Up to 32B: the card
An RTX 5090 reads memory at 1,792 GB/s, 6.6 times the Spark’s 273. It wrote Qwen3.8-27B at 334.4 tok/s with drafting in a LocalMaxxing run.
DGX Spark vs RTX 5090 and RTX PRO 6000
The DGX Spark is NVIDIA’s desktop AI machine with 128 GB of unified memory at 273 GB/s. An RTX 5090 is a 32 GB graphics card at 1,792 GB/s that needs a PC built around it, and the RTX PRO 6000 is its 96 GB workstation sibling. For DGX Spark vs RTX 5090, the deciding question is whether your local LLM fits in 32 GB.
About 100 GB: the Spark
DeepSeek V4 Flash (101.5 GB) and Qwen3.8-Flash-Next (106.3 GB) fit in the Spark’s 128 GB, not on one 32 or 96 GB card.
Cards cost more than they did
An RTX 5090 sells for about $7,681 for the card alone (list $1,999). An RTX PRO 6000 workstation is about $21,460.
Which LLMs fit in 32, 96 and 128 GB?
An RTX 5090 holds Qwen3.8-27B (15.2 GB) in its 32 GB, while DeepSeek V4 Flash (101.5 GB) and Qwen3.8-Flash-Next (106.3 GB) fit on neither card and do fit in the DGX Spark’s 128 GB.
| Model and size | RTX 5090 card32 GB on the card | RTX PRO 6000 workstation96 GB on the card | NVIDIA DGX Spark128 GB unified | Lucebox Zero 495, 128 GB32 GB card + 128 GB |
|---|---|---|---|---|
| Qwen3.8-27B15.2 GB | Yes, on the card | Yes, on the card | Yes | Yes, on the card |
| DeepSeek V4 Flash101.5 GB | No | No | Yes | Yes, card and unified |
| Qwen3.8-Flash-Next106.3 GB | No | No | Yes | Yes, card and unified |
| Qwen3.8-27B and DeepSeek V4 Flash at once116.7 GB | No | No | Just, little room left | Yes, card and unified |
| DeepSeek V4.1 Flash190.8 GB | No | No | No | Yes, part on the SSD |
RTX 5090 LLM speed: tokens per second
An RTX 5090 wrote Qwen3.8-27B at 334.4 tokens per second with drafting in a LocalMaxxing run by jabbatheduck on 19 October 2026 (NVFP4, 32K context), the fastest single card in our data.
Model
Measure
R9700 (Lucebox card) writes 2.9x as fast as Strix Halo with a 27-token prompt, same engine and file. The other runs used a different setup. Drafting speeds some prompts far more than others.
RTX 5090 against the R9700 on Qwen3.8-27B
An R9700, the 32 GB card in a Lucebox, wrote the same model at 303.38 tok/s with drafting in schuttdev’s LocalMaxxing run on hipfire. The RTX 5090 reads memory 2.8 times as fast, yet its lead on this model is about 10%, and the two runs used different engines and model files.
Why the DGX Spark and RTX PRO 6000 have no bar here
The public DGX Spark and RTX PRO 6000 runs of Qwen3.8-27B both report more tokens per second than their memory speed allows without drafting. We hold both back until the people who sent them explain how they were run.
DGX Spark vs RTX 5090 and RTX PRO 6000 on paper
NVIDIA powers the DGX Spark, with its 128 GB of LPDDR5X at 273 GB/s, from a 240 W adapter. The RTX 5090 has 32 GB of GDDR7 and a 575 W rating, the RTX PRO 6000 has 96 GB of GDDR7 with ECC and a 600 W rating, and both cards read memory at 1,792 GB/s.
| Spec | DGX Spark (GB10) | RTX 5090 | RTX PRO 6000 |
|---|---|---|---|
| Memory | 128 GB LPDDR5X, or 64 GB from other brands | 32 GB GDDR7 on the card | 96 GB GDDR7 with ECC on the card |
| Memory speed | 273 GB/s, 256-bit bus | 1,792 GB/s | 1,792 GB/s |
| Processor and GPU | 20 Arm cores and a Blackwell GPU, 1 PFLOP at FP4 | Blackwell GPU; the rest is the PC you build | Blackwell GPU in a workstation |
| Software | CUDA, DGX OS | CUDA | CUDA |
| Power | 140 W chip, 240 W adapter; 104 to 160 W at the wall under load | 575 W for the card | 600 W for the card (300 W Max-Q) |
| Networking | ConnectX-7, 200 Gbps, links two Sparks | - | - |
| Size | 150 x 150 x 50.5 mm | - | - |
| Sources | NVIDIA, ServeTheHome | NVIDIA | NVIDIA |
The RTX PRO 6000 Max-Q
NVIDIA’s RTX PRO 6000 page also names a Max-Q version of the card but lists no specs for it. The figures in this table are for the Workstation Edition.
RTX 5090 and RTX PRO 6000 price for an LLM build
On 30 October 2026, the cheapest RTX 5090 in stock on a US price tracker was $7,681 for the card alone, a complete RTX PRO 6000 workstation was $21,460, and the whole DGX Spark was reported at $6,950.
| Machine | Memory | Bandwidth, GB/s | Software | US price |
|---|---|---|---|---|
| NVIDIA GB10 | ||||
| NVIDIA DGX Spark | 128 GB | 273 | CUDA | $6,950 to $6,995 |
| GB10 box with 64 GB (Dell, ASUS and others) | 64 GB | 273 | CUDA | $4,999 |
| Dell Pro Max with GB10 | 128 GB | 273 | CUDA | $6,593 |
| Gigabyte AI TOP ATOM | 128 GB | 273 | CUDA | $8,000 |
| ASUS Ascent GX10 | 128 GB | 273 | CUDA | $6,060 backordered |
| Lenovo ThinkStation PGX | 128 GB | 273 | CUDA | $5,859 backordered |
| NVIDIA RTX build | ||||
| RTX 5090 card The card alone; the PC around it costs extra. | 32 GB | 1,792 | CUDA | $7,681 in stock |
| RTX PRO 6000 workstation A complete workstation; the card alone was $17,063 in stock. | 96 GB | 1,792 | CUDA | $21,460 |
US prices read on 30 October 2026.
What the rest of an RTX 5090 PC adds
The $7,681 buys the card only. NVIDIA recommends a 1,000 W power supply for a PC with an RTX 5090, and the processor, board, RAM and case come on top. The DGX Spark price covers the whole machine, storage included.
RTX PRO 6000 price: the card or a workstation
On 30 October 2026, an RTX PRO 6000 card alone was $17,063 in stock and a complete workstation with one was $21,460. NVIDIA publishes no price for the card.
Cost per GB of memory
Per GB the model can sit in, the RTX 5090 card costs about $240, the RTX PRO 6000 card about $178 and the DGX Spark about $54. The Spark’s memory is the cheapest per GB and also the slowest, at 273 GB/s against the cards’ 1,792 GB/s.
The RTX 5090 price gamers see
The RTX 5090 price is the same for a gaming PC as for an LLM build. NVIDIA launched the card at $1,999, and press reports from October 2026 put US listings between about $4,299 in Micro Center stores and about $10,000 on Newegg, with GDDR7 memory in short supply.
Which machine runs your model for your budget?
Set the model to Qwen3.8-27B and raise the budget to see where the cards come in. For single graphics cards ranked by VRAM, bandwidth and price, see the best GPU for local LLMs.
Model
Rank by
Our pick
- Holds Qwen3.8-27B (15.2 GB) on the card's 32 GB.
- Writes it at 334.4 tok/s, drafting on (LocalMaxxing).
- $7,681, $1,319 under your budget.
Also consider
- Holds Qwen3.8-27B (15.2 GB) in 64 GB of unified memory.
- Writes it at 104.32 tok/s on another box with the same chip, drafting on (LocalMaxxing).
- $2,200, $6,800 under your budget.
- RTX 5090 card32 GB on the cardOn the card 48% used1,792 GB/s334.4 tok/s drafting on, LocalMaxxing
- GMKtec EVO-X2, 64 GB64 GB unifiedFits 24% used256 GB/s104.32 tok/s drafting on, LocalMaxxing, on another box with the same chip
- GMKtec EVO-X2, 128 GB128 GB unifiedFits 12% used256 GB/s104.32 tok/s drafting on, LocalMaxxing, on another box with the same chip
- Mac Studio (M5 Max), 48 GB48 GB unifiedFits 32% used614 GB/s56.4 tok/s drafting on, public benchmark
- Mac Studio (M5 Max), 128 GB128 GB unifiedFits 12% used614 GB/s56.4 tok/s drafting on, public benchmark
- Mac Studio (M5 Ultra), 96 GB96 GB unifiedFits 16% used1,228 GB/sno run yet
- Lucebox Zero 495, 128 GB32 GB card + 128 GBOn the card 10% used640 GB/sno run yet
- Lucebox Zero 495, 192 GB32 GB card + 192 GBOn the card 7% used640 GB/sno run yet
- GB10 box with 64 GB (Dell, ASUS and others)64 GB unifiedFits 24% used273 GB/sno run yet
- Framework Desktop (Ryzen AI MAX+ PRO 495)192 GB unifiedFits 8% used273 GB/sno run yet
- NVIDIA DGX Spark128 GB unifiedFits 12% used273 GB/sno run yet
- Mac Studio (M5 Ultra), 256 GB256 GB unified$1,799 over budgetFits 6% used1,228 GB/sno run yet
- RTX PRO 6000 workstation96 GB on the card$12,460 over budgetOn the card 16% used1,792 GB/sno run yet
- tinybox red flash192 GB on 6 cards$13,000 over budgetOn the card 8% used3,840 GB/sno run yet
- Mac Studio (M5 Ultra), 512 GB512 GB unified$13,299 over budgetFits 3% used1,228 GB/sno run yet
The pick holds the model with room for the prompt, fits your budget and has the fastest run we can link to. Runs with drafting rank against each other first, then runs without it, then machines with no run, by memory speed. The amber band is the Lucebox's 32 GB card.
RTX 5090 build or Lucebox: where each fits
We make the Lucebox Zero 495, so read this box as ours.
Where it fits
- Qwen3.8-27B (15.2 GB) on the Lucebox’s own 32 GB R9700, a card that wrote it at 303.38 tok/s with drafting on a short code prompt on LocalMaxxing, with 128 or 192 GB of unified memory left for a second model.
- DeepSeek V4 Flash (101.5 GB) split between the card and the unified memory, a model that fits on no single RTX 5090 or RTX PRO 6000.
- $5,999 for the whole workstation, against $7,681 for an RTX 5090 card with no PC around it.
Where it doesn't
- Your code or tools need CUDA: the R9700 runs ROCm and Vulkan.
- You only run models up to 32 GB and want the most speed: the RTX 5090 reads at 1,792 GB/s, the R9700 at 640.
- You need a machine this month.
From $5,999, delivered from January 2027. See the hardware Configure and order
More local AI hardware comparisons
NVIDIA DGX Spark
AMD Ryzen AI MAX+ boxes
- DGX Spark vs Strix Halo Close on memory speed, about half the price, CUDA only on the Spark.
- Framework Desktop The same chip as the cheaper boxes, in a small case you can build.
- GMKtec EVO-X2 AI mini PC One of the cheapest 128 GB Strix Halo boxes, with 64 GB as an option.
- Minisforum MS-S1 MAX Strix Halo with a PCIe slot, and a 192 GB Gorgon Halo version.
- Beelink GTR9 Pro A 128 GB Strix Halo box with two 10 GbE ports.
- Corsair AI Workstation 300 Strix Halo from a big US brand, in stock.
Apple Mac Studio
Graphics card builds
DGX Spark vs RTX 5090 FAQ
Is the DGX Spark faster than an RTX 5090?
Not on models that fit in 32 GB: the RTX 5090 reads memory at 1,792 GB/s, 6.6 times the DGX Spark’s 273 GB/s, and token speed follows memory speed. The DGX Spark wins when the model needs more than 32 GB, because it has 128 GB.
Is an RTX 5090 good for running LLMs?
The RTX 5090 is the fastest single card in our data for LLMs that fit in its 32 GB: it wrote Qwen3.8-27B (15.2 GB) at 334.4 tok/s with drafting on LocalMaxxing. Larger models, such as DeepSeek V4 Flash at 101.5 GB, do not fit on it.
Is the RTX PRO 6000 worth it for LLMs?
The RTX PRO 6000 is worth it when your models need between 32 and 96 GB on one fast card: it has 3 times an RTX 5090’s memory at the same 1,792 GB/s. A workstation with one was $21,460 on 30 October 2026, about 3.1 DGX Sparks, and models above 96 GB, like Qwen3.8-Flash-Next (106.3 GB), still do not fit.
Mac Studio M3 Ultra or RTX 5090 for LLMs?
The Mac Studio M3 Ultra holds far bigger models: it reads memory at 819 GB/s, and the 512 GB version Apple sold at launch ran DeepSeek V4 Flash at about 33 tok/s on LocalMaxxing. The RTX 5090 is faster on anything that fits in 32 GB, and Apple no longer sells the M3 Ultra new. More on the Mac Studio for local LLMs.
DGX Spark or RTX PRO 6000?
The RTX PRO 6000 writes much faster on any model up to 96 GB, with 1,792 GB/s against the DGX Spark’s 273 GB/s. The DGX Spark holds DeepSeek V4 Flash in its 128 GB and was reported at $6,950, against $21,460 for an RTX PRO 6000 workstation.
Why is the RTX 5090 so expensive in 2026?
Press reports in October 2026 tie the RTX 5090’s price to the cost of its GDDR7 memory. The cheapest card in stock on a US price tracker was $7,681 on 30 October 2026, 3.8 times its $1,999 launch price.
Can an RTX 5090 run DeepSeek V4 Flash?
Not on the card alone: DeepSeek V4 Flash takes 101.5 GB and the RTX 5090 has 32 GB, so most of it would sit in the PC’s far slower system memory. For a box that holds it whole, read DeepSeek V4 Flash on Strix Halo.