AMD made several GPU-themed announcements at its Analyst Day today. While we’ll have coverage of the wider event, we wanted to focus specifically on the GPU announcements the company has made. AMD announced a pair of new architectures: CDNA and RDNA.
RDNA2: More Than Big Navi
AMD didn’t say a ton about RDNA2, but they did make some announcements about the kind of improvements users should expect. When AMD launched Navi 10, there were some discontented mutterings about the fact that the new 7nm GPUswere only on-par with Nvidia’s 12nm parts, as far as performance-per-watt was concerned. While this represented a very real improvement over GCN, AMD hadn’t delivered enough improvement to close the gap in a single bound.
According to the company, RDNA2 will deliver a substantial additional improvement in performance per watt.
It’s always wise to take claims like this with a grain of salt, but fortunately, we can look back to our original 5700 and 5700 XT review to see how Navi compared with GCN. The Radeon 5700 offers nearly identical performance to the Vega 64, but it draws substantially less power:
If you divide the Vega 64’s measured 347W by 1.5x, you’d expect the RDNA-based 5700 to draw 231W, compared with the 256W we observed. Our observed improvement, therefore, is somewhat smaller than AMD’s assumption, with a 1.36x improvement compared with a claimed 1.5x.
Keep in mind, however, that AMD undoubtedly measures performance-per-watt in a suite of titles, while we gather data in just one game / resolution. Yes, companies are going to pick data that puts their best foot forward, but our single data point is meant to be representative, not definitive. Regardless, AMD is claiming it can deliver the equivalent benefit of a full node shrink with further architectural improvements. That may be reasonable, given that the company didn’t have nearly as much time to develop RDNA after GCN as it had to develop Ryzen following Piledriver. AMD’s own timeline explicitly recalls that comparison:
A further 1.5x – 1.36x improvement in performance-per-watt would give AMD a decisive advantage versus Turing. We have absolutely no information on how it would compare against Ampere, and therefore won’t speculate. Even if Nvidia retained overall performance-per-watt leadership, however, delivering two generational improvements of 1.5x (or even 1.36x) in a row would indicate AMD was taking the issue far more seriously than it has in the past. GCN became modestly more power efficient over its long life, but AMD rode the ragged edge of its TDP ratings to maximize performance against Nvidia from virtually the moment the architecture launched in 2012 to its final dGPU iteration as the Radeon VII in early 2019.
The bottom line is this: Best-case, AMD will be on much stronger competitive footing with Nvidia as far as performance-per-watt. Worst-case — assuming AMD delivers on these figures but Nvidia gains more from the 7nm shrink — is that AMD will still have a solid story to tell about this aspect of its business.
AMD gave no information on segmentation for RDNA2 but emphasized that it was an enthusiast-class architecture. It’s obviously intended to take the fight to Nvidia’s upper-tier products and AMD has stated it won’t be calling the chip “big Navi,” because the level of advances from RDNA to RDNA2 deserve a better name. It’s been quite some time since AMD sunk much effort into quickly iterating on a GPU architecture, but there was a time when the company was known for this kind of rapid-fire iteration. The HD 2000 series hit market (or, more accurately, splatted into it) in 2007, to absolutely no one’s delight. AMD rapidly evolved the dead-letter HD 2000 family into the more-respectable HD 3000 family by late-year, launched the competitive HD 4000 family by mid-2008, and had the first DX11 GPUs in-market by September, 2009.
A fairly rapid iteration between RDNA (July 2019) and RDNA2 (?, 2020), in other words, would be more the historical exception rather than the rule.
CDNA: Data Center Driven
Although AMD didn’t explicitly make this connection during the event, the emergence of CDNA as a separate architecture may explain why we’ve seen no formal ROCm support for Navi 10 under Linux. (ROCm is AMD’s open source GPGPU computing platform that translates CUDA into code AMD GPUs can run).
AMD’s compute-centric roadmap starts with GCN in 2019 (Radeon Instinct MI50 and MI60), progresses through CDNA, and arrives at “CDNA2” in 2022. CDNA2, therefore, would be the architecture that supports the El Capitan supercomputer. CDNA is a compute-centric version of RDNA, but AMD didn’t give specifics about the changes between the two families beyond that statement. Sometimes it’s possible to draw conclusions about what one company will do by examining the behavior of a competitor, but Nvidia has actually used several different strategies for its high-end Tesla GPUs compared with its consumer cards. There have been times when Team Green deployed significantly different architectures for Tesla compared with GeForce and times when the company tapped the same core design for parts in both families.
Click to enlarge. This slide is cropped differently to maximize visibility of the small text elements. Gray = Unsupported, Dark blue = early, and bright blue = full production.
One major feature of CDNA? Fully connected, cache-coherent architecture between CPU and GPU, which CDNA will introduce. AMD did not specifically say if CDNA corresponds exactly to RDNA, while CDNA2 corresponds to RDNA2. In theory, CDNA might have a similar relationship to RDNA that Nvidia’s old GK104 had to GK110. Both chips implemented the Kepler architecture, but GK110 supported several features that GK104 didn’t, in addition to packing more GPU cores and supporting faster double-precision floating-point performance.