Percepta / Workstation GPU
SKU:90YV0MN0-M0NA00
32GB GDDR6 and 640GB/s memory bandwidth on an AMD RDNA 4 workstation card for local AI with ROCm. Blower-cooled and dual-slot for multi-card workstations. ASUS Turbo board, part number 90YV0MN0-M0NA00.
Explore the detailsIn stock
Couldn't load pickup availability
Workstation graphics / AMD RDNA 4
The ASUS Turbo Radeon AI PRO R9700 puts AMD's RDNA 4 workstation GPU and 32GB of memory on a full-height, dual-slot card. It suits local LLM inference, AI development with AMD ROCm and professional graphics. Its blower cooler pushes most of its heat out through the rear bracket, a design AMD describes as suited to workstations running several cards side by side.
This is the ASUS board, part number 90YV0MN0-M0NA00.
On Linux, AMD ROCm supports the R9700 with PyTorch, TensorFlow and vLLM, plus JAX and llama.cpp for inference. On Windows 11, PyTorch runs for inference, and desktop LLM apps such as LM Studio run through Vulkan. The 32GB memory budget has to cover the model, its context and working buffers.
128 AI accelerators handle FP16, FP8 and INT8 matrix work. AMD rates the card at 191 TFLOPS FP16 and 383 TOPS INT8 for matrix operations, doubling with structured sparsity. PyTorch on ROCm uses the familiar torch.cuda interface, so many PyTorch scripts run without code changes.
AMD's ROCm multi-GPU support on Linux covers up to four cards. There is no bridge: the cards communicate over PCIe, and AMD requires CPU-attached slots with identical lane widths that support PCIe atomics. The blower cooler exhausts most of its heat out through the rear bracket, so cards can sit next to each other in a case with good airflow.
Specifications for one card
| GPU | AMD Radeon AI PRO R9700, RDNA 4 architecture |
|---|---|
| Board | ASUS Turbo, part number 90YV0MN0-M0NA00 |
| Compute units | 64 |
| Stream processors | 4,096 |
| AI accelerators | 128, second generation |
| Ray accelerators | 64 |
| GPU memory | 32GB GDDR6; ECC supported on Linux |
| Memory interface | 256-bit |
| Memory bandwidth | 640GB/s |
| AMD Infinity Cache | 64MB |
| Boost clock | Up to 2,920MHz |
| FP32 vector | 47.8 TFLOPS |
|---|---|
| FP16 matrix | 191 TFLOPS; 383 TFLOPS with structured sparsity |
| FP8 matrix | 383 TFLOPS; 766 TFLOPS with structured sparsity |
| INT8 matrix | 383 TOPS; 766 TOPS with structured sparsity |
| INT4 matrix | 766 TOPS; 1,531 TOPS with structured sparsity |
AMD's peak figures for one GPU, not application benchmarks. Actual performance depends on the model, precision and software.
| Host interface | PCIe 5.0 x16 |
|---|---|
| Display outputs | 3x DisplayPort 2.1a, 1x HDMI 2.1b |
| Maximum displays | 4 |
| Video engines | H.264, HEVC and AV1 encode and decode |
| Multi-GPU | Up to 4 cards with ROCm on Linux, over PCIe; no bridge |
| Operating systems | Windows 10 and 11 64-bit; Linux x86 64-bit |
| Drivers and software | AMD Software: PRO Edition (Windows); Radeon Software for Linux; AMD ROCm |
| Total board power | 300W (AMD specification) |
|---|---|
| Power connector | 1x 16-pin 12V-2x6, at the far end of the card |
| Recommended PSU | 750W minimum for one card (AMD) |
| Cooling | Blower fan, front-to-back airflow |
| Form factor | Full height, dual slot |
| Dimensions | 266.7 x 111.1 x 40mm |
Included
One ASUS Turbo Radeon AI PRO R9700 32GB graphics card. ASUS lists a quick setup guide in the box; no power adapter is listed.
Plan your installation
Not directly: the R9700 uses AMD ROCm rather than NVIDIA CUDA. PyTorch on ROCm keeps the torch.cuda interface, so many PyTorch projects run unchanged, and vLLM, TensorFlow and llama.cpp support ROCm on Linux. Check that your applications and any custom CUDA code support AMD GPUs.
Yes. AMD Software: PRO Edition supports the card for professional applications on Windows 10 and 11. For AI work on Windows 11, AMD supports PyTorch for inference, not training, using the graphics driver named in AMD's PyTorch-on-Windows guide, and LLM apps such as LM Studio can run through Vulkan. For training, the wider ROCm framework stack and multi-GPU work, use Linux.
Yes. AMD's ROCm multi-GPU support on Linux covers up to four cards, and AMD advises running no more than two compute workloads at the same time; multi-GPU is not supported under WSL. The cards communicate over PCIe, so AMD requires CPU-attached slots with identical lane widths that support PCIe atomics. Size the PSU from AMD's 750W minimum for one card, adding at least 300W for each additional card, with one native 16-pin cable per card.
AMD designs the Radeon AI PRO R9700 for workstations and does not recommend it for data centre use. For servers, choose a passively cooled data centre GPU.
One 16-pin 12V-2x6 connector at the far end of the card. Use a PSU with a native 16-pin GPU cable; AMD recommends at least 750W for a system with one card.
Discuss the details
Tell us about your application or the system you are putting together.
Workstation GPU
Condition: New - OEM/bulk packaging
Workstation GPU
Condition: New - OEM/bulk packaging
Workstation GPU
Condition: New - OEM/bulk packaging
