RTX PRO 6000 Max-Q in a rack server: when it works and what to check before you order

RTX PRO 6000 Max-Q in a rack server: the short answer

An RTX PRO 6000 Max-Q runs in a rack server when the chassis is built, and listed by its maker, for actively cooled cards. The Max-Q is a 300 W dual-slot card with its own blower, so it does not rely on the server fans to push air through its heat sink as the passive Server Edition does. Four of them fit in a rack-mounted workstation that lists them, and some rack servers list the card in smaller numbers. The Server Edition is the card NVIDIA builds for the data centre: passive, configurable from 300 to 600 W by its product brief, and the only RTX PRO 6000 that supports vGPU. Both carry 96 GB of GDDR7 with ECC, so the models that fit are the same.

NVIDIA’s product page for the Max-Q, read on 16 October 2026, gives 300 W, 96 GB and 1,792 GB/s and says the card “enables up to four GPUs in a single system”. The datasheet (5349650, June 2026) lists the thermal solution as “Active”, a dual-slot, full-height card of 4.4 by 10.5 inches and one PCIe CEM5 16-pin power connector. The product page also speaks of “up to four GPUs in a workstation” and names no server or rack. Some server makers list the card, so whether a given chassis takes it is decided by its maker’s list. Our comparison of the Workstation, Max-Q and Server editions covers all three in full.

Max-Q vs Server Edition in a rack

The two cards differ in cooling, power range, memory bandwidth and in what NVIDIA lists them for.

NVIDIA Max-Q product page and datasheet 5349650 (June 2026); Server Edition product brief SP-12355-001_v02 (June 2025); NVIDIA GPUs supported by vGPU (2 October 2026); NVIDIA AI Enterprise 8.2 support matrix (2 September 2026); NVIDIA RTX PRO workstation warranty page; all read on 16 October 2026.

The Max-Q keeps the full memory bandwidth of the workstation cards at half their power. Token generation for LLMs reads every weight for each token, so it tracks memory bandwidth, and the Max-Q’s 1,792 GB/s is about 12 per cent above the Server Edition’s 1,597 GB/s. The Server Edition has more power to spend on compute: its product brief gives a 2,430 MHz boost clock, at a 600 W limit it has twice the Max-Q’s power, and prompt processing, which is bound by the Tensor Cores, gains from that. We found no published test of the two cards in the same server, so treat both statements as direction, not as a measured gap.

Blower cooling and airflow direction in a server

A passive card has only a heat sink, and the server fans push air through it. The Server Edition’s product brief gives a minimum airflow “whenever input power is provided to the card”, assumed to be fully ducted through the heat sink: 43 CFM at a 25 °C heat-sink inlet, 65 CFM at 35 °C and 125 CFM at 45 °C. The card accepts that air from either side, left to right or right to left.

The Max-Q moves its own air. Puget Systems, in its comparison of July 2025, describes a blower-style cooler that “takes air from inside the system and blows it out the back of the case”, so the heat of one card does not significantly affect its neighbours. The same article adds that cards placed directly next to each other leave little room for the fans to draw in air, and that a single-slot gap between cards is ideal, or high-airflow system fans that force cool air into the gaps.

The blower exhausts through the card’s bracket, so the bracket has to face the hot aisle, as it does in a riser at the rear of a front-to-back chassis. In chassis where the GPU brackets face the front, the blower would push against the server’s own airflow. In 2U risers with adjacent double-width slots, the blowers sit next to each other and depend on the server fans for intake air, which is the arrangement Puget, writing about workstations, warns against.

Rack workstations and servers that take the Max-Q

Rack-mountable workstations and tower servers that list the Max-Q come from more than one manufacturer. One maker specifies its dual-socket workstation, in a document dated 9 October 2026, as “Tower (54L) or 5U rack, optional rack rail kit needed”, with support for “up to four NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (300W)”. The same document offers an IPMI 2.0 BMC as an option, so remote management comes only if it is ordered. Our comparison of rack servers, towers and workstations lists a tower server from another maker that converts to a rack and names the Max-Q in its GPU list. Tower builds are in our guide to four RTX PRO 6000 Max-Q in one workstation.

Rack servers that list the Max-Q are fewer. One server maker’s product guide, updated on 8 May 2026, offers the card as a server option, “96GB PCIe Gen5 Active GPU”, for two of its rack servers, with the figure 3 for each in its compatibility table, and marks every other current server as not supported. By the maker’s own product guide, one of the two is a 2U server that takes up to three double-wide GPUs. The option guide states “vGPU software support: No support” and “NVLink support: No”. It ships the option without power cables, which are server-specific; the parts it lists for them are thermal kits that include the other components needed to install the card. The DisplayPort outputs are disabled by default.

For passive cards the server lists are much longer, and the models that list the Server Edition, with two to eight cards each, are in our article on RTX PRO 6000 Server Edition compatible servers.

Checklist before four Max-Q go into a chassis

The chassis maker’s documents answer each item.

Compiled from the Max-Q datasheet 5349650 (dimensions, connector), one server maker’s Max-Q product guide (8 May 2026, cables), the Server Edition brief (airflow) and Puget Systems (July 2025, card spacing).

The card is full height and 10.5 inches long, about 267 mm, so the slot needs room for that length. Lanes are the next limit: four cards at x16 take 64 PCIe lanes before the network cards and NVMe drives take theirs, and our article on how many GPUs fit in one server works this out per chassis together with power and airflow. With redundant supplies, the server should carry the 1,200 W of four Max-Q at full load when one supply fails.

The BMC item needs a direct answer from the maker: ask whether the BMC reads the card’s temperature and raises the fans with it, because the server fans also supply the blowers’ intake air. After installation, read the clock event reasons under load with nvidia-smi -q -d PERFORMANCE and the temperatures with nvidia-smi -q -d TEMPERATURE; a thermal slowdown shows that the card lacks air.

We supply the Max-Q and build on your own chassis after a compatibility check of the platform, power and cooling. Tell us the chassis model and how many cards it should hold, and we say in advance if it will not work.

Worked example: four Max-Q or four Server Edition for 50 to 200 users

Take a company that plans a private assistant for 50 to 200 staff on one rack server with four cards. Users of an assistant are not all generating at the same moment, and we found no published ratio between staff and concurrent requests, so size for the peak number of requests in flight. With Qwen3-32B in FP8 and an FP8 KV cache, one 96 GB card holds about 51 concurrent conversations at a full 8K context, by our sizing rule of 0.9 × the 95.6 GiB the driver reports, less 3 GiB and the weights. Four cards running one copy of the model each hold about 204. Llama 3.3 70B in FP8, with the same FP8 cache, leaves room for 12 per card, 48 across four.

Both editions have 96 GB, so these counts are the same and the choice rests on power, cooling and software. Four Max-Q draw 1,200 W for the cards, and by the same estimate as for a four-card tower, about 2 kW at the wall with one processor, which a 16 A single-phase feed at 230 V carries. Four Server Edition cards at 600 W draw 2,400 W for the cards alone, and from four 600 W cards upwards we plan 32 A or three-phase. The Server Edition can be limited to 300 W, the minimum in its product brief (NVIDIA’s family page gives 400 to 600 W), which brings its power to the Max-Q’s level while keeping passive cooling and vGPU; we found no published measurement of its speed at that limit.

For one service on bare metal in a chassis its maker lists for four Max-Q, the Max-Q gives the higher memory bandwidth at half the power. Virtual machines with vGPU, NVIDIA AI Enterprise support or a choice among many server models point to the Server Edition.

We size AI servers from the model and the concurrent users and check the rack, power and airflow before we quote. Write to us with the model, the peak users and the rack position for a configuration and quote within one business day.

Workstation cards in a data centre: warranty and NVIDIA support

NVIDIA’s Max-Q page speaks of “up to four GPUs in a workstation”, while the Server Edition’s product brief describes a card for “data center and edge server deployments”. NVIDIA’s warranty page for RTX PRO workstation graphics cards, read on 16 October 2026, lists the “NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition” with “Three (3) years from the date of purchase”, for original purchases from NVIDIA or an NVIDIA-authorised seller, and sends claims to the local distributor or reseller first. For cards that go into a rack, have the seller state the warranty term and the claims route in writing. When a server maker sells the Max-Q as its own option, its terms apply instead: the product guide quoted above gives a “One year limited warranty” and adds that a card installed in one of that maker’s servers takes on the server’s base warranty and any warranty upgrades.

In a data centre, the Max-Q has no NVIDIA support for vGPU or NVIDIA AI Enterprise. NVIDIA’s list of GPUs supported by vGPU, updated on 2 October 2026, names the Server Edition from release 19.0 and does not name the Max-Q. The NVIDIA AI Enterprise 8.2 support matrix, updated on 2 September 2026, lists the RTX PRO 6000 and RTX PRO 4500 Server Editions as “supported with servers listed on the NVIDIA-Certified Systems”, and the Max-Q does not appear in it. The NVIDIA-Certified Systems list, updated on 4 October 2026, names systems with the Server Edition and none with the Max-Q. MIG splits each Max-Q into up to four 24 GB instances, according to its datasheet, so a team can share the cards without virtual machines.

What we supply

We supply the NVIDIA RTX PRO 6000 Blackwell in all three editions, Max-Q, Workstation Edition and Server Edition, with manufacturer warranty on one EU contract and invoice, as cards for machines you already run or in AI servers built to order. For four Max-Q in a rack workstation or server, we check the platform, power and cooling of the chassis first and say in advance if something will not work. For a rack server with four or eight passive cards, we build around the Server Edition, with vGPU and NVIDIA AI Enterprise licences on the same invoice where the cards will run virtual machines. Our RTX PRO 6000 page compares the three editions.

FAQ