RTX PRO 6000 Max-Q vs Workstation vs Server Edition for LLMs: power, speed and price in November 2026

The RTX PRO 6000 Blackwell Max-Q and Workstation Edition have the same 96 GB of GDDR7 and 1,792 GB/s; the Max-Q is rated at 300 W with a blower, the Workstation at 600 W with a flow-through cooler, and the passive Server Edition at up to 600 W with 1,597 GB/s. For one user generating text, 300 W costs little: a Workstation card capped at 300 W kept 94% of its llama.cpp generation speed on Llama 3.3 70B Q8_0, but its prompt processing fell 48% (Level1Techs, December 2025). On 4 November 2026 the cheapest new cards on gpuprix were $17,063 for a Max-Q and $17,479 for a Workstation Edition, so for most LLM builds the Max-Q is the one to buy.

Sources: NVIDIA's pages for the Workstation Edition, Max-Q and Server Edition; Lenovo Press LP2263 (24,064 CUDA cores, 512-bit bus, up to 4 MIG instances of 24 GB, no NVLink; updated 28 July 2026); PNY's Max-Q page for the "Blower Active Fan" and single 16-pin connector. Prices from gpuprix US, both Amazon listings of PNY cards. All three have 24,064 CUDA cores per Lenovo and HOSTKEY's comparison table; clocks, power, cooling and, on the Server Edition, memory speed differ. NVIDIA quotes the Workstation and Max-Q FP4 figures with sparsity.

Is the Max-Q slower than the Workstation Edition for LLMs?

Barely for generation, clearly for prompt processing. Decode is limited by memory bandwidth, which both cards share at 1,792 GB/s; prefill and batched serving are limited by compute, which scales with power. Nobody I could find has published a same-model test of a retail Max-Q against a Workstation card, so the evidence comes from Workstation cards power-limited with nvidia-smi -pl.

tok/s. Level1Techs figures are llama-bench runs by user thr3e on a Workstation Edition. Naveen's runs used a 20k-token window and a 4k-token prompt, and his single figures are not split into prefill and generation. local-inference-lab reports 500 W and 600 W as nearly identical.

The two decode sources disagree, 6% against 18 to 31%, and the method explains most of it. thr3e's tg128 isolates generation; Naveen's numbers come from requests that included a 4k-token prompt, and that prompt is the part a power cap slows down. My reading is that pure generation loses under 10% at 300 W and anything prompt-heavy loses a quarter to a half.

A retail Max-Q should do somewhat better than a capped Workstation card, because its clocks are tuned for 300 W. NVIDIA rates it at 88% of the Workstation Edition's peak (3,511 against 4,000 FP4 TOPS, 110 against 125 TFLOPS FP32). A matrix-multiply test quoted in the same Level1Techs thread measured 423, 361 and 260 TFLOPS at 600, 450 and 300 W, or 61% at 300 W; the thread does not make clear whether the 300 W run was a capped Workstation card or a Max-Q. Real Max-Q prefill likely sits between 61% and 88% of a Workstation card's; no published LLM test pins it down.

What do the three editions cost in November 2026?

About $17,000 to $20,000 new in the US and 16,500 euros or more in Germany, roughly double early 2025 prices, and stock is thin. Igor's Lab reported on 14 August 2026 that NVIDIA's US marketplace listed the Workstation Edition at $16,000, marked unavailable, against early 2025 retailer listings near $8,565.

Sources: gpuprix US and gpuprix Germany listing pages (updated 4 November 2026, prices as listed by the retailer, German prices include VAT); Scan's Server Edition page, read 7 November 2026; Igor's Lab, 14 August 2026.

The Max-Q and Workstation Edition now cost about the same, so price no longer decides between them. To use one without buying, Hetzner's GEX131 rents a Max-Q server at 1,499 euros a month or 2.4022 euros an hour, plus a 1,499-euro setup fee, excluding VAT (read 7 November 2026); the Hetzner comparison covers that offer against a rented Spark.

How does the Server Edition differ?

It is a passive card for servers with ducted airflow, and it has 11% less memory bandwidth: 1,597 GB/s against 1,792, per NVIDIA and Lenovo. Since decode follows bandwidth, expect single-stream generation about a tenth below a Workstation card's on the same model.

Its power limit depends on the chassis. One owner found a Server Edition in a Supermicro SYS-521GE-TNRT capped at 450 W with nvidia-smi -pl 600 refused; a forum member traced it to the power cable, and Supermicro named cable CBL-PWEX-1364Y-30 for 600 W (NVIDIA developer forum, September to October 2025). Lenovo's ThinkSystem variant is slot-capped at 450 W. HOSTKEY's August 2025 test saw at most 83°C in standard mode and about 90°C in 600 W mode, and Windows drivers did not detect the card at the time; Linux worked.

The one recent LLM test on a Server Edition: Paralon Cloud served Qwen3.8-27B at W4A16 on vLLM 0.27.1 and measured 106 tok/s for one request, 636 tok/s total at 8 and 1,342 at 32 (Paralon, 28 September 2026).

Which edition suits a multi-GPU build?

The Max-Q, by design. NVIDIA's Max-Q page says up to four per system: 384 GB of VRAM at 1,200 W of board power, against 2,400 W for four Workstation cards. A blower exhausts through the card's bracket, while the Workstation card's flow-through cooler blows its heat into the case, where the next card breathes it in.

Reported four-card runs, both over PCIe (the vLLM issue states no NVLink, and Lenovo lists NVLink as not supported on the Server Edition):

• Four Max-Q cards capped at 250 W on PCIe Gen4 x16, vLLM nightly 0.28.1rc1, DeepSeek-V4-Flash-0731 with tensor parallelism 4 and FP8 KV cache: 220.9 tok/s single-stream with DSpark speculative decoding, 525.4 tok/s total for 8 agents at 150k context, 373.0 without speculation (vLLM recipes issue 967, 13 September 2026).
• Four cards, Qwen3.5-397B at NVFP4 on vLLM with 2-token MTP: 130 tok/s single-stream, 1,127 tok/s total at 50 users with 1k context. The same repo measured peer-to-peer at 55 to 56 GB/s and 0.36 to 0.45 µs latency, against 14 µs with peer-to-peer off (local-inference-lab, undated).

Each Max-Q takes one 16-pin connector (PNY), and four of them draw 1,200 W before the CPU and drives, so size the power supply for the whole box, not the cards alone.

What fits in 96 GB?

Every 100-120B-class mixture-of-experts model at 4-bit, with 20 to 30 GB left for cache. That is the main reason to pay for this card over a 32 GB one.

Weight sizes from the Hugging Face file listings of ggml-org/gpt-oss-120b-GGUF and Unsloth's Qwen3.5-122B-A10B MXFP4_MOE folder, read 7 November 2026. Spare memory is my subtraction from 96 GB before cache and overhead. The Llama 3.3 70B Q8_0 size is my estimate (70B parameters at 8.5 bits); the run itself shows it fit. Hardware Corner does not name the edition or power limit it tested. DeepSeek-V4-Flash's 284B total comes from the DeepSeek-V4-Flash model page.

Models between about 85 GB and 120 GB, or 120B-class models with long contexts for several users, outgrow one card; the DGX Spark against RTX PRO 6000 comparison covers the 128 GB alternative and the speed it gives up. To size a cache for your own context, the memory calculator uses the standard formula.

Is one RTX PRO 6000 better than three RTX 5090s?

At November 2026 prices, yes. Three 5090s also total 96 GB, but as three 32 GB pools, with 1,725 W of board power and no NVLink (NVIDIA's 5090 spec page lists "No"). vLLM's parallelism guide points uneven splits and GPUs without NVLink toward pipeline parallelism, which adds capacity rather than single-stream speed.

When tensor parallelism does work, split cards can be faster per request, because each card reads only its slice of the weights. Paralon measured two 5090s (400 W each, vLLM 0.31.0) on Qwen3.8-27B at 149 to 156 tok/s for one request, against 106 for one Server Edition, while both landed near 1,320 to 1,340 tok/s at 32 requests (Paralon, 12 October 2026). The engine versions differ (0.31.0 against 0.27.1), so treat the single-request gap as approximate.

My arithmetic from gpuprix's RTX 5090 page and its RTX PRO 6000 page, both updated 4 November 2026. The 5090's 3-month median includes offers that are now out of stock. Excludes the larger PSU, board and case three 575 W cards need.

If you can actually buy 5090s near the median, three of them are cheaper and faster for batched serving of a model that splits cleanly. At in-stock prices the PRO 6000 is cheaper, draws under a fifth of the power as a Max-Q, and holds a 75 GB model without any parallelism. The sibling article on dual RTX 5090 builds covers the multi-card route, and the planned piece on RTX 5090 power limits shows the same decode-versus-prefill pattern on the consumer card.

Which edition should you buy?

For one card in a desktop running LLMs, I would buy the Max-Q. It has the same memory and bandwidth, generation within single digits of the Workstation card on the best evidence available, half the heat, and a blower that leaves room for a second card later, at about the same price.

• Workstation Edition: one card, prompt-heavy work (RAG over long documents, coding agents resending 100k-token contexts, batched serving), and a case and PSU that handle 600 W.
• Max-Q: everything else, and every build with more than one card.
• Server Edition: rack servers with ducted airflow. Check that the chassis cable allows 600 W, and accept 11% less bandwidth.

Before spending $17,000, check the model itself. A 128 GB DGX Spark runs every single-card model in the table above, at a fraction of the PRO 6000's speed; GPUwerk rents one by the minute ($0.79 an hour), which is enough to confirm the model is the right one before you buy the faster card.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA
• NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, NVIDIA
• NVIDIA RTX PRO 6000 Blackwell Server Edition, NVIDIA
• ThinkSystem NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB PCIe Gen5 GPU, Lenovo Press, 2026-07-28
• NVIDIA RTX PRO 6000 Blackwell Max-Q, PNY
• RTX PRO 6000 Blackwell price in United States, gpuprix, 2026-11-04
• RTX PRO 6000 Blackwell price in Germany, gpuprix, 2026-11-04
• PNY NVIDIA RTX PRO 6000 96GB Blackwell Server Edition, Scan Computers, 2026-11-07
• RTX PRO 6000 Blackwell now costs 16,000 US dollars, Igor's Lab, 2026-08-14
• GEX131 dedicated GPU server, Hetzner, 2026-11-07
• RTX Pro 6000 600W vs 300W performance?, Level1Techs Forums, 2025-12-05
• 96GB VRAM on a single GPU: benchmarking NVIDIA's RTX PRO 6000, Naveen Kulandaivelu, 2025-08-07
• RTX 6000 Pro Server power limit of 450W instead of expected 600W, NVIDIA Developer Forums, 2025-09-15
• NVIDIA RTX 6000 Blackwell Server Edition: tests, benchmarks and comparison, HOSTKEY on Hugging Face, 2025-08-30
• NVIDIA RTX PRO 6000 Workstation GPU review, StorageReview, 2025-10-07
• RTX PRO 6000 Blackwell local LLM benchmarks, Hardware Corner, 2026-03
• DeepSeek-V4-Flash-0731 DSpark on 4x RTX PRO 6000 (issue #967), vllm-project/recipes on GitHub, 2026-09-13
• rtx6kpro benchmark results, local-inference-lab on GitHub
• RTX PRO 6000 vs two RTX 4090s on a 27B model, Paralon Cloud, 2026-09-28
• Two RTX 5090s on a 27B model: 150 tok/s alone, 1,300 at 32, Paralon Cloud, 2026-10-12
• GeForce RTX 5090 price in United States, gpuprix, 2026-11-04
• GeForce RTX 5090 specifications, NVIDIA
• Parallelism and scaling, vLLM documentation
• gpt-oss-120b-GGUF files, ggml-org on Hugging Face
• Qwen3.5-122B-A10B-GGUF, MXFP4_MOE files, Unsloth on Hugging Face