NVIDIA RTX PRO 5000 gets a high memory variant

NVIDIA has quietly added a 72GB version of the RTX PRO 5000 Blackwell to its lineup. The product page for the 48GB model has been updated to include the new version with increased memory capacity, but there has been no formal announcement so far.
The 72GB version is listed with a 512-bit memory bus, the same as the 48GB model. The 48GB variant uses 16x GDDR7 chips, each with 3GB capacity. With a 512-bit bus and 1.344 TB/s bandwidth, the card uses 21 Gbps modules, slower than gaming variants, though this memory includes ECC support.
For the RTX PRO 5000 72GB to achieve the same bandwidth, NVIDIA would need to use one of the following configurations:
- 24 × 3GB chips (12 channels in full clamshell, 384-bit) with 28 Gbps memory, or
- 24 × 3GB chips (16 channels, with clamshell on 8 channels, 512-bit) at 21 Gbps. ⬅️
If NVIDIA’s listed specs are accurate, this model uses the second option with clamshell mode. The company mentions the updated memory in its newly published datasheet.
GDDR7 Memory: New and improved GDDR7 memory significantly boosts bandwidth and capacity, empowering your applications to run faster and work with larger, more complex datasets. With 48 or 72 GB of GPU memory and 1.3 TB/s of bandwidth, tackle massive 3D and AI projects, fine-tune AI models locally, explore large-scale VR environments, and drive larger multi-app workflows.
The official specs confirm that both models have the same CUDA core count of 14,080, meaning both are based on the GB202 GPU with a 512-bit bus.

Source: NVIDIA
There’s no change to TDP, both are rated at 300W , and no difference in TFLOPS, suggesting identical clock speeds. The new 72GB model is likely aimed at AI inference workloads requiring a larger memory buffer.
The current 48GB version is listed at around $4,250 to $4,600, so the 72GB model could be priced close to $5,000. For reference, the flagship RTX PRO 6000 costs over $8,300.
| NVIDIA RTX Workstation GPUs | |||||
|---|---|---|---|---|---|
| VideoCardz.com | Picture | GPU | CUDA Cores | Memory | TDP |
| BLACKWELL | |||||
| RTX PRO 6000 Workstation Edition | ![]() | GB202 | 24064 | 96GB G7 ECC | 600W |
| RTX PRO 6000 Server Edition | | GB202 | 24064 | 96GB G7 ECC | 600W |
| RTX PRO 6000 Max-Q Edition | ![]() | GB202 | 24064 | 96GB G7 ECC | 300W |
| RTX PRO 5000 72GB⬅️ | | GB202 | 14080 | 72GB G7 ECC | 300W |
| RTX PRO 5000 | ![]() | GB202 | 14080 | 48GB G7 ECC | 300W |
| RTX PRO 4500 | ![]() | GB203 | 10496 | 32GB G7 ECC | 200W |
| RTX PRO 4000 | ![]() | GB203 | 8960 | 24GB G7 ECC | 140W |
| RTX PRO 4000 SFF 🆕 | ![]() | GB203 | 8960 | 20GB G6 ECC | 70W |
| RTX PRO 2000 🆕 | ![]() | GB206 | 4352 | 16GB G7 ECC | 70W |
Source: NVIDIA via TechPowerUP





