NVIDIA slotted a new card between the RTX PRO 5000 and the RTX PRO
6000 in September 2026: 84GB GDDR7 ECC at 1,398 GB/s, 21,760 CUDA
cores (the RTX 5090 count), 600W, PCIe Gen 5. Listed as “Coming Soon”
on nvidia.com with specs marked preliminary and no MSRP published.
The price field here stays empty until NVIDIA names a number.
What it does well:
-
84GB single-card capacity: 70B dense at 8-bit, 235B-class MoE
at 4-bit, with ECC GDDR7 and full CUDA. MIG can split it into two
isolated 42GB instances.
-
2.6x the RTX 5090’s memory at the same core count: the 32GB
consumer flagship’s compute with the headroom it lacks - at
1,398 GB/s instead of 1,792.
Where it falls short: the bandwidth cut. 1,398 GB/s is 22% under
both the RTX 5090 and the RTX PRO 6000 (1,792 GB/s). Same core count
as the 5090, but LLM decode is memory-bandwidth-bound, so tok/s
tracks the bus, not the cores. Tom’s Hardware reads the board as 28
GDDR7 chips clamshelled on a 448-bit bus, downclocked from 28 to
25 Gb/s - the same bandwidth as the China-exclusive RTX PRO 6000D,
the only other 84GB card NVIDIA has shipped. (PNY’s sheet lists
416-bit, which cannot hold 84GB clamshelled, so treat Tom’s as the
working read until NVIDIA publishes.) No price yet: the RTX PRO 6000
above it has 12GB more memory, 1,792 GB/s, and 24,064 cores at
$8,565 MSRP, so the 5500 has to undercut that by enough to matter.
Specs are preliminary.
The fit case, from a Spark owner: “it is the first single card
that holds my entire single spark board of open model” - the
100B-class checkpoints @sudoingX runs (Laguna at 67GB, Ling int4 at
72GB, Qwen 3.5 122B at 74GB) all fit with context to spare, on a
bus 5x the Spark’s 273 GB/s. His caveat stands: “the rest is a
benchmark nobody has run.” And the 180B-at-FP8 and 1M-context
builds still need 96GB or more, which keeps them on the desk boxes.
Run it locally: full CUDA stack - llama.cpp, Ollama, vLLM. The
modeldex results rank checkpoints by fit
and estimated tok/s.