The card is built around AMD's Navi 48 GPU, the most powerful RDNA 4 GPU this generation, with 64 compute units comprising 4,096 shading units and 128 AI accelerator cores. RDNA 4 has no discrete matrix cores as AMD's server-focused CDNA chips do; matrix operations run on the same compute units via Wave Matrix Multiply-Accumulate. Even so, AMD doubled floating-point and quadrupled INT8 matrix performance moving from RDNA 3 to RDNA 4, and the card reaches 382.7 TFLOPS of dense FP8 throughput at its 2.92GHz peak clock, or double that with sparsity.

Memory is the workstation headline: the card carries 32GB of GDDR6 on a 256-bit bus. One caveat: the review's description of the chip layout, 16 2Gbit chips in clamshell mode with chips on both sides of the board, does not add up arithmetically to 32GB, and the source text is cut off mid-sentence, so the exact chip organisation is not fully evidenced. The 32GB total is the figure to rely on; the chip-level detail is best treated as unresolved, possibly a typo for 16Gbit chips.

New FP8 support also lets the card run smaller weight formats such as FP6 and FP4, which are processed at FP8 precision. That means larger parameter counts in the same VRAM, with the review noting a roughly 30B parameter model fitting on the 32GB card with room to spare.

Because AMD is not selling a reference design this generation, partners handle production, and AMD keeps them on what the review describes as a short leash. The result is a family of near-identical dual-slot blower cards that follow PCIe dimensions closely, so several can fit in one workstation chassis. ASRock, a long-time AMD partner, is among the closest partners now offered the option.