AMD has begun shipping the Radeon AI PRO R9700, a new RDNA4-based graphics card designed for AI workloads, priced starting at $1299 USD. The card features 32GB of GDDR6 memory and targets professional users with enhanced compute capabilities. Initial benchmarks under Linux show its performance in single and dual GPU setups.
The AMD Radeon AI PRO R9700 officially went on sale today, following its announcement at Computex. This AI-focused professional graphics card offers 32GB of GDDR6 video memory, 128 AI accelerators, and is rated for 96 TFLOPs peak half-precision compute and up to 1531 TOPS INT4 sparse. It has a 300 Watt TDP and supports DisplayPort 2.1a outputs. AMD positions it as delivering 2x the performance of the Radeon PRO W7800.
Compared to the prior flagship Radeon PRO W7900, the R9700 is a cut-down version with 32GB of 256-bit GDDR6 versus 48GB of 384-bit, 128 AI accelerators against 192, 4096 stream processors versus 6144, and 96 compute units compared to 64. However, it benefits from the newer RDNA4 architecture and PCI Express 5.0 interface, aiming for superior AI workload performance over the RDNA3-based W7000 series. The GDDR6 memory is ECC capable, though wired up driver-side under Linux. Board power remains similar at 300 Watts versus the W7900's 295 Watts.
Pricing starts at $1299 USD, making it competitive against the NVIDIA RTX PRO 4000 Blackwell at $2499 USD with 24GB VRAM, and far below the W7900's launch price of $3999 USD (now around $3600). The NVIDIA RTX 6000 Ada Generation, used for comparison, retails for over $5300 USD.
Testing involved two R9700 cards on Ubuntu 24.04.3 LTS with ROCm 7.0.2 and the latest AMDGPU DKMS driver. The setup worked smoothly, supporting upstream drivers from Ubuntu 25.10 and 24.04.3 LTS. Benchmarks focused on vLLM performance for large language models in single and dual GPU configurations, compared to the Radeon PRO W7900 and NVIDIA RTX 6000 Ada. The NVIDIA RTX 4000 Ada with 20GB VRAM could not handle the tested models.