NVIDIA · Local AI machine
NVIDIA DGX Spark (128 GB)
NVIDIA machine: NVIDIA GB10 Grace Blackwell Superchip, 128 GB memory, 273 GB/s memory speed, about 4.2 tok/s on a 70B model (estimate).
82.3Score
- #6 of 41 ranked
- #1 of 9 NVIDIA machines
Why this score
The score adds up these parts. Each part has a weight. A part counts as the average of its sources, pulled toward 80 when there are few sources.
AI power
50% weightcounts as 82.4Average of 2 sources: 83.6
Experts
30% weightcounts as 82.4Average of 2 sources: 83.6
Users
20% weightcounts as 81.9Average of 1 source: 83.8
Good: 3 reviews and buyer ratings.
80 is the field average. A part counts as 80 + (average − 80) × N ÷ (N + 1), where N is its number of sources. One source counts half of the gap from 80. No sources count as 80.
Price
€6922.98 at BA-Computer · Geizhals.at listing, not checked live
Founders Edition, 128 GB with 4 TB SSD (940-54242-0005-000), cheapest of 13 Geizhals Austria offers. e-tec.at, DiTech.at and Alternate.at ask EUR 6,923.
Prices checked 8 Oct 2026.
Specs
- Chip
- NVIDIA GB10 Grace Blackwell Superchip
- CPU
- 20 cores (10 Cortex-X925 + 10 Cortex-A725)
- Memory
- 128 GB
- GPU memory
- 128 GBNVIDIA GB10: 128 GB LPDDR5X coherent unified memory shared by CPU and GPU (NVIDIA spec page). No separate GPU limit is documented; the OS draws from the same pool. Linux (DGX OS) only.
- Memory speed
- 273 GB/s
- FP16 TFLOPS
- —
- NPU
- —
- OS
- Linux
- Software
- CUDA
- Ethernet
- 10 Gbit/s
- Thunderbolt / USB4
- —
- QSFP port
- Yes
- Idle power
- 35 W
- Load power
- 160 W
- Noise
- 37.5 dB(A)
- Volume
- 1.1 L
- Weight
- 1.2 kg
- SSD
- 4 TB
- Release
- 2025-10
Measured
- gpt-oss-120b, MXFP4
- 55.4 tok/s writing, 1725.5 tok/s prompt · llama.cpp (NVIDIA test, ISL 2048 / OSL 128, batch 1)
- gpt-oss-120b, MXFP4
- 60.6 tok/s writing, 1956 tok/s prompt · llama.cpp CUDA build 6922, llama-bench pp2048 / tg32
- Qwen3 Coder 30B A3B, Q8_0
- 60 tok/s writing, 2933.4 tok/s prompt · llama.cpp b6767, llama-bench pp2048 / tg32
- Llama 3.1 70B (dense), FP8
- 2.7 tok/s writing, 803 tok/s prompt · SGLang
Notes
Runs 100B-class MoE models such as gpt-oss-120b at about 55-60 tok/s, with full CUDA support (Tom's Hardware: new workflows run on day one) and a 200 Gbit ConnectX-7 for clustering two units. Dense 70B models decode at only about 3 tok/s because 273 GB/s of memory bandwidth is the limit; it is Linux only and the price has risen sharply.
gpt-oss-120b decode differs by build: 55.37 (NVIDIA blog), 52.87 (JetsonHacks, b6767), 60.57 (llama.cpp discussion, b6922 after a kernel fix); a hands-on Ollama test cited in a search excerpt gave about 38.7. Idle power: Tom's Hardware 35 W headless (and 22-25 W after a later NVIDIA update per a search excerpt of Tom's article, not confirmed on the page), ServeTheHome 40-45 W. Load: Tom's 160 W typical at the wall, ServeTheHome just under 200 W with CPU and GPU together. NVIDIA US price rose from USD 4,699 to USD 6,950-6,995 (HotHardware 6,995); an undated forum post earlier cited EUR 4,800 for the EU, now superseded by Geizhals Austria EUR 6,922.98. Starry Hope's spec table lists the SSD as 2 TB but its body text says 4 TB; Tom's Hardware and Geizhals confirm 4 TB for the Founders Edition. HotHardware's score comes from the page's review markup. The Amazon rating was read via Starry Hope.
More reading:servethehome.comlmsys.orgstoragereview.comtheregister.comyoutube.comhardwareluxx.decomputerbase.deheise.dedigit.intheregister.com