Four NVIDIA RTX PRO 5000 Blackwell 48 GB in one workstation: what 192 GB runs

Four 48 GB cards: what the build holds

A workstation with four NVIDIA RTX PRO 5000 Blackwell 48 GB cards has 192 GB of GPU memory, and by our sizing rule each card holds about 40 GiB of weights and cache. That makes it a machine for models up to 32B per card, four separate services side by side, or eight developers on isolated 24 GB MIG instances; a 70B model in FP8 runs only split over all four cards, for about 73 conversations at once. It draws about 1.8 to 2 kW at the wall by our estimate, which one 230 V office circuit carries.

Each card has 48 GB of GDDR7 with ECC at 1,344 GB/s and 300 W, in an actively cooled dual-slot board of 4.4 × 10.5 inches, according to NVIDIA's product page read on 24 October 2026; NVIDIA rates it at 2,064 AI TOPS in FP4 with sparsity. The 48 GB and 72 GB versions share every speed figure, as our comparison of the RTX PRO 5000 with 48 GB or 72 GB shows, so the choice between them is a question of memory per card.

NVIDIA RTX PRO 5000 product page (read 24 October 2026) and MIG user guide (13 October 2026). Driver figure scaled from the 95.6 GiB of a 96 GB card; usable memory by our users-per-card rule with ECC off.

The rule takes 90 per cent of the memory the driver reports and subtracts 3 GiB per card for activations and the runtime: 0.9 × 47.8 − 3 = 40.0 GiB. The rest after the weights is KV cache, the memory each running conversation needs. Our guide to how many users one RTX PRO 6000 can serve derives the rule; if ECC reserves memory, about 37.3 GiB per card remain.

Where 48 GB per card draws the line

The smaller card changes which models stay on one card. With 40 GiB usable, Qwen3-14B in FP8 at 15.2 GiB leaves plenty of cache, Qwen3-32B in FP8 at 32.0 GiB leaves 8 GiB, and Llama 3.3 70B in NVFP4 at 39.8 GiB loads with 0.2 GiB to spare, which holds no conversation. Anything larger is split.

Our estimates by the users-per-card rule with 40.0 GiB usable per card; conversations of 8K tokens with an FP8 KV cache, gpt-oss-120b at 32K with a 16-bit cache of 1.125 GiB. Qwen3-235B cache 0.73 GiB per 8K conversation from its config.json (94 layers, 4 KV heads, head size 128), read 24 October 2026. Memory arithmetic, not measured throughput.

The table shows why this build suits small and mid-sized models. A 70B model in FP8 on a pair leaves cache for nine conversations, too few for a department, and only over all four cards does it reach 73. Qwen3-235B-A22B in NVFP4, about 124.8 GiB, loads over all four cards with 35 GiB of cache; the FP8 version, about 219.8 GiB, does not fit 192 GB at all. For other models, see our LLM hardware requirements by model.

Layouts for a team: copies, pairs and MIG

Four cards allow three kinds of layout for one model, and MIG adds a fourth for many users with small needs.

Our estimates by the users-per-card rule, 8K tokens with an FP8 KV cache.

The counts are requests in flight at once. The example values in our guide to a private ChatGPT server by company size put about 20 requests in flight at peak for 500 staff and 80 for 2,000. By memory, four copies of Qwen3-14B cover a company of the first size with each card independent, and Qwen3-32B as two pairs covers the second. Four copies keep running when one card fails; a TP=4 copy stops.

A common plan for an engineering or data science team combines the two. One pair serves Qwen3-32B in FP8 to the team, 48 conversations, while the other two cards are split into four MIG instances of 24 GB for development and tests. Send us the models and the number of developers and service users, and we plan the split between service and MIG before quoting the machine.

Speed on a split model

The RTX PRO 5000 has no NVLink, so a split model exchanges data over PCIe 5.0 after every layer. By default vLLM's source disables its custom all-reduce on more than two PCIe-only GPUs, so TP=2 pairs keep the faster kernel and TP=4 uses NVIDIA's NCCL; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576), and vLLM's parallelism guide advises pipeline parallelism on GPUs without NVLink "for higher throughput and lower communication overhead". Our article on TP=2, TP=4 or replicas without NVLink explains both paths.

Bandwidth sets a ceiling for one user. By our arithmetic, Qwen3-32B in FP8 on one card, about 34.4 GB read per token, is held to about 39 tokens per second; Qwen3-14B in FP8, about 16.3 GB per token, to about 82. Llama 3.3 70B in FP8 under TP=4 reads about 17.7 GB per card and token, a ceiling of about 76 before the cost of the all-reduce over PCIe. These are upper bounds from the specifications, not measurements. We found no published measurement of tensor-parallel LLM serving on four RTX PRO 5000 cards, and the published single-card results are in our RTX PRO 5000 benchmarks.

Eight 24 GB GPUs with MIG

NVIDIA's MIG user guide, updated on 13 October 2026, lists the RTX PRO 5000 48 GB with the profiles 1g.24gb and 2g.48gb, each also as a +gfx variant with graphics support. Each card splits into two isolated instances of 24 GB, so the workstation offers eight separate GPUs with their own memory and compute, or any mix of whole cards and halves. A 24 GB half holds gpt-oss-20b, or Qwen3-14B in FP8 with cache for about five conversations of 8K tokens by our rule. That suits one developer, one notebook server or one test instance per half.

MIG on this card needs Linux, driver 575.51.03 or later, a vBIOS of 98.02.73.00.00 or later and the display mode switched to compute with DisplayModeSelector 1.72 or later, which turns the card's display outputs off. NVIDIA's MIG guide also states that NCCL is not supported with MIG, so a model split over cards needs whole cards. Our RTX PRO 6000 MIG profiles describe the same procedure on the larger card.

The 48 GB card is not on NVIDIA's vGPU list of 2 October 2026; among RTX PRO 5000 cards only the 72 GB version is, on Red Hat KVM. Virtual desktops on vSphere need a Server Edition card, as our guide to MIG and vGPU: VMs per GPU explains.

Rendering, CAD and AI on one machine

The 48 GB card is also a graphics card with four DisplayPort outputs per card. In Puget Systems' professional GPU roundup of 18 December 2025 it ran about 50 per cent ahead of the RTX 5000 Ada in Blender and 35 per cent in V-Ray GPU, as our buy page for RTX PRO 5000 48 GB workstations reports; that roundup tested single cards. Rendering and visualisation use the cards in graphics mode, while MIG needs compute mode, which turns the display outputs off; NVIDIA's documentation we read does not say whether one card may stay in graphics mode while the others run MIG, so a machine meant for both is a point to settle with us before the order.

Towers and power for four cards

As of 24 October 2026 Dell's Precision 7960 configurator offers four RTX PRO 5000 48 GB, BIZON lists the card for its four-GPU X5500 G2, and Lenovo lists it for the ThinkStation PX without a maximum. Two of the Precision 7960's four GPU slots are PCIe 4.0 x16, which matters for a TP=4 layout more than for copies or MIG. NVIDIA's configuration guide for certified systems asks for at least six cores per GPU and system memory of at least twice the GPU memory, so we plan 24 cores and 384 GB, eight 48 GB modules on an eight-channel Threadripper PRO.

By our estimate four 300 W cards, a 350 W processor and 15 per cent for memory, drives, fans and supply losses come to about 1.8 to 2 kW at the wall, 7.8 to 8.7 A on 230 V. A 16 A circuit planned at 80 per cent carries 2.94 kW, so the tower fits one circuit if it does not share it with other large loads. About 6,800 BTU/h of heat at 2 kW belongs in a technical room or an air-conditioned office. Our article on RTX PRO 6000 Max-Q power on 230 V covers cords and UPS sizing for the same 300 W class.

Four RTX PRO 5000 48 GB or two RTX PRO 6000 Max-Q

Two Max-Q cards bring the same 192 GB into a tower at half the board power, which makes them the comparison to settle before ordering four 48 GB cards.

NVIDIA product pages, Max-Q datasheet and MIG user guide; conversations by our users-per-card rule; Max-Q figures as in our two-card Max-Q article.

For one large model the Max-Q pair is ahead: a 70B model in FP8 fits each card, and as one TP=2 copy it holds more conversations than four 48 GB cards under TP=4, while the pair keeps vLLM's faster all-reduce. It draws about 1.1 kW at the wall instead of about 2 kW and leaves two slots free for a later upgrade. The four-card build is ahead where four independent services run at once, since the four cards have 5,376 GB/s between them against 3,584 GB/s, and on summed FP4 compute. Both give eight 24 GB MIG instances. Our article on two RTX PRO 6000 Max-Q in one workstation covers the pair, and we quote both builds side by side when the choice is open.

Other builds to consider

• Four RTX PRO 5000 72 GB: 288 GB, Llama 3.3 70B in NVFP4 on each card, two FP8 pairs for 88 conversations, vGPU on Red Hat KVM; see four RTX PRO 5000 72 GB in one workstation.
• Four RTX 6000 Ada: the same 192 GB at 960 GB/s per card, no MIG, but on NVIDIA's vGPU list from 15.2; see our page to buy an RTX 6000 Ada workstation.
• Four RTX PRO 6000 Max-Q: 384 GB at the same 1,200 W, for 70B and 235B-class models with many users.
• A rack server: for a service in continuous operation, RTX PRO 6000 Server Edition cards with vGPU, or the RTX PRO 5000 in a rack chassis, as our page on an RTX PRO 5000 rack server describes.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build four-card RTX PRO 5000 48 GB workstations to order on Threadripper PRO or Xeon W platforms, assembled and tested before delivery, with the operating system, driver, CUDA, a container runtime and MIG instances set up on request, and we supply the cards for towers your company already runs. Tell us the tower or platform you prefer, and we check slots, power supply and cables before quoting. Models, retrieval and the platform on top are our Private AI/ML service.

FAQ