DGX Spark alternatives: which machine fits your models?
A DGX Spark has 128 GB of unified memory at 273 GB/s. Most alternatives trade one of those for the other: more bandwidth and less memory, or similar memory for less money. The table below sums up eight comparison pages, one row per machine. Every figure in a row comes from the linked page, where the sources and dates are.
Eight machines against the Spark
| Machine | Memory | Bandwidth | Where it wins | Where the Spark wins |
|---|---|---|---|---|
| NVIDIA H100 SXM | 80 GB HBM3 | 3.35 TB/s | Throughput per request, and batch serving of dense 70B models to many users | Memory per dollar; models above about 70 GB for one user; cheap evaluation by the minute |
| NVIDIA H200 SXM | 141 GB HBM3e | About 4.8 TB/s | A shared production endpoint with real concurrent traffic; close to 18 times the bandwidth | Models under about 100 GB for one person or a small team; development and evaluation |
| RTX PRO 6000 Blackwell | 96 GB GDDR7 ECC | 1,792 GB/s | Speed on models that fit in 96 GB; a reported 215.2 against 60.9 tok/s on gpt-oss-20b | Models of 100 GB or more; a complete box instead of a card that needs a host PC |
| GeForce RTX 5090 | 32 GB GDDR7 | 1,792 GB/s | Anything under about 30 GB; reported 575 tok/s against 25 to 32 tok/s on a Qwen3-class 27B at concurrency 8; gaming | Models that will not load on 32 GB, such as 70B and gpt-oss-120b; a complete 240 W system |
| Mac Studio (M5 Ultra) | 96 to 512 GB unified | 1.2 TB/s | Token generation on models that fit; more than 128 GB in one box | Prefill and long prompts (3.8 times faster than an M3 Ultra in EXO Labs' test); CUDA; two-unit clustering |
| AMD Strix Halo (HP Z2 Mini G1a) | 128 GB LPDDR5x | 256 GB/s | Price ($2,949 as tested); a near tie on single-user token generation in llama.cpp | 2 to 3 times faster time to first token; about 2 times faster fine-tuning; about 2.5 times on FLUX.1; CUDA |
| ASUS Ascent GX10 | 128 GB LPDDR5x | 273 GB/s (same chip) | Same GB10 chip; compare price and stock for the storage tier you need on the day you order | The Founders Edition gets DGX OS releases from NVIDIA directly |
| Jetson AGX Thor | Not stated; that page does not verify Thor specs | Not stated | Robotics and embedded edge inference inside another product | LLM development and serving on a desk; two ConnectX-7 ports for a two-unit cluster |
Each row is taken from the linked comparison page. Rival speed figures are third-party reports cited there, not GPUwerk measurements. The Spark's own figures are on the specs page.
How we compare
I run GPUwerk, which rents DGX Sparks and none of the other machines here, so read every page with that in mind. Spec figures come from each vendor's own pages. Spark speeds are GPUwerk's own, measured with vLLM 0.27.1 on one fleet node on 2 September 2026, with raw logs on the benchmarks page. GPUwerk has not run any of the rivals, so their speeds are third-party reports, named on each page, or bandwidth arithmetic labelled as such. Prices carry the date they were read. Purchase prices have moved a lot this year: the 128 GB Spark lists at $6,950 in the US, and the price page has the source, date and history.
Pick by job, not by spec
- A shared endpoint with many concurrent users: a datacenter GPU such as the H100 or H200.
- A model that fits in 32 GB or 96 GB, used daily: a fast card, the RTX 5090 or RTX PRO 6000.
- Models of 100 GB or more, CUDA, long prompts: the Spark. The GB10 machines comparison covers the other makers' versions.
- Not sure yet: rent a Spark by the minute and measure your own model before you buy anything.