RTX PRO 6000: Workstation, Max-Q and Server Edition — the full comparison

What NVIDIA’s own table leaves out

The official family page compares the three versions in six rows: memory, DisplayPort count, power, bus, form factor, cooling. Everything that actually separates them in production is missing — no bandwidth, no CUDA counts, no MIG or vGPU. A buyer relying on the primary source simply cannot compare the SKUs: the numbers sit scattered across three product pages, three PDFs and a whitepaper that never mentions Server Edition.

The full table

1,792 vs 1,597 GB/s: what the gap changes

The bus is 512-bit on all three, so the whole difference is memory clock: desktop versions run GDDR7 at 28 Gbit/s, the server card at roughly 25 — a clock NVIDIA publishes nowhere and which has to be derived backwards. What it means in practice: token generation reads every weight per token and tracks bandwidth almost linearly, so figure ~10% fewer tokens per second on Server Edition. Prompt processing is Tensor-Core-bound and rated identically in FP4.

What the gap does not change: whether a model fits. All three carry 96 GB — and that threshold is where the big jumps live. GamersNexus logged a 928% gain over an RTX 5090 on Llama 3.3 70B: the pure effect of 32 GB being too little and 96 GB being enough.

The cooler decides how many cards you get

Workstation breathes through the chassis (two axial fans): excellent for one card — 82 °C core, 88 °C memory, 32.5 dBA measured. With two cards the first card’s exhaust feeds the second, and the airflow arithmetic is the real limit. Two is the realistic ceiling. Max-Q is the same board with a blower and half the power limit: up to four per workstation, at the cost of ~12% FP32. Server Edition has no fan at all and relies on chassis airflow: in a certified server it holds 83–90 °C; on an open bench it throttles to the floor within minutes (thermal shutdown at 104 °C) — owners fix it with a shroud and a high-static-pressure fan that sounds like a vacuum cleaner.

The cable that quietly caps the card at 450 W

The most expensive trap in the line-up looks exactly like a defective card. All three versions use the 16-pin 12V-2×6 connector, whose SENSE pins tell the card how much power the supply is prepared to deliver — and the card obediently lowers its own ceiling. Real case: a Server Edition in a Supermicro chassis reports Max Power Limit: 450.00 W, and nvidia-smi -pl 600 changes nothing. The bundled cable was configured for 450 W; the full 600 W needs a different part (on Supermicro: CBL-PWEX-1364Y-30, ordered separately — put it in the BOM). Sometimes 450 W is the vendor’s deliberate choice to double card density; the measured cost is small — capping a Workstation card from 600 to 300 W lost ~6% on Llama 3.3 70B Q8.

MIG and vGPU: the hardest split

MIG (up to four isolated instances) is listed on all three — but on desktop versions it does not come enabled: driver R575+, vBIOS 98.02.55.00.00+ and compute display mode are required, after which video output disappears. Launch retail units shipped with an older vBIOS, so “Unable to enable MIG Mode: Not Supported” was the predictable first experience. vGPU leaves no room for debate: Server Edition only, up to 48 VMs per card (4 MIG instances × 12 time-sliced VMs). Lenovo’s Max-Q guide states it verbatim: “vGPU software support: No support.”

What nobody knows

NVIDIA publishes neither the memory clock nor the boost clock of Server Edition; no verified A/B test of two versions on one bench exists (the question has sat unanswered on the developer forum since February 2026); card weight appears in no datasheet; NVLink is neither confirmed nor denied.

What we supply

Eurokommerz delivers all three versions — NVIDIA RTX PRO 6000 Workstation Edition 96 GB, RTX PRO 6000 Max-Q and RTX PRO 6000 Server Edition — retail or OEM (tray), across the EU with manufacturer warranty.

FAQ