AMD Introduces Radeon AI PRO R9700 — 32GB and 1531 TOPS for Local AI Models, 5x Faster than RTX 5080
AMD has officially announced Radeon AI PRO R9700 video card, the first model in a new series aimed exclusively at AI workloads and AI-accelerated computing. The new product is built on architecture RDNA4, is equipped with 32 GB of GDDR6 video memory and offers performance up to 1531 TOPS (INT4 Sparse) — 2 times more than the previous generation.
Radeon AI PRO R9700 uses the chip Navi 48 with 64 compute units (4096 stream processors) and 128 hardware AI accelerators. The card has TBP up to 300W, 256-bit memory bus and total FP16 performance 96 TFLOPS, making it particularly effective for running advanced local models.
According to AMD, the R9700 is aimed at running neural networks of the level DeepSeek R1 Distill Qwen 32B Q6, Mistral Small 3.1 24B Instruct, Flux 1 Schnel and SD 3.5 Medium, where 32 GB of video memory becomes the necessary minimum. In practice, the R9700 demonstrates 2x performance compared to Radeon PRO W7800 32 GB, and also outperforms the RTX 5080 with 16GB up to 5x in VRAM-sensitive models.
The special feature of the card is support 4-fold multi-GPU-configurations via PCIe 5.0, allowing you to combine up to 128 GB of video memory - this is enough to load models of the level Mistral 123B and DeepSeek R1 70B, which require more than 110 GB.
The model replaces the old Radeon PRO and WX series and marks transition to the new Radeon AI PRO brand line, which is entirely aimed at developers, researchers and enthusiasts of local AI. Radeon AI PRO R9700 is made in dual-slot form factor with a turbine cooler, and will go on sale in July 2025 through partners such as ASUS, ASRock, Gigabyte, PowerColor, Sapphire, XFX and Yeston.
