XFX confirms AMD’s Radeon AI Pro R9700 GPU will only launch in pre-built PCs

It is scheduled to come to the DIY market sometime later, but if you want it now, you will have to settle for a complete workstation build.

XFX has announced that its professional AMD Radeon AI Pro R9700 graphics cards will soon launch as part of complete workstations from select system builders. These will be powered by AMD’s latest Ryzen Threadripper Pro processors, offering up to 96 cores, 2TB of DDR5 memory, and multi-GPU configurations.

Showcased during Computex, the Radeon AI Pro R9700 is AMD’s latest offering to professionals working with mid-to-large AI models or VRAM-heavy computation. Based on the RDNA 4 Navi 48 chip, it succeeds the RDNA 3 Radeon Pro W7700, unlocking more potential thanks to a doubling in memory capacity. This allows it to run large models entirely in VRAM without falling back on system RAM, enabling up to 5x the performance of 16GB-class GPUs.

The Radeon AI Pro R9700 officially released on July 23, 2025, but we were wondering about its absence in retail channels. Turns out, it will only be available through AMD’s add-in board partners, according to a new announcement from XFX.

XFX didn’t take this decision of its own accord however; it’s a choice made by AMD. For this generation, Team Red preferred to focus on the professionals who may need big setups, such as quad-GPU Threadripper systems. That said, AMD has also promised to release standalone Radeon AI Pro R9700 graphics cards later in Q3 2025, so users awaiting it for their AI and computational experiments will need to be patient.

As a recap, the Radeon AI Pro R9700 features 64 RDNA 4 compute units, 128 AI accelerators, and 32GB of GDDR6 memory linked via a 256-bit bus. This results in up to 191 TFLOPs of FP16 Dense and 1,531 TOPS of INT4 Sparse at 300W TDP. The card takes only two slots and uses a blower-type cooler. Furthermore, it can be powered by a single 12V-2×6 cable.

AMD claims that the R9700’s 32GB memory pool allows it to reach much higher performance levels in large AI models compared to Nvidia’s RTX 5080. To put this in numbers, AMD’s testing shows the R9700 outperforming the RTX 5080 by 361% in Phi 3.5 MoE Q4, 437% in Mistral Small 3.1 24B Instruct 2503 Q8, and 496% in Qwen 3 32b Q6 Large Prompt (3000+ tokens). Not bad for a card that takes less space and power.

Question is, how much will it cost when finally available to Joe Public?

Fahd Temsamani
Fahd Temsamani
Senior Writer at Club386, his love for computers began with an IBM running MS-DOS, and he’s been pushing the limits of technology ever since. Known for his overclocking prowess, Fahd once unlocked an extra 1.1GHz from a humble Pentium E5300 - a feat that cemented his reputation as a master tinkerer. Fluent in English, Arabic, and French, his motto when building a new rig is ‘il ne faut rien laisser au hasard.’
SourceXFX

Deal of the Day

Hot Reviews

Preferred Partners

Related Reading