Radeon VII – Paper Launch and Thoughts
On February 7, 2019, the Radeon VII GPU was allegedly released. Many reviews with differing opinions yet mostly consistent data hit the web. Attempting to purchase this card within the United States was a non-starter; AMD’s site was sold out within minutes and none of the other major e-tailers had any stock available. While a number of manufactures such as AsRock, MSI, XFX, and others announced an offering, the associated marketing materials confirm that they’re effectively reselling the reference design with the “value add” of alternative packaging and some stickers on the fans.
At $699, the defining feature of this solution involves the most up-to-date manufacturing process and memory allocations that are better suited for professional workloads. While the readily-available GeForce RTX series has plenty of offerings with custom cooling solutions, better-than-Founder’s Edition performance, and next-generation technologies (ray tracing, DLSS), the Radeon VII didn’t launch with equivalent commitment. Is it a rushed product that effectively shoehorns the Radeon Instinct MI50 components into a prosumer-usable package? Based on availability, the answer is a resounding yes.
The thermal aspects of this “RTX 2080” competitor (in rasterization and frame rate only) provided by Steve Burke’s analysis over at Gamers Nexus demonstrate that advanced fabrication processes can only take technology so far if the core fundamentals related to optimizing performance aren’t achieved. The GCN architecture as a whole is getting long in the tooth. While it’s still sufficient for modern workloads such as gaming, content creation, and 3D rendering, consumers of these solutions have had to deal with ever-increasing heat generation and ever-shrinking performance gains with each successive iteration. The Radeon VII and RX 590 both lend service to this fact.
The hype train is slowly building for Navi, yet there are some in the media that aren’t aware that the inherent limits of GCN will still be retained for the new GPUs. If the move from 14nm to 7nm for the Vega architecture wasn’t enough to provide consistently superior performance against the GeForce 1080 Ti and GeForce RTX 2080, the potential performance envelope for a Polaris-class GPU with a die shrink and GDDR6 memory will most likely meet the same fate against the competition.
The net-new GPU hotness (pun intended) from AMD won’t release until 2021. Arcturus, the code name for this effort, will be the first product that moves on from GCN. The following article from Anandtech (circa-2011) demonstrates AMD’s transition from pre-GCN VLIW designs to what we’ve effectively been using for a very long time now. We’re pessimistic in our belief that Navi will not be sufficient for maintaining competitiveness with nVidia. With a renewed secondary front from Intel in the graphics arena, AMD can’t afford to have Arcturus flop.
A salvage of the 7nm Vega architecture for gaming purposes would most likely require additional capital that doesn’t make a lot of sense to spend at this point. More efficient implementations of Vega as part of the Ryzen G APUs demonstrate that acceptable performance and low power consumption are indeed possible. The commitment to HBM2 via silicon interposer has created additional challenges from the thermal perspective. The die consistency noted in a number of reviews is a positive development which is immediately offset by the poor contact that the reference cooling solution establishes with the die. The improvements identified at Gamers Nexus when the stock thermal pad is replaced with an application of quality thermal interface material show that the potential for more headroom exists.
As the Vega64 can be had in reference designs at $399, we don’t expect an uptake in Radeon VII adoption outside of media and prosumer use cases. Based on the lack of custom PCBs and cooling solutions from AMD partners, it’s not outside of the realm of possibility that these partners came to the same conclusion.