32 GB of GDDR7 and MIG for AI inference in the rack at 165 W
The RTX PRO 4500 Server Edition covers the task where many parallel AI workloads are required on limited power: data center, cloud, edge nodes. The NVIDIA Blackwell architecture provides 10,496 CUDA cores, 328 5th-generation Tensor cores and 82 4th-generation RT cores, while 32 GB of GDDR7 with ECC on a 256-bit bus push 800 GB/s. This SKU, TCSRTXPRO4500SE-PB, is the retail version in PNY branded packaging with the full warranty package from the distributor ASBIS.
The card is made single-slot in a passive FHFL form factor with a 165 W TGP, so several accelerators fit into a node without competing for slots and power. Connection - PCI Express 5.0 x16, power through a single PCIe CEM5 16-pin connector. There are no display outputs: this is a headless card for compute, not for a monitor.
- NVIDIA Blackwell: 10,496 CUDA / 328 Tensor (5 gen) / 82 RT (4 gen)
- 32 GB GDDR7 with ECC, 256-bit, 800 GB/s
- Single-slot passive, FHFL, 165 W TGP, PCIe 5.0 x16
- MIG: up to 2 isolated instances of 16 GB each for vGPU/VDI
- NVENC 9 gen / NVDEC 6 gen with 4:2:2; NVLink is not supported
Confidential Computing and the 2nd-generation Transformer Engine keep the inference of generative models secure, MIG splits the resource between two independent users, and the hardware video engines offload 4:2:2 transcoding. We will advise on vGPU profiles and compatibility with your server platform - an engineer is on the line at +380 93,594 00 77.
