RTX PRO 6000 vs RTX 6000 Ada for AI work

RTX 6000 Ada has 48 GB of GDDR6 at 960 GB/s and 300 W. RTX PRO 6000 Server Edition has 96 GB of GDDR7 at 1,597 GB/s, FP4 and MIG. Specs and prices.

Four NVIDIA RTX 6000 Ada Generation cards with blower coolers installed side by side on an open workstation board
Photo: 4× NVIDIA RTX™ 6000 der Ada-Generation (极客湾Geekerwan) 01 by 极客湾Geekerwan, Wikimedia Commons, CC BY 3.0

The names are close, which causes a lot of confusion. RTX 6000 Ada Generation is NVIDIA's professional card of the Ada generation, built on the Ada Lovelace AD102 chip with 48 GB of memory. RTX PRO 6000 Blackwell is its successor, with 96 GB of GDDR7. For AI work the short version is simple: the new card doubles the memory, adds about two thirds more memory bandwidth and supports FP4 and MIG. The older card still makes sense for smaller models and existing workstations.

Which one for whom

  • Models up to about 30B parameters, a single workstation, or a lower hourly price → RTX 6000 Ada. 48 GB at 960 GB/s, 300 W, a blower card that fits in ordinary workstations. On 14 November 2026 RunPod lists it at $0.74 an hour (Community Cloud) and $1.09 (Secure Cloud).
  • 70B-class models, gpt-oss-120b, long context, or several isolated users on one card → RTX PRO 6000. 96 GB at 1,597 GB/s, FP4, up to four MIG instances per card. RunPod lists it at $1.69 and $2.69 an hour. Our server in Sofia has eight RTX PRO 6000 Server Edition cards and rents from $1.80 per GPU-hour, or about $10,510 a month for the whole machine. It is rented into early 2027 and we take reservations for the period after. See the server.

The specifications side by side

RTX 6000 Ada RTX PRO 6000 Server Edition
Architecture Ada Lovelace Blackwell
GPU memory 48 GB GDDR6 with ECC 96 GB GDDR7 with ECC
Memory bandwidth 960 GB/s 1,597 GB/s
FP4 Tensor Core not listed 4 PFLOPS (1)
FP8 Tensor Core 1,457 AI TOPS with sparsity 2 PFLOPS (1)
FP32 91.1 TFLOPS 120 TFLOPS
Max power 300 W up to 600 W, configurable
MIG not listed up to 4 instances
PCIe Gen 4 x16 Gen 5
Form active blower, dual slot dual slot air-cooled, or single slot liquid-cooled
Display outputs 4 × DisplayPort 1.4a 4 × DisplayPort 2.1

(1) NVIDIA's page for the RTX PRO 6000 Server Edition does not say whether these figures assume sparsity. The RTX 6000 Ada figure is with sparsity. Compare the compute rows as rough, not exact.

There are also Workstation and Max-Q versions of the RTX PRO 6000, with the same 96 GB but different coolers and power limits. How they differ is in Max-Q vs Workstation vs Server Edition.

Memory: 48 GB against 96 GB

When a model generates text, its weights must sit in GPU memory, plus a cache for every conversation (the KV cache). As a rough rule, weights take two bytes per parameter in FP16, one in FP8 and about half a byte in FP4. On one card:

  • 7B to 14B models: fit on either card in FP16 or FP8 with room for the cache. Here the older card is enough.
  • About 30B in FP8 (around 30 GB of weights): fits on one RTX 6000 Ada with about 18 GB left for the cache. On RTX PRO 6000 about 66 GB stays free, so many more parallel conversations or a much longer context.
  • 70B in FP8 (around 70 GB of weights): does not fit on one RTX 6000 Ada; you need two cards. On one RTX PRO 6000 the weights fit, but the cache is tight. In FP4 a 70B model takes about 35 GB and runs comfortably on one RTX PRO 6000.
  • gpt-oss-120b: OpenAI's model card says it fits on a single 80 GB GPU. That means one RTX PRO 6000, or at least two RTX 6000 Ada cards.

In an eight-card server the totals are 384 GB against 768 GB. For a per-model guide, see Local LLM vs private LLM server.

Bandwidth sets the speed for one user

For each new token the GPU reads the model's weights again, so for one conversation the speed depends mainly on memory bandwidth. RTX PRO 6000 has 1,597 GB/s against 960 GB/s, about 1.66 times more. With the same model on one card, expect roughly that much faster token generation for one user. This is an estimate from the specifications; we have not run the two cards side by side.

With many users at once, compute and the free memory for the cache matter more. RTX PRO 6000 is ahead on both and adds FP4, which halves the memory per parameter against FP8. NVIDIA lists no FP4 figure for the Ada card.

Splitting one card between users

NVIDIA's page for the RTX 6000 Ada does not list MIG, so one card runs one set of jobs or shares its time between them. RTX PRO 6000 splits into up to four isolated instances, each with its own memory and compute. In an eight-card server that is up to 32 separate GPUs, useful for separate development, test and production environments, or for serving several small models.

Workstation card or server card

RTX 6000 Ada was made for workstations: a blower fan and 300 W, so it runs in a desk tower with a normal power supply. The RTX PRO 6000 Server Edition is built for server chassis, air-cooled or liquid-cooled, and can draw up to 600 W. Eight of them need a rack built for that. On our own machine we measured about 6.5 kW from the wall at full GPU load; the details are in What an 8x RTX PRO 6000 server really draws.

If you already own RTX 6000 Ada cards and your models fit in 48 GB, there is little reason to replace them. The case for the newer card is bigger models, longer context, more users per card, or FP4.

Price: by the hour or a whole machine

On 14 November 2026 RunPod's pricing page lists RTX 6000 Ada at $0.74 an hour on Community Cloud and $1.09 on Secure Cloud, and RTX PRO 6000 at $1.69 and $2.69. Per gigabyte of memory that is about 1.5 cents for the Ada card and 1.8 cents for RTX PRO 6000 on Community Cloud. The newer card costs a little more per gigabyte but keeps the memory in one piece, without splitting a model over PCIe, and gives more bandwidth per card. Prices change often and depend on region and commitment.

Hourly cards suit tests and short jobs. For a model that runs all day for a team or a product, a whole machine at a fixed monthly price is simpler to plan. Our server costs about $10,510 a month for eight RTX PRO 6000 cards, with power, cooling and a fixed address in Sofia, in the EU. Compare the options on our pricing page or read about a dedicated GPU server.

Questions

Is RTX PRO 6000 just a renamed RTX 6000 Ada? No. It is a new generation (Blackwell instead of Ada Lovelace), with twice the memory, GDDR7 instead of GDDR6 and PCIe Gen 5.

Is RTX 6000 Ada the same as RTX A6000? No. RTX A6000 is the older Ampere-generation card, also with 48 GB. RTX 6000 Ada replaced it, and RTX PRO 6000 replaced RTX 6000 Ada.

Can two RTX 6000 Ada cards replace one RTX PRO 6000? For memory, yes: 2 × 48 GB is 96 GB. But the model is then split over PCIe Gen 4, each card has 960 GB/s, and you power and pay for two cards.

How does it compare with L40S? L40S uses the same AD102 chip with 48 GB, aimed at servers. See RTX PRO 6000 vs L40S.

Sources

  • NVIDIA, RTX 6000 Ada Generation: 48 GB GDDR6 with ECC, 91.1 TFLOPS FP32, 1,457 AI TOPS FP8 with sparsity, 300 W, PCIe Gen 4 x16, dual slot, 4 × DisplayPort 1.4a, checked 14 November 2026.
  • PNY, NVIDIA RTX 6000 Ada Generation datasheet: 960 GB/s memory bandwidth, checked 14 November 2026.
  • NVIDIA, RTX PRO 6000 Blackwell Server Edition: 96 GB GDDR7, 1,597 GB/s, 4 PFLOPS FP4, 2 PFLOPS FP8, up to 600 W, up to four MIG instances, PCIe Gen 5, checked 13 November 2026.
  • RunPod, GPU pricing: RTX 6000 Ada $0.74 and $1.09 an hour, RTX PRO 6000 $1.69 and $2.69 an hour (Community and Secure Cloud), checked 14 November 2026.
  • OpenAI, gpt-oss-120b model card: fits on a single 80 GB GPU.
  • NEOXIS measurements of our 8 × RTX PRO 6000 server, published 20 October 2026.
SHARE