NVIDIA RTX PRO 6000 Blackwell is coming to FlexiCloud

It is the data-centre version of NVIDIA’s RTX PRO 6000 Blackwell: a passively cooled, dual-slot PCIe Gen 5 card with 96 GB of GDDR7 memory, fifth-generation Tensor Cores with FP4, fourth-generation RT Cores and dedicated video engines. NVIDIA positions it as a universal data-centre GPU for AI inference, graphics, simulation and video, and it ships in servers from the major manufacturers with two, four or eight GPUs.

It suits inference of small and mid-size models (a 70B-parameter model fits on one card at FP8), multimodal and video AI, rendering, simulation and digital twins, and virtual workstations. Light fine-tuning of smaller models on a single GPU is fine. It is not the right choice for pre-training or large fine-tunes across many GPUs: it has no NVLink, so the GPUs talk over PCIe and cannot pool memory at speed. For that work, our B300, B200, H200 and H100 nodes are available now.

Per GPU, from NVIDIA’s published specifications: 96 GB of GDDR7 with ECC on a 512-bit bus, 1,597 GB/s of memory bandwidth, 24,064 CUDA cores, 752 fifth-generation Tensor Cores and 188 fourth-generation RT Cores. NVIDIA’s peak figures are up to 4 PFLOPS at FP4 and 2 PFLOPS at FP8 (both with sparsity), and about 120 TFLOPS at FP32. It draws up to 600 W (configurable), connects over PCIe Gen 5 x16, has four NVENC and four NVDEC engines, supports up to four MIG instances of 24 GB each, and has no NVLink.

The H100 and H200 are SXM parts with HBM memory and NVLink, so eight GPUs work together as one machine for training and for large models. The RTX PRO 6000 is a PCIe card with GDDR7: more memory per GPU than the H100 (96 GB against 80 GB) but lower memory bandwidth, no NVLink, and RT Cores and video encoders that the H100 and H200 do not have. Choose the H100 or H200 for training and for models split across GPUs; choose the RTX PRO 6000 for inference that fits on one card, graphics, video and simulation.

The card itself is shipping now. We are finalising pricing, terms and regions with our partner providers, so pricing is to be announced and we will not quote a number until it is firm. Register your interest and you will be among the first to get a firm quote. If you need compute sooner, we can put you on an H100 or H200 node now.

Yes. Multi-Instance GPU (MIG) splits each card into up to four fully isolated instances of 24 GB, each with its own memory, cache and compute, so an eight-GPU server can run up to 32 isolated workloads. The card also supports NVIDIA’s virtual workstation and virtual PC software (vWS and vPC) for VDI, which NVIDIA licenses separately.

FlexiCloud sources GPU capacity across a network of more than ten partner providers in India, the EU and the US. You get one relationship and one quote across all of them, at the same price as going direct or lower, because we are paid by the providers, not by you. Your requirement stays in confidence instead of going to ten sales teams, and you are not tied to one data centre or region.