/s (UD-Q8_K_XL, llama.cpp/vulkan-radv) | source | 50.2 tok/s (Q8_0, llama.cpp/cuda) | source | 1x |
| Qwen3.8-Flash-Next | 4-bit | 41.3 tok/s (NVFP4, backend not stated) | source | 36.7 tok/s (NVFP4, atlas) | source | 1.1x |
How to read this
Each cell is the best known single-stream number for that box, taken from its source. It is not a controlled test: backends, quant formats, context length and the person measuring differ between the two sides, and the source link on each cell shows which. A group counts only when both boxes have a plain-decoding run (no speculative decoding) of the same model family at the same bit-width. Models under 7B parameters are left out.
Price in Singapore
- Strix Halo: from S$4,329 today
- DGX Spark: from S$7,452 today
Questions and answers
- Is Strix Halo faster than DGX Spark for local LLMs?
- It depends on the model and the engine. Where both run llama.cpp: Strix Halo ahead on 3, DGX Spark on 1, within 5% on 3. 3 rows pair different engines. Across all 10 shared model and bit-width pairs, the median ratio Strix Halo to DGX Spark, best known number per box, is 1x.
- Which model shows the biggest gap between Strix Halo and DGX Spark?
- Nemotron-3-Super-120B-A12B at 4-bit: DGX Spark 24.5 tok/s against Strix Halo 18.9 tok/s, a ratio of 1.3x.
- Is this a fair benchmark of Strix Halo against DGX Spark?
- No. These are the best known numbers per box, taken from different sources. Backends, quant formats, context length and who measured it all differ, and each row links to its source. Speculative-decoding runs and models under 7B parameters are left out.
- How much do Strix Halo and DGX Spark cost in Singapore?
- Strix Halo from S$4,329 and DGX Spark from S$7,452, the lowest listed price across the sellers TokenMark reads daily. The listings are different configurations and memory sizes, so check the where-to-buy list on each hardware page.
Strix Halo hardware page · DGX Spark hardware page
Other comparisons: Strix Halo vs Mac Ultra.
Last updated: snapshot 2026-10-29; newest source check 2026-10-29.
More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.