RTX PRO 5000 Blackwell: 48 GB or 72 GB, and what the extra 24 GB actually buys

One card, two memory sizes

The RTX PRO 5000 Blackwell arrived with 48 GB in 2025. The 72 GB version appeared on NVIDIA’s site in October 2025 and NVIDIA announced its general availability on 18 December 2025; the current datasheet covers both in one table.

NVIDIA RTX PRO 5000 Blackwell datasheet (June 2026), MIG user guide and vGPU documentation (September 2026). NVIDIA’s December 2025 launch blog quotes 2,142 TOPS for the 72 GB card; every NVIDIA specification table gives 2,064.

NVIDIA’s MIG guide lists the RTX PRO 5000, in its 48 GB version, on the GB202 chip, the same die as the RTX PRO 6000, and the datasheet gives both versions the same 14,080 CUDA cores. Everything that decides speed is identical between the two versions. Everything that decides what fits is not.

The same speed when the model fits

Token generation reads the weights once per token, so with equal bandwidth the two versions generate at the same rate on any model that fits both. Puget Systems tested both in Topaz Video’s Starlight models and, in Starlight Mini, found them “less than 1% apart, showing that VRAM capacities this size did not affect performance in our testing”. For language models we found no published figure for either version that states the engine and its version, the precision, the concurrency and the test system in text: a vLLM benchmark of the 48 GB card names the engine, the precision and the concurrency, but not the vLLM version or the test system. By our arithmetic, 1,344 GB/s limits generation of Llama 3.3 70B in NVFP4, which reads about 40.6 GB per token, to about 33 tokens per second for a single user, on both cards alike; that is a ceiling from the specifications, not a measurement.

The same holds for rendering. Puget measured the RTX PRO 5000, tested with 48 GB, at 50 per cent above the RTX 5000 Ada in Blender and 35 per cent above it in V-Ray GPU, where it also beat the older RTX 6000 Ada by 11 per cent. A scene that fits in 48 GB should render no faster on 72, although no rendering test of the 72 GB card has been published yet.

What the extra 24 GB holds

The difference shows up when the weights and the cache no longer fit. The table uses the method of our 70B sizing example: weights from the model cards, a tenth of the card reserved for the runtime, and FP8 cache at 160 KiB per token for Llama 3.3 70B, which is 1.25 GiB per 8,192-token conversation.

Our arithmetic from model-card sizes; 32B figures for a 32.8-billion-parameter model such as Qwen3-32B. An FP16 cache instead of FP8 halves the conversation counts. Production servers keep more in reserve than 10 per cent.

The 70B row is the one most buyers care about. On 48 GB a 4-bit 70B model runs for one or two people with a working context; on 72 GB it serves a small team. gpt-oss-120b is the other dividing line: at 60.8 GiB it loads on the 72 GB card with little to spare, and it does not load on the 48 GB card at all.

MIG: two halves of 24 or 36 GB

Both versions split into two isolated instances. NVIDIA’s MIG user guide documents the 48 GB card with the profiles 1g.24gb and 2g.48gb, plus variants with graphics support; the 72 GB card splits into two instances of 36 GB according to NVIDIA’s datasheet, but its profiles are not yet in the MIG guide. NVIDIA’s MIG guide sets the requirements for the 48 GB card: Linux, driver 575.51.03 or later, a vBIOS of 98.02.73.00.00 or later, and the display mode switched from graphics to compute with DisplayModeSelector 1.72 or later, which switches off the card’s display outputs. The 72 GB card is not yet in the guide; NVIDIA staff confirmed in May 2026 that it needs DisplayModeSelector 1.76 or later for that switch.

A 24 GB half takes gpt-oss-20b or a 14B model in FP8. A 36 GB half takes a 32B model in FP8 with a small cache, which is a good reason to pick the larger card for a two-person setup.

vGPU: the 72 GB card’s quiet advantage

Only the 72 GB card is on NVIDIA’s vGPU supported-GPU list. Support starts with vGPU 20.2, released in August 2026, and the release notes limit it to Red Hat Enterprise Linux with KVM 9.6; NVIDIA’s support matrix of 22 September 2026 still lists no other hypervisor for it, so VMware vSphere, Nutanix and Windows Server are not covered. The vGPU 20 user guide mentions the card only in its table of CUDA developer tools, under time-sliced and MIG-backed modes; its MIG profiles are not yet in NVIDIA’s MIG guide, so check the profile list for your vGPU release before planning MIG-backed virtual machines. The 48 GB version has no vGPU support.

For virtual machines on other hypervisors, the cards to look at are the RTX PRO 6000 Server Edition, supported since vGPU 19.0, and the RTX PRO 4500 Server Edition, supported since 20.0. Our MIG and vGPU guide compares them.

Two 72 GB cards or one bigger card

Two 72 GB cards give 144 GB at 600 W of board power, the same power as one RTX PRO 6000 Workstation Edition with 96 GB. The pair suits two separate models or two teams better than one model split across both, because NVIDIA lists no NVLink for the RTX PRO 5000 and a split model talks over PCIe. When one model needs more than 72 GB on one card, the RTX PRO 6000 with 96 GB and 1,792 GB/s is the next step, and the new RTX PRO 5500 with 84 GB sits in between.

Decision rules

Choose 48 GB for models up to 32B in FP8, a 4-bit 70B model for one or two users, rendering scenes that fit, and CAD, where Puget found the professional tiers perform much alike.

Choose 72 GB for a 4-bit 70B model serving a team, for gpt-oss-120b on one card, for 32B models in BF16, for two MIG instances of 36 GB, and for vGPU on Red Hat KVM.

Otherwise the two cards are the same product. Pay for the extra 24 GB only if something will live in it.

What we supply

Eurokommerz supplies both versions of the RTX PRO 5000 Blackwell EU-wide with manufacturer warranty, on their own or in a configured workstation, alongside the rest of the RTX PRO family. Send us the model and the number of users, and we will tell you which memory size the arithmetic points to.

FAQ