Introduction

PowerColor Radeon RX 7900 GRE Hellhound is the company’s second premium custom-design implementation of the RX 7900 GRE, positioned a notch behind the company’s Red Devil flagship. The Radeon RX 7900 GRE (Golden Rabbit Edition) started out as a China-exclusive limited edition product to mark the Year of the Rabbit, but found itself growing in importance to AMD in January 2024, with NVIDIA’s brisk launches of the GeForce RTX 4070 SUPER. The Radeon RX 7800 XT beats the original RTX 4070 in raster gaming workloads, but NVIDIA plugged this gap with the RTX 4070 SUPER. The RX 7900 GRE had been around since Summer 2023, and so it was a matter of simply pricing it right for the global market, and giving it a wider launch outside China.

There is an interesting story behind the RX 7900 GRE. AMD had designed a compacted version of the Navi 31 package to drive the mobile versions of the RX 7900 series. This package is rumored to be pin-compatible with the Navi 32 package that powers the RX 7800 XT, allowing AMD’s add-in board partners to reuse their RX 7800-series PCBs with compacted Navi 31. This package only has pins for a 256-bit memory bus even though the Navi 31 chip is capable of 384-bit, and so the RX 7900 GRE gets 16 GB of GDDR6 memory across this 256-bit memory bus. With the RX 7900 GRE having a similar 260 W total board power to the RX 7800 XT and its 265 W, AMD gets to minimize the R&D costs for its partners by not only letting them reuse the RX 7800 XT PCB, but also carry over cooling solutions from their RX 7800 XT products.

The Navi 31 is a chiplet-based GPU. AMD identified components on the GPU that tangibly benefit from the switch to the newer 5 nm EUV foundry node—basically the GPU’s front-end and shader engines, or all its logic-heavy components; and clumped them into a large central chiplet called the graphics compute die (GCD). All the memory-heavy components that don’t benefit as much from 5 nm, namely the memory controllers and the Infinity Cache, have been disaggregated to smaller chiplets called the memory cache dies (MCDs), built on the older 6 nm process. Each MCD has a 16 MB segment of the Infinity Cache, and a 64-bit portion of the memory bus. There are six of these on the Navi 31 (96 MB cache, 384-bit memory bus), and four on the Navi 32 (64 MB cache, 256-bit memory bus). On the RX 7900 GRE, AMD enabled four of these MCDs.

The RX 7900 GRE has been carved out of the Navi 31 silicon by not just disabling two of its six MCDs, but also enabling just 80 out of the 96 compute units, for 5,120 stream processors, 160 AI accelerators, 80 Ray accelerators, 320 TMUs, and 160 out of the 192 available ROPs. The GPU is clocked at 1880 MHz Game clock, which PowerColor has further overclocked; while the memory, interestingly, ticks at 18 Gbps, which is slower than the 19.5 Gbps of the RX 7800 XT. At 18 Gbps, the GPU enjoys 576 GB/s of bandwidth.

The underlying graphics architecture is AMD’s latest RDNA 3, which the company engineered to take advantage of the 5 nm process. The new dual instruction issue rate compute unit comes with a 17% IPC increase over the RDNA 2 compute unit, and supports newer AI-relevant instructions. The AI accelerator prepares matrix math for execution on the stream processors, providing a large speedup in AI DNN building and training. The 2nd generation Ray accelerator comes with a 50% improvement in ray intersection performance. The new multi-draw indirect accelerator (MDIA) is an exotic new on-silicon accelerator that can significantly speed up Direct3D 12 workloads that use MDI instructions.

The PowerColor RX 7900 GRE Hellhound looks very similar to the company’s RX 7800 XT GRE, as it’s mostly reusing the board design. The large aluminium fin-stack heatsink uses a trio of LED illuminated fans, while enthusiasts can benefit from features such as dual-BIOS. The card comes with a healthy factory overclock of 2013 MHz Game clock (vs. 1880 MHz reference), while leaving the memory speed untouched. PowerColor is pricing the RX 7900 GRE at $580, a slight premium over the $550 AMD baseline.