Last year, AMD launched the Radeon Pro W7900 as a 3-slot workstation GPU based on full Navi 31 (RDNA 3). Today, at Computex 2024, AMD reintroduces the Radeon Pro W7900 albeit in a more compact 2-slot form factor as the Radeon Pro W7900 Dual Slot (DS) and with a lower MSRP.
The specifications of the W7900 DS remain identical to the previous iteration. The W7900 DS features 96 compute units (CUs) or 6,144 stream processors (SPs) with boost up to 2.5 GHz and 192 AI accelerators. The card supports 48GB of 384-bit GDDR6 ECC memory, which offers a bandwidth of 864GB/s. There are a total of four DisplayPort 2.1 outputs, including a mini-DisplayPort.
Power requirements remain unchanged at 295W total card power (TBP). The Radeon Pro W7900 DS is expected to offer 61 TFLOPs of single-precision FP32 peak, which is significantly lower than the 91 TFLOPs of the Nvidia RTX 6000 Ada.
On the other hand, Nvidia doesn’t officially specify peak performance in half-precision FP16 for the RTX 6000 Ada, but AMD claims 123 TFLOPs with the Radeon Pro W7900 DS. AMD also showed a roughly 2x relative advantage in performance per dollar for the W7900 DS over the RTX 6000 Ada in enterprise workloads. The flagship Navi 31 workstation is also said to offer 38% more performance per dollar in Llama3 70 billion Q4 LLM inference compared to the RTX 6000 Ada.
The W7900 DS also gets a price cut this time around and will sell for $3,499 MSRP starting June 19, up from the $3,999 MSRP of the original, which currently retails for $3,663 on Amazon.
ROCm 6.1 for Radeon Pro
ROCm, traditionally made for AMD Instinct datacenter cards, is also officially supported on Radeon VII Radeon VII (deprecated) and RX 7900 series GPUs, such as the RX 7900 XTX on the consumer side and on professional cards such as Radeon Pro V620, W6800, W7800 and W7900.
AMD is updating the ROCm software stack for Radeon Pro to version 6.1, which now offers Ubuntu 22.04.3 (Jammy Jellyfish) support both natively and via WSL 2 on Windows. ROCm 6.1 also adds rocDecode which allows ROCm HIP post-processing of decoded compressed video streams within the GPU’s VRAM itself, instead of having to constantly ferry them across the PCIe bus.
