NVIDIA RTX PRO 5500 vs RTX PRO 5000 72 GB: 84 GB at up to 600 W or 72 GB at 300 W

The decision in one table

The NVIDIA RTX PRO 5500 Blackwell has 84 GB of GDDR7 at 1,398 GB/s and up to 600 W; the RTX PRO 5000 Blackwell 72 GB has 72 GB at 1,344 GB/s and 300 W. The 5500 adds 12 GB per card, enough to hold a 70B model in FP8 with a little cache, at twice the power ceiling, while the RTX PRO 5000 72 GB has a full datasheet, vGPU support on Red Hat KVM and fits four cards into a tower on one office circuit. As of 24 October 2026 NVIDIA's product page marks the 5500's specifications as preliminary, so every 5500 figure here is dated and the final specifications are confirmed with the quote.

NVIDIA RTX PRO 5500 and RTX PRO 5000 product pages, read 24 October 2026; NVIDIA vGPU supported-GPU list (2 October 2026); conversations and power by our estimate as explained below.

The rest of this article sets out what NVIDIA has published for each card, what fits in 84 and 72 GB, and how two- and four-card builds compare on power and sharing. Our RTX PRO 5500 overview follows the launch status of the newer card.

What NVIDIA publishes for each card

For the RTX PRO 5500, NVIDIA's product page, read on 24 October 2026, gives the memory, bandwidth, power and interface, and the MIG layout in its FAQ; these are the only 5500 figures we use. A footnote reads "Preliminary product specifications, subject to change", and the page shows "Coming Soon" with a "Notify Me" sign-up. The RTX PRO 5000 page carries a "Buy Now" link and, with the datasheet, the full set of figures.

NVIDIA product pages for both cards, read 24 October 2026; RTX PRO 5000 cores and MIG size from its datasheet. All RTX PRO 5500 figures are preliminary per NVIDIA.

Figures for the 5500's core count and memory bus circulate in partner listings and press reports. NVIDIA has not confirmed them, so we leave them out, as we do board size, cooler, video engines and outputs until a datasheet appears; for a tower whose card length is close to the limit, we confirm the dimensions of the ordered version before quoting.

Memory: what 84 GB holds that 72 GB does not

We size both cards with our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card, must hold the weights and the KV cache. A 96 GB card reports 95.6 GiB, and we scale in proportion, 71.7 GiB for the 72 GB card and 83.7 GiB for the 84 GB card, since no driver figure for the 5500 is published yet. That leaves 61.5 GiB and 72.3 GiB usable. The rule is derived in our guide to how many users one RTX PRO 6000 can serve.

Our estimates by the users-per-card rule; concurrent conversations of 8K tokens with an FP8 KV cache, gpt-oss-120b at 32K with a 16-bit cache of 1.125 GiB. The 5500 figures rest on NVIDIA's preliminary 84 GB; weights as in our model guides.

The 12 GB make a difference at two points. Llama 3.3 70B in FP8 fits a 5500 with about 4.6 GiB of cache, three conversations, which suits one developer or a test but not a team; on the 72 GB card it does not load. gpt-oss-120b leaves 11.5 GiB on the 5500 and 0.7 GiB on the 72 GB card, so only the 5500 serves it from one card. For a team using a 70B model, both cards need a pair or the 4-bit checkpoint, and a 96 GB RTX PRO 6000, in any of its three editions, holds the FP8 model with cache for twelve conversations. Tell us the model and the number of people using it, and we check which card holds it before quoting.

Bandwidth, compute and speed

Token generation reads all active weights once per token, so on a model that fits both cards the rate follows bandwidth, and the 5500 has only 4 per cent more. By our arithmetic one stream of Llama 3.3 70B in NVFP4, about 40.6 GB per token, has a ceiling of about 34 tokens per second on the 5500 and 33 on the RTX PRO 5000. Llama 3.3 70B in FP8 on one 5500, 70.6 GB per token, is held to about 20. These are upper bounds from the specifications; our comparison of GPU memory bandwidth sets both cards against the rest of the range.

Prompt processing, batch throughput, rendering and fine-tuning depend on compute, and there the comparison cannot be made from NVIDIA's figures. NVIDIA rates the RTX PRO 5000 at 2,064 AI TOPS in FP4 with sparsity and publishes nothing comparable for the 5500. The 600 W ceiling suggests more sustained compute, but how much the extra 300 W buys stays open until NVIDIA publishes clocks and TOPS, and we found no independent test of the 5500. For the RTX PRO 5000, published single-card results are in our RTX PRO 5000 benchmarks.

Two and four cards: power and the circuit

The power ceiling separates the two cards most clearly in multi-card builds. Our estimate adds the cards, a 350 W workstation processor and 15 per cent for memory, drives, fans and supply losses, at full load.

Our estimate from NVIDIA's board power figures (the 5500's "up to 600 W" is preliminary) and the 350 W TDP of the Threadripper PRO 9975WX, plus 15 per cent; usable memory by our users-per-card rule, ECC off.

Two 5500 cards draw what four RTX PRO 5000 draw. Four 5500 at full power come to about 3.2 kW by our estimate, above the 2.94 kW we plan per 230 V 16 A circuit at 80 per cent, so they belong on a rack position with a 32 A or three-phase feed, with the heat of about 10,800 BTU/h going into a server room. Four RTX PRO 5000 72 GB stay on one office circuit. NVIDIA's 5500 page says nothing about the number of cards per system, and our searches on 24 October 2026 found no workstation or server maker's specification that lists the card, so the count per chassis is a question for the system maker once it qualifies the 5500. For the RTX PRO 5000 72 GB, Dell's Precision 7960 configurator offers four, as our page to buy an RTX PRO 5000 72 GB workstation shows.

On memory, two 5500 hold Llama 3.3 70B in FP8 as a TP=2 pair with cache for 61 conversations by our rule, two RTX PRO 5000 72 GB for 44. Over four cards, Qwen3-235B-A22B-2507 in FP8, about 219.8 GiB, leaves cache for about 94 conversations on 5500 cards and about 35 on 72 GB cards. Our guide to four RTX PRO 5000 72 GB in one workstation covers that build in detail.

The rest of the system scales with the cards either way. NVIDIA's configuration guide for certified systems asks for system memory of at least twice the GPU memory and at least six processor cores per GPU: 576 GB and 24 cores for four RTX PRO 5000 72 GB, 672 GB and 24 cores for four 5500, both of which we would plan as 768 GB in eight modules. Four cards at x16 take 64 PCIe 5.0 lanes, which a Threadripper PRO 9975WX with its 128 lanes provides. The power supply is the part that differs: a four-card 5500 build needs about twice the supply capacity of the RTX PRO 5000 build, and two supplies on separate feeds where the chassis offers them.

MIG, vGPU and shared use

NVIDIA's 5500 page gives "one instance with up to 84 GB, or two instances with up to 42 GB each". The RTX PRO 5000 page gives "up to two (2) fully isolated instances", which its datasheet sizes at 36 GB on the 72 GB card. NVIDIA's MIG user guide, updated on 13 October 2026, lists neither the 5500 nor the 72 GB RTX PRO 5000, only the 48 GB version, so profile names and prerequisites for both cards are still to come. By our rule a 42 GB half keeps about 34.6 GiB for weights and cache and holds Qwen3-32B in FP8 with room for about two conversations; a 36 GB half, about 29.3 GiB, does not hold that model in FP8 but takes Qwen3-14B in FP8 for about 22.

vGPU is where the RTX PRO 5000 72 GB has the documented lead. NVIDIA's vGPU list, updated on 2 October 2026, includes it from vGPU 20.2 on Red Hat Enterprise Linux with KVM, and does not list the 5500. NVIDIA describes the 5500 as "designed for rack-mounted workstation deployment", where "IT teams can centralize GPUs into managed racks, share them across users". Without vGPU, sharing a 5500 means MIG instances or whole cards passed to users. For virtual desktops on VMware vSphere neither card is listed; our guide to MIG and vGPU: VMs per GPU covers the Server Edition cards that are.

Which card for which use

• A 70B model in FP8 on one card, for one user or tests: the RTX PRO 5500; for a team on one card, a 96 GB RTX PRO 6000 Max-Q in a workstation or Server Edition in a rack.
• Llama 3.3 70B in NVFP4 or 32B models per card for a team: the 5500 holds about 40 to 50 per cent more conversations per card; the 72 GB card does the same job for fewer users at 300 W.
• Four cards in a tower on an office circuit: the RTX PRO 5000 72 GB.
• Graphics virtual machines on Red Hat KVM: the RTX PRO 5000 72 GB, the only one of the two on NVIDIA's vGPU list.
• Remote workstations in a rack, or liquid cooling: the 5500, which NVIDIA designed for this; plan it once a system maker lists it and NVIDIA's final specifications are out.
• Compute-bound rendering and simulation: open until NVIDIA publishes compute figures for the 5500.
• An order with a complete datasheet today: the RTX PRO 5000 72 GB.

Our RTX PRO 5500 vs RTX PRO 6000 Max-Q comparison sets the 5500 against the 300 W 96 GB card, and RTX PRO 5000 with 48 GB or 72 GB covers the smaller memory version.

What we supply

We supply the RTX PRO 5000 72 GB and quote the RTX PRO 5500, as cards or in workstations and servers built to order, on one EU contract and invoice with manufacturer warranty. For the 5500, the final specifications are confirmed with the quote, together with quantities and lead time. Send us the models, scenes or solvers and the number of users, and we compare the two cards for your workload in one quote. Our page to buy an RTX PRO 5500 workstation collects what is known about systems with the newer card.

FAQ