NVIDIA Quietly Lists the RTX PRO 5500 Blackwell: 84 GB, Built for the Rack

What NVIDIA and PNY Have Published

Two vendor pages carry specifications. NVIDIA's own page gives the memory, power and media engine details. The page from PNY, NVIDIA's professional board partner, adds the numbers NVIDIA leaves out: core count, die size, transistor count and board dimensions.

NVIDIA's table carries the footnote “Preliminary product specifications, subject to change.” Both pages were read on September 17, 2026.

RTX 5090 Silicon, Professional Memory

NVIDIA does not name the die, but PNY's numbers do it for them. 92.2 billion transistors on 750 mm² is the full-size GB202, the same die under the RTX PRO 6000 Blackwell and the GeForce RTX 5090. And 21,760 CUDA cores is exactly the RTX 5090's count.

So the simplest way to read the RTX PRO 5500 is as the RTX 5090's compute with professional memory and features: 84 GB instead of 32 GB, ECC, MIG partitioning and enterprise drivers. For local LLM work that is the difference that matters. A 70B model in FP8 needs about 70 GB for weights alone, which does not fit on a 32 GB card and leaves room for KV cache on an 84 GB one. You can compare the two side by side.

Where It Sits in the RTX PRO Blackwell Line

Until now the line had a wide gap. The RTX PRO 5000 tops out at 72 GB and 300 W. The next step was the RTX PRO 6000 at 96 GB and 600 W, with 71% more cores. The 5500 lands between them on cores and memory, but at the 6000's power limit.

Figures for the other four cards are from the Flopper database and trace to vendor datasheets. FP32 is peak single precision.

Two things stand out. First, memory bandwidth goes down against the RTX PRO 6000 Workstation Edition, from 1,792 to 1,398 GB/s, a 22% cut. LLM decode speed tracks memory bandwidth closely, so expect the 5500 to generate tokens more slowly than the 6000 even where the model fits on both. Second, the 5500 has the RTX 5090's core count, and the 5090 is rated at 104.8 TFLOPS FP32. That is a reasonable ballpark for the 5500, but it is not a published figure, so it is not in our database.

Side by side: RTX PRO 6000 vs RTX PRO 5500 and RTX PRO 5500 vs RTX PRO 5000 72GB.

The 84 GB Question

84 GB is an unusual capacity, and the published numbers do not quite add up. PNY lists a 416-bit memory interface. GDDR7 devices each use a 32-bit channel, so 416 bits is 13 channels. Even with two devices on every channel, 13 channels hold 26 devices. With today's common 3 GB GDDR7 devices that is 78 GB, not 84.

There are two ways to square it.

• Take 416-bit at face value. Then the board must use 4 GB devices, 21 of them, running at about 26.9 Gbps. That speed matches no other card in the line.
• Read 416 as a typo for 448. HotHardware raised this, noting that NVIDIA's China-market RTX PRO 6000D reaches 84 GB with 28 devices of 3 GB on a 448-bit bus. Our own arithmetic supports it. 1,398 GB/s over 448 bits is 24.96 Gbps per pin. The RTX PRO 6000 Server Edition runs 1,597 GB/s over 512 bits, which is 24.95 Gbps. On this reading the 5500 has exactly seven eighths of the Server Edition's memory system: 28 devices instead of 32, 84 GB instead of 96, 1,398 GB/s instead of 1,597.

We think the second reading is more likely, but it is an inference. Our database stores the 416-bit figure because that is what the only vendor page to print a width says. We store no device count and no memory clock, and we will update the row when NVIDIA publishes a datasheet.

A Workstation Card Sold Into the Rack

The positioning is the real story. NVIDIA describes the RTX PRO 5500 as “designed for rack-mounted workstation deployment with a choice of air- or liquid-cooled thermal solutions, letting teams centralize, share, and scale workstation GPU power across the enterprise.” The liquid-cooled version comes in what NVIDIA calls the RXM form factor, a name with no published specification yet.

In practice that means IT puts the cards in a managed rack, splits each one into two isolated 42 GB MIG instances, and serves workstation sessions to people who never touch the hardware. It is a datacenter sales motion under a workstation product name. Note that the 5500 offers two MIG instances where the RTX PRO 6000 offers four, so the 6000 remains the denser card for serving many small sessions.

What It Means for Buyers

Nothing to act on yet. There is no price, no date and no performance figure, so there is no performance-per-dollar or performance-per-watt case to make in either direction. If you need more than 72 GB on one professional card today, the RTX PRO 6000 Blackwell is the shipping option, and it is available to rent by the hour.

The number to watch is price. The 5500 draws the same 600 W as the 6000, with 10% fewer cores, 12.5% less memory and 22% less bandwidth than the Workstation Edition. It only makes sense if it is priced well below the 6000. If it is, it becomes the cheapest way to put more than 72 GB of ECC memory behind a Blackwell GPU.

Source Confidence

Sources

• NVIDIA (primary): RTX PRO 5500 Blackwell Workstation Edition product page
• PNY: NVIDIA RTX PRO 5500 Blackwell product page
• HotHardware (Sept 14, 2026): NVIDIA's RTX Pro 5500 Blackwell Packs 84GB GDDR7 and RTX 5090 Core Count
• VideoCardz (Sept 14, 2026): NVIDIA introduces RTX PRO 5500 Blackwell with 84GB GDDR7 and 600W power limit