AMD fans have been curious about Big Navi since the company teased its arrival last year, but AMD hasn’t officially said much about the GPU since, beyond confirming that it’ll launch this year and that the same technology is used in the PS5 and Xbox Series X. Now, a Twitter user is claiming to have documents from SK Hynix that give some of the official stats on the upcoming graphics card.
— CyberPunkCat (@CyberCatPunk) February 24, 2020
The first question to ask when evaluating any rumor is how much sense the chip makes in context. From that perspective, this is a pretty reasonable rumor. A straight doubling of Navi’s width makes sense for a larger card, while the 4,096-bit memory bus reportedly delivers up to 2TB/s of memory bandwidth. That’s double Radeon VII, for a GPU that would also pack a 1.5x larger frame buffer. Hat-tip to Hot Hardware for seeing the story.
To be clear, there’s going to be zero near-term need for GPUs with 24GB of dedicated RAM or 2TB/s of memory bandwidth for mainstream gaming. While both of these specs are impressive and we are on the edge of a new console generation, consoles like the PS5 and Xbox Series X are expected to offer 16-24GB of RAM in total, with some percentage of it reserved for the host operating system. PC game developers are going to target their software at the hardware people actually own, and the most popular GPUs currently in-market are all in the 4GB to 8GB range. The one metric that does reliably increase every console generation is total game storage requirements, which tend to jump at the introduction of each new generation. PC gamers worried about storage might want to look to their SSDs, but VRAM needs shouldn’t surge dramatically.
The bandwidth improvement would be enormous. Compared with the GCN-based Radeon VII, the Navi 2.0 RDNA GPU would be 1.25x larger (5,120 cores versus 4,096) with 2x the memory bandwidth and an assumed equivalent 1.25x expected uplift courtesy of the new RDNA architecture. We can’t speculate on fill rate or texture rates without knowing the GPU clock, but the increase in texture units and ROPs would provide significant uplift compared with the Radeon VII as well. A three-high HBM2 stack at 2GB/layer would work out to 6GB per stack, for 24GB total in four stacks.
Mathematically, everything makes sense. A Navi built around specs like these would easily take the fight to Nvidia’s Turing family, given that we’re looking at significant improvements in width and architecture. Any additional gains AMD picked up for RDNA 2.0 over 1.0 would also factor in, though we don’t actually have many details on how different the two architectures are, beyond the fact that RDNA 2.0 supports ray tracing and variable rate shading. Power efficiency is also a question — RDNA 1.0 moved the ball downfield, with AMD’s 7nm competing with Nvidia’s 12nm, but it’s not clear if RDNA 2.0 will improve things further. Nvidia will also be releasing Ampere this year, so Big Navi doesn’t just need to tackle Turing — it needs to have enough gas in the tank to be well-positioned against future Nvidia products as well.
I’m not going to say this rumor is accurate — we genuinely don’t know — but the GPU it describes at least makes sense relative to what AMD might be releasing, and we’d expect strong performance scaling compared with the 5700 XT if this is indeed the GPU AMD intends to launch.