DGX Spark vs RTX PRO 6000 Blackwell for local LLMs
The RTX PRO 6000 Blackwell has 1,792 GB/s of memory bandwidth against the Spark's 273 GB/s, about 6.6 times more, so it decodes several times faster on any model that fits in its 96 GB. The Spark holds 32 GB more (128 GB against 96 GB), is a whole computer rather than a card, and rents by the minute for $0.79 an hour. I run GPUwerk, which rents DGX Sparks and does not rent a PRO 6000, so weigh everything below with that in mind.
Side by side
| DGX Spark (GB10) | RTX PRO 6000 Blackwell | |
|---|---|---|
| Memory | 128 GB LPDDR5x, unified (advantage) | 96 GB GDDR7 ECC |
| Memory bandwidth | 273 GB/s | 1,792 GB/s (about 6.6 times) (advantage) |
| Form factor | Whole system: CPU, memory, storage, network (advantage) | Dual-slot PCIe Gen 5 x16 card; needs a host PC |
| Power | 240 W power supply for the whole system (NVIDIA) (advantage) | 600 W for the card alone (Workstation Edition); 300 W for the Max-Q Edition (NVIDIA) |
| Decode ceiling, 70B model at 4-bit (37 GB) | About 7.4 tok/s; GPUwerk-measured 6.0 tok/s (Llama 3.3 70B AWQ) | About 48 tok/s. Arithmetic ceiling, not a measurement (advantage) |
| Reported gpt-oss-20b decode | 60.9 tok/s (reported) | 215.2 tok/s (reported) (advantage) |
| Fits gpt-oss-120b at MXFP4 | Yes | Yes |
| Fits GLM-5.3-Flash at 2-bit (about 107 GB loaded) | Yes, with little headroom (GPUwerk runs a smaller 93 GB build) (advantage) | No, 96 GB is below the weight size |
| Buy price | Check vendor; GPUwerk store lists the Lenovo PGX from €5,330 excl. VAT | Check vendor; I did not verify a retailer price today |
| Rent options | GPUwerk, $0.79/h, per minute | Hetzner GEX131: hourly rate up to a monthly maximum, plus a setup fee, no minimum term; GPUwerk does not rent one |
| Where it wins | Models above 96 GB, a complete box, rented by the minute | Speed on models that fit in 96 GB; lower latency per request |
Checked 2026-10-11. PRO 6000 memory, bandwidth, power and form factor: NVIDIA RTX PRO 6000 Blackwell Workstation Edition page. Max-Q Edition (same 96 GB and 1,792 GB/s, 300 W): NVIDIA Max-Q page. Spark figures are NVIDIA's (128 GB, 273 GB/s, 240 W power supply; see the specs page); the 6.0 tok/s is GPUwerk-measured with vLLM 0.27.1 on 2 September 2026, raw log, method on the benchmarks page. The two decode ceilings are bandwidth divided by 37 GB of weights (273 / 37 and 1,792 / 37); real engines land below the ceiling, and GPUwerk has not run a PRO 6000. The gpt-oss-20b figures are reported, not GPUwerk-measured: they come from the LMSYS spreadsheet (Ollama, MXFP4, batch 1) as quoted by AIMultiple, which I read on 2026-10-11; neither the spreadsheet nor AIMultiple gives a run date. gpt-oss-120b size is my arithmetic (117B parameters at about 4.25 bits, roughly 62 GB). GLM-5.3-Flash loads at roughly 107 GB at UD-Q2_K_XL, per GPUwerk's model page. Hetzner: GEX131 page, fetched 2026-10-11. Street prices for the PRO 6000 and the Spark are left as "check vendor" because I could not fetch a retailer page today.
What the speed gap looks like
Generating a token means reading the active weights from memory once, so single-stream decode is roughly bandwidth divided by active-parameter bytes. For a 70B model at 4-bit (about 37 GB) that is 7.4 tok/s on a Spark and about 48 tok/s on a PRO 6000. GPUwerk measured 6.0 tok/s for Llama 3.3 70B AWQ on a Spark; the 48 is a ceiling I computed, and a real engine would come in lower.
The one reported head-to-head I found is smaller than the bandwidth ratio. On gpt-oss-20b, a small mixture-of-experts model, the PRO 6000 shows 215.2 tok/s and the Spark 60.9 tok/s, about 3.5 times. gpt-oss-20b reads only about 3.6B active parameters per token, so fixed per-token overhead takes a larger share of the time and the extra bandwidth counts for less. My expectation, not a measurement: the gap grows toward the 6.6 times bandwidth ratio as the active weights grow, as with a dense 70B model.
For a single user on a model that fits in 96 GB, the PRO 6000 is the faster machine. The Spark's case is capacity and packaging. See the RTX 5090 comparison for the same trade at 32 GB, and the H100 comparison for 80 GB at 3.35 TB/s.
Which models fit where
Fit is arithmetic: weights plus KV cache plus runtime overhead must stay under the memory size. A Spark addresses about 121.6 GiB; the PRO 6000 has 96 GB. The rules on GPUwerk's models page apply to both: a 70B model is about 140 GB at FP16, 70 GB at FP8 and 37 GB at 4-bit.
- Both fit: 70B dense at 4-bit or FP8 (70 GB leaves little cache on the PRO 6000), gpt-oss-120b at MXFP4 (about 62 GB by my arithmetic), and every 30B-class model. On these the PRO 6000 is faster.
- Spark only: weights between roughly 85 GB and 110 GB. GLM-5.3-Flash at UD-Q2_K_XL (about 107 GB loaded) is the example from GPUwerk's own testing, and it already left too little headroom on a Spark, which is why GPUwerk now runs the 93 GB UD-IQ1_S build.
- Neither: anything whose weights plus cache pass the Spark's roughly 121.6 GiB, such as a 235B-parameter model at 4-bit (about 125 GB of weights by the same arithmetic). A two-node Spark cluster gives 256 GB across two machines, not one pool.
The Spark specs page has the memory details. The gap between 96 GB and 128 GB is narrow, so most models people run fit both. The extra 32 GB matters at the top of the range, and for cache: long contexts and several concurrent sessions eat memory quickly.
Rent before deciding
GPUwerk rents a dedicated Spark at $0.79 an hour, billed per minute from prepaid credit. An hour of your own prompts on a Spark tells you whether a model's quality is enough, and whether 6 tok/s or 60 tok/s is tolerable for your use. The first top-up is matched up to $10.
GPUwerk does not rent a PRO 6000. In the EU, Hetzner's GEX131 carries a PRO 6000 Blackwell Max-Q 96 GB in Helsinki and Falkenstein, with no minimum contract term. Hetzner's page quotes a monthly maximum, an hourly rate and a one-time setup fee, but the fetch on 11 October 2026 did not return the figures, so check the page. The setup fee makes a short test cost more than the hourly rate alone suggests. NVIDIA lists the Max-Q Edition with the same 96 GB and 1,792 GB/s as the Workstation Edition, at 300 W instead of 600 W. My Hetzner GPU comparison covers that provider in more detail.
When to pick which
- Model fits in 96 GB and you use it daily: PRO 6000, if you can host and power a 600 W card. It is several times faster per request.
- Model is 100 GB or more: Spark. The PRO 6000 cannot load it.
- You want one box that works when plugged in: Spark. The card needs a PC around it.
- You have not picked a model yet: rent a Spark by the minute first. The fit table shows what loads, and the benchmarks page shows what GPUwerk measured.
On speed per euro of hardware the PRO 6000 may well come out ahead for models under 96 GB; I have no verified retail price today to settle it, and that comparison leaves out the models only a 128 GB machine loads. The before you buy guide covers the buying side.
FAQ
Is an RTX PRO 6000 faster than a DGX Spark?
Yes, on any model that fits in its 96 GB. NVIDIA lists 1,792 GB/s of memory bandwidth for the RTX PRO 6000 Blackwell Workstation Edition against the Spark's 273 GB/s, about 6.6 times more, and decode speed tracks bandwidth. A reported gpt-oss-20b run (Ollama, MXFP4, batch 1, from the LMSYS spreadsheet as quoted by AIMultiple) shows 215.2 tok/s on the PRO 6000 and 60.9 tok/s on a Spark, about 3.5 times. GPUwerk has not run a PRO 6000 itself.
Can an RTX PRO 6000 run gpt-oss-120b?
Yes. gpt-oss-120b at MXFP4 is roughly 62 GB of weights (117B parameters at about 4.25 bits, my arithmetic), which leaves room for cache inside 96 GB. It also fits a 128 GB Spark. GPUwerk has not measured gpt-oss-120b on either card.
Which is better value for local LLMs?
It depends on the model. If your model fits in 96 GB and you run it all day, the PRO 6000 gives several times the decode speed per request, but it is a card you must put in a host PC, and I have not verified a retail price for it today. If you need more than 96 GB, want a complete machine, or only need to test a model for a few hours, a Spark at $0.79 an hour is the cheaper way to find out.
Does the PRO 6000 need a separate PC?
Yes. The Workstation Edition is a dual-slot PCIe Gen 5 x16 card rated at 600 W by NVIDIA, so it needs a host with a suitable power supply, cooling and CPU. A DGX Spark is a complete computer on a 240 W power supply, with its own CPU, memory and storage.
Can I rent an RTX PRO 6000 in the EU?
Hetzner lists the GEX131 server with an RTX PRO 6000 Blackwell Max-Q 96 GB in Helsinki and Falkenstein, with no minimum contract term. Hetzner's page quotes a monthly maximum, an hourly rate and a one-time setup fee; I could not retrieve the figures on 11 October 2026, so check the page. The Max-Q card has the same 96 GB and 1,792 GB/s as the Workstation Edition at 300 W. GPUwerk does not rent a PRO 6000; it rents DGX Sparks only.