The NVIDIA RTX PRO 6000 Blackwell Server Edition pairs 96 GB of GDDR7 with 24,064 CUDA cores. That is more GPU memory than an H100 carries, and double what the 48 GB professional cards of the previous two generations offered.
For model work the 96 GB is the headline. It is double the 48 GB that the RTX A6000, L40S and RTX 6000 Ada have offered as the practical ceiling, and that difference decides which models run on one card rather than two. This guide covers the specifications, what actually fits in 96 GB, and how the card compares to the 48 GB generation and to the H100.
RTX PRO 6000 specs and VRAM
NVIDIA's datasheet for the RTX PRO 6000 Blackwell Server Edition lists 96 GB of GDDR7 with ECC, a 512-bit memory interface, 24,064 CUDA cores and 1,597 GB/s of memory bandwidth.
Source: NVIDIA RTX PRO 6000 Blackwell Server Edition datasheet, April 2025.
What fits in 96 GB
The jump from 48 GB to 96 GB changes which models run on one card rather than two.
- 70B-class models at 4-bit fit comfortably, with room for a working context.
- 70B at 8-bit fits, which is the practical reason to want this card over a 48 GB one. You stop trading quantization quality for capacity.
- 30B-class models at 16-bit fit without quantization at all.
- Long-context inference benefits more than parameter count does. The key-value cache grows with context length and concurrency, and 96 GB is what lets you hold a large cache instead of evicting it.
Treat these as capacity guidance rather than benchmarks. We have not published measured throughput for this card, and we are not going to imply numbers we have not run.
RTX PRO 6000 vs RTX A6000
These two get confused constantly, including by search engines, so it is worth being blunt: they are not the same card and not the same generation.
If a model fits in 48 GB and you are not memory-bound, the A6000 remains the far cheaper way to run it. The PRO 6000 earns its rate when you need the 96 GB or the bandwidth, not as a general upgrade.
RTX PRO 6000 vs L40S and RTX 6000 Ada
Both are 48 GB Ada-generation cards, at $1.09/hr and $1.09/hr on Runpod. Against them the PRO 6000 offers double the memory and a real bandwidth advantage, though the gap is narrower than the memory gap: 1,597 GB/s against 864 GB/s on the L40S and 960 GB/s on the RTX 6000 Ada.
The decision is the same one as above. If your model and its key-value cache fit in 48 GB, the cheaper cards win on cost per unit of work. If they do not, you are otherwise looking at splitting across two GPUs, which costs you both money and inter-GPU overhead.
RTX PRO 6000 vs H100
This is the comparison most people actually mean, and the answer is not a straight ranking.
The H100 has HBM and NVLink. The PRO 6000 has more raw memory, 96 GB against the H100's 80 GB, but GDDR7 and no NVLink. For single-card inference where the model fits, the PRO 6000's extra capacity can matter more than the H100's bandwidth advantage. For multi-GPU training, the H100's interconnect is the point, and the PRO 6000 is not trying to compete there.
On Runpod the H100 PCIe is $2.89/hr and the H100 SXM $4.39/hr, against $2.69/hr for the PRO 6000.
Server Edition, Workstation Edition and MIG slices
Runpod offers both editions. The Server Edition is the passively cooled card intended for rack deployment. The Workstation Edition carries the same 96 GB of GDDR7, and the difference you will notice on Runpod is the host around the card rather than the GPU itself. System RAM and vCPU vary by machine on both editions, so check the configuration shown on the specific instance at deploy time rather than assuming one edition is always the larger host.
You can also take the card in slices. NVIDIA supports up to four fully isolated MIG instances per card, and on Runpod MIG partitions are available at 48 GB ($1.39/hr) and 24 GB ($0.69/hr). That is the cheaper route when you want Blackwell-generation compute without needing the whole 96 GB.
Running the RTX PRO 6000 on Runpod
Deploy the RTX PRO 6000 from the GPU models page, billed per second with no commitment. Current rates for every card are on the pricing page.
RTX PRO 6000 FAQs
How much VRAM does the RTX PRO 6000 have?
96 GB of GDDR7 with ECC, on a 512-bit interface, with 1,597 GB/s of memory bandwidth.
What is the difference between the RTX PRO 6000 and the RTX A6000?
Two generations and double the memory. The PRO 6000 is Blackwell with 96 GB GDDR7 and 1,597 GB/s of bandwidth. The A6000 is Ampere with 48 GB GDDR6 and 768 GB/s. They are frequently confused because of the similar name.
What is the difference between the Server Edition and the Workstation Edition?
Both carry the same 96 GB of GDDR7. The Server Edition is passively cooled for rack deployment. On Runpod the practical difference is the host configuration around the card, and system RAM and vCPU vary by machine, so check the instance you are about to deploy.
How many CUDA cores does the RTX PRO 6000 have?
24,064, alongside 752 fifth-generation Tensor cores and 188 fourth-generation RT cores.
Is the RTX PRO 6000 better than an H100?
For single-card work where you need more than 80 GB, the PRO 6000's 96 GB can be the deciding factor. For multi-GPU training, the H100's HBM and NVLink matter more. Neither is strictly better; they are built for different shapes of job.
Can I run a 70B model on an RTX PRO 6000?
Yes. A 70B-class model fits at 4-bit with room for context, and at 8-bit as well, which is the main practical advantage over a 48 GB card.
Can I rent part of an RTX PRO 6000?
Yes. NVIDIA supports up to four fully isolated MIG instances per card, and Runpod offers 48 GB and 24 GB partitions, which are cheaper than taking the full card when you do not need all 96 GB.
What is the power draw of the RTX PRO 6000?
Up to 600 W, configurable, per NVIDIA's datasheet. The card uses a passive thermal solution, so the deployed figure depends on the chassis it sits in.
Related guides
- NVIDIA RTX A6000 GPUs
- NVIDIA RTX 5090
- NVIDIA H100
- NVIDIA H200
- Best GPU for LLM inference
- GPU memory sizing guide for LLM inference
Author profile: The Runpod Team
