NVIDIA DGX Spark FE 128 GB
Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
- llama-2-7b · dgx-spark-128g
llama.cpp (CUDA) · Q4_0 · tg128
Recorded 2026-10-12 · historical source record · repeat count unknownSourced
Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table. Run source - gpt-oss-20b · dgx-spark-128g
llama.cpp (CUDA) · MXFP4 · tg32 @ 8k depth
Recorded 2026-10-12 · historical source record · repeat count unknownSourced
llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark. Run source - gpt-oss-120b · dgx-spark-128g
llama.cpp (CUDA) · MXFP4 · tg32 @ 8k depth
Recorded 2026-10-12 · historical source record · repeat count unknownSourced
llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark. Run source - qwen3-coder-30b-a3b-instruct · dgx-spark-128g
llama.cpp (CUDA) · Q8_0 · tg32 @ 8k depth
Recorded 2026-10-12 · historical source record · repeat count unknownSourced
llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark. Model is Qwen3-Coder-30B-A3B-Instruct. Catalogue correction 2026-10-13: model identifier corrected from Qwen3-30B-A3B; shared architecture does not establish identical performance. Run source
Evidence labels · Freshness policy · Calculation methodology
CUDA on 128 GB of unified memory. Similar bandwidth to Strix Halo, more compute for prefill.
Decode-speed ceiling for a model is 273 GB/s ÷ GB read per token. For a 16.6 GB file that is 16 tok/s; real runtimes reach roughly 45% of it. Why bandwidth is the limit.
Specifications
Memory | 128 GBSourcedVendor spec page, checked 2026-10-12. Source |
|---|---|
Bandwidth | 273 GB/sSourcedVendor spec page, checked 2026-10-12. Source |
FP16 dense compute | 100 TFLOPSSampleSample FP16 dense compute; pending vendor-sheet citation. |
Vendor TDP / board power | 140 WSourcedGB10 TDP (CPU+GPU), nvidia.com; system power supply is 240 W. Source |
Load power at wall | 170 WSampleSample wall power under load. |
Price, USD | $4,699SourcedUS$4,699 (secondary source). Source |
Price, CAD | CA$6,706EstimateUSD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured. Source |
Models that run
Default quant, single stream, 8k context. Fastest first; estimates are ±25% until measured.
Model | Quant | Memory | Verdict | Decode, tok/s | Max context |
|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | 12.9 GB | Fits | 75.2Sourced | 128k |
| gpt-oss-120b | MXFP4 | 64.3 GB | Fits | 52.9Sourced | 128k |
| Qwen3 30B-A3B | Q4_K_M | 19.9 GB | Fits | 43.7Estimate | 32k |
| Llama 3.1 8B Instruct | Q4_K_M | 6.6 GB | Fits | 20.5Estimate | 128k |
| Qwen3 8B | Q4_K_M | 6.8 GB | Fits | 19.7Estimate | 32k |
| Qwen3 14B | Q4_K_M | 10.9 GB | Fits | 11.9Estimate | 32k |
| Mistral Small 3.2 24B | Q4_K_M | 16.2 GB | Fits | 7.8Estimate | 128k |
| Gemma 3 27B IT | Q4_K_M | 18.2 GB | Fits | 7.0Estimate | 128k |
| Qwen3 32B | Q4_K_M | 22.5 GB | Fits | 5.6Estimate | 32k |
| Llama 3.3 70B Instruct | Q4_K_M | 45.8 GB | Fits | 2.7Estimate | 128k |
Published reference runs
Third-party llama.cpp measurements on this device, with conditions and source. Llama 2 7B Q4_0 runs are the community calibration test; they are not catalogue models.
Model | Quant | Date | Runtime · test | Decode, tok/s | Conditions |
|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | 2026-10-12 | llama.cpp (CUDA) · tg128 | 57.2Sourced | Llama 2 7B Q4_0, llama-bench tg128, full GPU offload, flash attention off, single GPU. Community scoreboard entry; commit and submitter in the source table. Source |
| gpt-oss-20b | MXFP4 | 2026-10-12 | llama.cpp (CUDA) · tg32 @ 8k depth | 75.2Sourced | llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark. Source |
| gpt-oss-120b | MXFP4 | 2026-10-12 | llama.cpp (CUDA) · tg32 @ 8k depth | 52.9Sourced | llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark. Source |
| qwen3-coder-30b-a3b-instruct | Q8_0 | 2026-10-12 | llama.cpp (CUDA) · tg32 @ 8k depth | 48.0Sourced | llama-bench -fa 1 -p 2048 -n 32 -ub 2048, ggml-org GGUF, commit 3795cc1e8 (2026-02-05), run by the llama.cpp maintainers on a DGX Spark. Model is Qwen3-Coder-30B-A3B-Instruct. Catalogue correction 2026-10-13: model identifier corrected from Qwen3-30B-A3B; shared architecture does not establish identical performance. Source |