AI ComparisonsNov 21, 20269 min read

DGX Spark vs RTX PRO 6000: Bigger Models or Faster Answers

DGX Spark vs RTX PRO 6000: the Spark holds 128 GB read at 273 GB/s, the card 96 GB at 1,792 GB/s. One fits the bigger model, the other answers faster.

On this page
  1. The two machines at a glance
  2. Which one holds the bigger model?
  3. Why does memory bandwidth set the speed of the answer?
  4. What do measured numbers show?
  5. What did we run on the card?
  6. What does each cost, and when does buying beat renting?
  7. Can you rent an RTX PRO 6000 in Canada?
  8. Where each one falls short
  9. FAQ
  10. Sources

The DGX Spark holds more model and the RTX PRO 6000 graphics card answers faster. The Spark has 128 gigabytes (GB) of memory, read at 273 GB per second. The card has 96 GB read at 1,792 GB/s. On 9 November 2026 the Spark listed at $6,950 on NVIDIA Marketplace. On 11 November the card cost $16,399 at Newegg, before the computer it goes in.

If the model is bigger than 96 GB and one person uses it, pick the Spark.

If the model fits in 96 GB and people wait on its answers, pick the card.

If you need it for one project, rent the card by the hour first.

The two machines at a glance

The Spark is a whole computer, 150 mm square and 50.5 mm tall. The card needs a workstation or server around it. Throughout this post, "GPU memory" means the fast memory the chip reads the model from while it works.

DGX SparkRTX PRO 6000 Workstation EditionRTX PRO 6000 Server Edition
What it isComplete desktop computer (GB10 chip)Graphics card for a workstationGraphics card for a data centre server
GPU memory128 GB LPDDR5X, shared with the CPU96 GB GDDR796 GB GDDR7
Memory bandwidth273 GB/s1,792 GB/s1,597 GB/s
Peak AI compute (FP4, NVIDIA's sparse figure)1 PFLOP4,000 TOPS4 PFLOPS
Power140 W chip, 240 W supply600 WUp to 600 W
Price (dated)$6,950, NVIDIA Marketplace, 9 Nov 2026$16,399, Newegg, 11 Nov 2026, card onlyRent: AWS g7e.2xlarge, $3.363/h in us-east-1
Rentable in Canada from the big threeNot a cloud productNoNo

Specs are from NVIDIA's own pages for the DGX Spark, the Workstation Edition and the Server Edition. The rental is the Server Edition: AWS lists 96 GB at 1,597 GB/s for the g7e family.

Which one holds the bigger model?

The Spark holds the bigger model: 128 GB against 96 GB. NVIDIA says one Spark runs inference on models up to 200 billion parameters.

The arithmetic behind that claim is simple. A model stored at 4 bits per parameter takes half a byte per parameter, so 200 billion parameters is about 100 GB of weights. That fits in 128 GB and does not fit in 96 GB. The Spark's 128 GB is also the computer's main memory, so the operating system takes a share. Long conversations need working room on top of the weights too.

For a model in the 100 to 120 GB range, the Spark runs it on one box and the card does not. To run it on cards you need two, which AWS sells as the g7e.12xlarge with 192 GB across two GPUs. For anything that fits in 96 GB with room to spare, the memory size stops deciding and speed takes over. Our sibling post on how much GPU memory an open model needs works through sizes model by model.

Why does memory bandwidth set the speed of the answer?

A model writes its answer one token at a time, about three quarters of a word each, and for every token it reads its weights out of memory again. Memory bandwidth is how many gigabytes per second the chip can read. So the faster the memory, the faster the words arrive.

Bandwidth is therefore the ceiling on writing speed for one user. The card reads memory 6.6 times faster than the Spark. The Server Edition you rent on AWS reads it 5.9 times faster. Real speeds land below that ceiling, because software, batch size and the model's design all take a cut. A mixture-of-experts model, which only reads the part of itself it needs for each token, writes faster on both machines than a dense model of the same size.

Reading the prompt is a different job. Before it writes, the model reads your whole input at once, and that part is limited by compute, not memory. The card leads there too, with about four times the Spark's quoted peak compute.

What do measured numbers show?

LMSYS, the research group behind the SGLang serving software, timed both on 13 October 2025. The test model was OpenAI's gpt-oss-20b, stored at 4 bits per weight. Ollama served it to one user at a time.

Measured by LMSYS, 13 Oct 2025DGX SparkRTX PRO 6000 Blackwell (MXFP4 weights, Ollama, one user)
Reading the prompt (tokens per second)2,05310,108
Writing the answer (tokens per second)49.7215

The card wrote 4.3 times faster in that test, below the 6.6 times its bandwidth allows and in the same direction. Read it as one model, one serving tool and one user, a year old. Serving software has improved since, on both machines. We have not timed a Spark ourselves.

What did we run on the card?

Our DecideBench v2 self-hosted runs, published 9 November 2026, used one of these cards, rented from AWS at $3.36 an hour. That benchmark tests decision and classification tasks only, so these numbers say nothing about chat or coding.

The largest model we served there was pplx-decider, a 27-billion-parameter model whose card asks for about 52 GiB of memory for its weights. It answered all 450 test items in 186.1 seconds with 4 requests in flight. Its Flash version did the same 450 in 60.4 seconds. Both fit in 96 GB with room for several requests at once. For models like these, the card's speed pays off and the Spark's extra memory would sit unused.

What does each cost, and when does buying beat renting?

The Spark costs $6,950 as a complete computer. The card costs $16,399 on its own, 2.4 times the Spark before you buy the workstation, power supply and cooling for a 600 W card. Renting the card on AWS costs $3.363 an hour.

The break-even, as plain division:

ComparisonArithmeticHours of rental
Rent the card vs buy a Spark$6,950 ÷ $3.363 per hour2,067 hours
Rent the card vs buy the card$16,399 ÷ $3.363 per hour4,876 hours

2,067 hours is 86 days running around the clock, or about a year of 40-hour weeks. If you need the card for less than that, renting costs less than a Spark and gives you the faster chip. The comparison leaves out what sits on each side: storage and data transfer on AWS, and on the other side the power bill, the workstation around a card, and the hours someone spends keeping a box running. Both prices have moved this year, so date any quote you get.

Can you rent an RTX PRO 6000 in Canada?

No. As of 11 November 2026 none of the big three clouds lists this card in a Canadian region, so owning the hardware in Canada is the route to keeping this class of GPU inside the country.

  • AWS: the g7e family costs $3.363 an hour in Northern Virginia and is not offered in Montreal, per AWS's price list read on 11 November 2026. Its price list shows no GPU machines in Calgary.
  • Azure: its Canada Central catalogue has no size with this card, per our read of Azure's VM sizes on 11 November 2026.
  • Google Cloud: its G4 machines carry the card, and its GPU regions page lists none in Montreal or Toronto, which offer the L4 and, in Toronto, the H100.

Owning one puts inference wherever the hardware sits: a workstation with the card in a Canadian office or data centre, or a DGX Spark on a desk. You buy the hardware, pay for its power and keep it patched. A single box on a desk is not a redundant production server, so if it fails, the work waits until it is fixed. If you would rather rent inside Canada, the nearest rentable chips are the L4 and H100, compared in RTX PRO 6000 vs H100 and in the Canadian residency guide.

XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account or on-premises, sizes the hardware to the model, and builds the workflow on top. The on-premise, client cloud or managed cloud post sets out how to choose between them, and how we deploy covers the rest. If you need a model running with inference kept in Canada, bring one job to a free 30-minute call.

Where each one falls short

The Spark is slow at writing for its size class. It can hold a 100 GB model, and every token of that model waits on 273 GB/s. A desk box shares one power socket and one network cable with everything else in the office.

The card is fast and expensive, and 96 GB is a hard limit. A model just over it needs a second card, not a setting. It draws up to 600 W, so the workstation around it needs the power supply and airflow to match. Renting it avoids both problems and puts the model in a US region.

We have not measured either machine on chat or coding work, and we have not timed a Spark at all. For other desk-sized options see DGX Spark alternatives. For the next step up, DGX Spark vs DGX Station.

FAQ

Is the DGX Spark faster than an RTX PRO 6000?

No. The card reads its memory at 1,792 GB/s against the Spark's 273 GB/s, and memory speed sets how fast a model writes. In LMSYS's October 2025 test on gpt-oss-20b, the card wrote 215 tokens per second and the Spark 49.7. The Spark's advantage is size: 128 GB of memory against 96 GB, so it holds models the card cannot.

Can a DGX Spark run a 200B model?

NVIDIA says one 128 GB Spark runs inference on models up to 200 billion parameters. At 4 bits per parameter, 200 billion parameters is about 100 GB of weights, which fits. It will write slowly, because every token reads those weights at 273 GB/s. Two linked Sparks reach 400 billion parameters on NVIDIA's figures, and the 64 GB Spark tops out at 100 billion.

How much does an RTX PRO 6000 cost to rent?

AWS rents one Server Edition card as the g7e.2xlarge at $3.363 an hour on demand in us-east-1, read on 11 November 2026. That size comes with 8 vCPUs and 64 GiB of system memory. AWS does not offer the g7e family in its Montreal region, and Azure and Google Cloud list none in Canada either.

Should I buy a DGX Spark or rent a GPU?

Divide the Spark's price by the hourly rent. At $6,950 and $3.363 an hour, the break-even is 2,067 hours, about a year of 40-hour weeks. Below that, renting the faster card costs less. Above it, owning wins, and owning is also the way to keep inference in Canada for this class of hardware, since the big three don't rent it there.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…