Computer Hardware Personal AI hardware
Rank 05
NVIDIA DGX Spark
Still the CUDA-native desk AI appliance: 128 GB coherent memory and ConnectX-7 clustering in a 1.1 L box. It is now second in this subcategory, because the M5 Ultra Studio generates tokens much faster for single users.
Photograph by Daniel Lu, CC BY-SA 4.0, via Wikimedia Commons, resized
Why it is here
Formerly Project DIGITS (CES 2025), renamed DGX Spark at GTC March 2025, Founders Edition shipping since 2025-10-15. LMSYS measured Llama 3.1 8B FP8 at 7,991 tok/s prefill / 20.5 tok/s decode (batch 1) scaling to 368 tok/s aggregate at batch 32; Llama 3.1 70B FP8 at 803 tok/s prefill / 2.7 tok/s decode; gpt-oss-20b MXFP4 via Ollama at 2,053 / 49.7 tok/s. StorageReview vLLM: Llama 3.1 8B FP4 ~924 tok/s at concurrency 128; Qwen3 Coder 30B-A3B FP8 ~46.5 tok/s (c=1) to 482.6 tok/s (batch 64); gpt-oss-120b NVFP4 31.4 tok/s (c=1). llama.cpp maintainer benches: gpt-oss-120b MXFP4 ~58.7 tok/s generation / ~2,444 tok/s prefill. NVIDIA claims 1 PFLOP sparse FP4 and models up to 200B (405B on two clustered units). Memory is the product: 128 GB LPDDR5X at only 273 GB/s. Tom's Hardware: beats Ryzen AI Max+ 395 on prefill, long-context decode, Flux.2 Klein (Spark ~60 s vs Strix Halo ~4× slower), idle ~35 W (later firmware ~22–25 W headless), load ~160 W, 37.5 dBA, GPU ~82 °C. No MLPerf Client submission found.
Who it is for
CUDA developers, labs, and researchers who need to prototype/fine-tune 8B–70B and run MoE 120B locally with a path to DGX Cloud — not buyers chasing fastest single-stream dense-70B chat.
Sources
- DGX Spark review (Jan 2026)independent professional review; llama.cpp vs Strix Halo, thermals, power
- DGX Spark review (Oct 2025, updated 2026-08-16)independent professional review; vLLM tok/s, MAMF, GDS
- DGX Spark in-depth review (2025-10-13)independent lab SGLang/Ollama tok/s vs RTX 5090 / PRO 6000
- DGX Spark reviewindependent professional hands-on; clustering/networking
- 2026-02-23 price changeofficial $3,999 → $4,699
- DGX Spark prices are already climbing before RTX Spark mini PCs launchOut of stock on NVIDIA Marketplace at $4,699; cheapest available units $4,999.99 (Amazon / Micro Center); ASUS GX10 about $6,000.
