DGX Spark vs an RTX PRO 6000 server

DGX Spark has 128 GB at 273 GB/s on your desk. An 8x RTX PRO 6000 server has 768 GB at 1.6 TB/s per card. Specs, benchmarks and prices compared.

An NVIDIA DGX Spark, a small gold-coloured desktop computer with a perforated metal front
Photo: Nvidia DGX front view dllu by Daniel Lu (User:dllu), Wikimedia Commons, CC BY-SA 4.0

NVIDIA DGX Spark is a 1.2 kg desktop computer with 128 GB of memory that runs models up to 200 billion parameters. A server with eight RTX PRO 6000 Blackwell cards has six times the memory and about six times the memory bandwidth per card. Both are Blackwell and both run the same NVIDIA software, so the question is how big your model is and how many people use it.

Which one for whom

  • One developer, prototyping, fine-tuning small models on a desk → DGX Spark. 128 GB of unified memory, up to 1 PFLOP of FP4 (with sparsity), 240 W from a wall socket. Price on NVIDIA Marketplace on 10 November 2026: $6,950, shown as out of stock. A 64 GB version from Acer, ASUS, Dell, Gigabyte, HP and MSI starts at $4,999.
  • A team, many users at once, or models above 200B → a server with eight RTX PRO 6000. 8 × 96 GB of GDDR7, 1.6 TB/s per card, up to 600 W per card in a data centre. Our server in Sofia rents from $1.80 per GPU-hour, or about $10,510 a month for the whole machine. It is rented into early 2027 and we take reservations for the period after. See the server.

The specifications side by side

DGX Spark (128 GB) 1 × RTX PRO 6000 Server Edition 8 × RTX PRO 6000 server
Chip GB10 Grace Blackwell superchip Blackwell GPU 8 Blackwell GPUs
Memory 128 GB LPDDR5X, shared by CPU and GPU 96 GB GDDR7 with ECC 768 GB GDDR7
Memory bandwidth 273 GB/s 1,597 GB/s 1,597 GB/s per card
FP4 up to 1 PFLOP (sparse) 4 PFLOPS (1) 8 × 4 PFLOPS
CPU 20 Arm cores the server's x86 CPUs the server's x86 CPUs
Network ConnectX-7, 200 Gbps data centre uplink
Power 240 W supply, GB10 TDP 140 W up to 600 W about 6.5 kW at full GPU load (2)
Size 150 × 150 × 50.5 mm, 1.2 kg dual-slot PCIe card 4U rack server
Largest model, NVIDIA's figure up to 200B parameters

(1) NVIDIA's page for the Server Edition does not say whether 4 PFLOPS assumes sparsity. The DGX Spark page says its 1 PFLOP does. (2) Measured from the wall on our own server. We published the measurements in What an 8x RTX PRO 6000 server really draws.

Memory bandwidth decides how fast answers come out

When a model writes an answer, the GPU reads the model weights for every new token. That makes memory bandwidth the limit for one user. DGX Spark has 273 GB/s. One RTX PRO 6000 has 1,597 GB/s, almost six times more.

The LMSYS team measured this when DGX Spark came out (October 2025). On gpt-oss-20b with Ollama, one request at a time, DGX Spark wrote 49.7 tokens per second and an RTX PRO 6000 Blackwell 215 tokens per second. Prefill, the reading of the prompt, was 2,053 against 10,108 tokens per second. LMSYS put the gap down mainly to the memory bandwidth.

On a 70B model the difference is easier to feel. LMSYS ran Llama 3.1 70B in FP8 with SGLang on DGX Spark at 2.7 tokens per second for decoding. That is slower than most people read. The model fits in 128 GB, but it is a test, not a service. LMSYS's own conclusion is that the Spark suits prototyping and experiments, not production.

Memory size decides which models fit

As a rough rule, model weights take one byte per parameter in FP8 and half a byte in FP4, plus room for the KV cache of each conversation.

  • Up to about 30B parameters: fits on DGX Spark and on one RTX PRO 6000 with room for long contexts.
  • 70B: about 70 GB in FP8. It fits on both, but on DGX Spark it runs at the speed above, and little memory is left for many users.
  • 120-200B in FP4: fits on one DGX Spark (NVIDIA says up to 200B), or on two to four RTX PRO 6000 cards.
  • 400B and above: NVIDIA says two linked DGX Spark units handle up to 400B, and four up to 700B through a 200 GbE switch. Eight RTX PRO 6000 cards hold 768 GB in one machine, enough for a 671B model such as DeepSeek-V3 in FP4 (about 336 GB of weights) with room for the cache.

We compare running models on a laptop or one card with a server in Local LLM vs private LLM server.

Many users: where a server pulls ahead

A desk computer serves one person well. A team of 20 developers or a product with customers sends many requests at once. Batching helps both machines: LMSYS saw Llama 3.1 8B on DGX Spark go from 20.5 tokens per second with one request to 368 with 32 requests. But the ceiling is set by memory and bandwidth, and a server has eight cards, each with six times the bandwidth.

On eight cards you can also run several models side by side, for example a large model on four cards and smaller ones on the others, or split a card into up to four isolated MIG instances. A server stays on 24 hours a day in a data centre with power, cooling and a fixed address. A desktop needs someone to keep it running.

For serving software, see vLLM vs Ollama vs SGLang.

Price: buying a box against renting a machine

On 10 November 2026 NVIDIA Marketplace lists the 128 GB DGX Spark Founders Edition at $6,950 and shows it as out of stock. The price has gone up twice. The list price was $3,999 until February 2026, when NVIDIA raised it to $4,699 and said the reason was worldwide constraints in memory supply. NVIDIA then announced the 64 GB version on 2 October 2026, from $4,999 through partners, on sale from 23 October.

A DGX Spark makes sense when one person needs a machine for a long time. When the work needs more than 128 GB or more speed, two or four linked units still share the 273 GB/s per unit. Renting a dedicated GPU server by the month gives the whole machine to one team at a fixed price, without buying hardware whose price moves with the memory market. You can compare the options on our pricing page.

Questions

Is DGX Spark the same as Project DIGITS? Yes. NVIDIA showed it as Project DIGITS on 6 January 2025 and announced it as DGX Spark on 18 March 2025.

Does DGX Spark run the same software as a server? It runs DGX OS, an Ubuntu-based Linux, with the NVIDIA AI software stack. NVIDIA presents it as a way to build on the desk and then move the work to DGX Cloud or a data centre.

Can I fine-tune on DGX Spark? NVIDIA says up to 70B parameters. For larger models or faster runs, a server with more memory and bandwidth finishes the same job sooner.

What is the difference between the 64 GB and the 128 GB DGX Spark? The same GB10 chip and software, half the memory. NVIDIA says the 64 GB version runs models up to 100B, and two linked units up to 200B.

Sources

SHARE