ASRock's Radeon AI Pro R9700: 382.7 FP8 TFLOPS in a 300W Blower

The ASRock Creator Radeon AI Pro R9700 delivers 382.7 TFLOPS of dense FP8 matrix throughput inside a 300-watt, dual-slot blower card. That number is the headline. It is also the only headline AMD supplies this generation, because the card carries 32GB of VRAM — 16GB less than the previous-generation Radeon Pro W7900 it nominally succeeds.

What does the card actually offer?

The R9700 is ASRock's version of AMD's top-tier RDNA 4 workstation card, built on the Navi 48 GPU. AMD is not selling a reference design this generation. Board partners carry the product instead, under tight design constraints from AMD.

Core specifications:

• 64 CUs (compute units, AMD's primary shader blocks), 4,096 shading units, 128 AI accelerator cores
• Boost clock of 2.92GHz
• 32GB GDDR6 on a 256-bit bus, 20Gbps, 640GB/s bandwidth, clamshell layout with 8 chips per side
• 64MB Infinity Cache (AMD's on-die memory cache)
• PCIe Gen5 x16, ECC under Linux only, manually enabled
• 271 x 111 x 39 mm, PCIe CEM air-cooled form factor

Peak FP8 throughput doubles with sparsity. FP32 vector throughput sits at 47.8 TFLOPS. AMD doubled floating-point performance and quadrupled INT8 performance relative to RDNA 3, which partially explains the positioning. But the register matters more than the rate of improvement.

Where does the memory question land?

Navi 48 does not use discrete matrix cores, unlike AMD's CDNA server chips. It relies on WMMA (Wave Matrix Multiply-Accumulate), which runs matrix operations on the same compute units as vector work. The result is competent, not comparable to server silicon.

FP8 support is new this generation. FP6 and FP4 execute at FP8 speed, not faster. The practical benefit is smaller weights. AMD's own example: FP6 weights fit a roughly 30-billion-parameter model into 32GB with headroom to spare. That is a real advantage. It is also a workaround for a memory ceiling AMD chose not to raise.

At AMD's Advancing AI 2025 event, ServeTheHome's Patrick Kennedy asked why the card stops at 32GB. AMD's response, per his account, treated 32GB as the maximum the memory configuration supports — which is true, and not the question. Kennedy's stated view: the market would likely absorb a higher-cost, higher-memory part between this card and the MI350P. No such product exists. AMD has not announced one.

How does ASRock build it?

The card is a conventional dual-slot blower, held to PCIe dimensional specifications so four fit in a standard ATX chassis. The open-backed blower pulls air from both sides. ASRock specifies a vapor chamber and Honeywell PTM7950 thermal interface material.

The power connector is a rear-mounted ATX 12V-2×6, rated for 600W against the card's 300W TBP (total board power). Four cards means 1,200W of GPU load, which strains a single power supply. ASRock notes nothing about PSU sizing.

Display output is unusual: four DisplayPort 2.1a ports and no HDMI. With DSC (Display Stream Compression), the card drives displays up to 12K. No other R9700 vendor makes that choice.

What is not said?

Navi 48 measures 357mm² with 53.9 billion transistors — a mid-to-high-end die. AMD shipped no true high-end GPU this generation. The R9700 therefore trades memory capacity for compute throughput against its predecessor, and AMD frames that as progress. The performance data will decide whether the trade holds. ASRock has not published pricing in the material reviewed.