In the early hours of March 6, Beijing time, AMD CEO Su Zifeng announced at the AMD Financial Analyst conference that the Radeon RX GPU with the next-generation RDNA2 architecture will be 50 percent more energy efficient than the existing RDNA GPU. AMD will focus on optimization at the microarchitecture level, and with changes from Zen to Zen 2, RDNA 2 will improve in terms of IPC, circuit design, and clock frequency.
Simply put, the RDNA2 architecture has three key points to improve energy efficiency than:
1, redesigned GPU microframe: by increasing logic functions, reduce design complexity and switching power, while improving the operating frequency, thus achieving the Improvement of IPC performance.
2, the new generation of 7nm plus EUV process: RNDA2 GPU will be converted from 7nm process to 7nm plus EUV process, the new process will further enhance the density of transistors, in a smaller package area to achieve higher performance, but also at the same power consumption to provide higher performance.
3, variable rate coloring VRS technology and hardware ray-tracing features added: in fact, these two functions in the NVIDIA Turing architecture has been implemented, which is also turing GPU energy efficiency than one of the important reasons.
VRS Variable simply put sgares the GPU’s focus on rendering what the user can see, reducing the rendering accuracy of the edge area appropriately, to improve the efficiency of the GPU. According to AMD, variable rate coloring VRS technology can increase the IPC of the RDNA2 architecture by 20%.
PS: “Variable Rate Shading” technology is called “Adaptive Rendering Technology” on the N card, which is essentially similar, except that it is called differently.
As for the next generation of RDNA3 architectures, whose core code name is NAVI 3X, AMD says it will use more advanced process processes. If the RDNA2 structure is 7nm plus EUV, RDNA3 should be 5nm and will be on the market by 2022.
AMD has never increased GPU energy efficiency by 50% in the past decade with little change in process processes, and the current RDNA architecture has been comparable to or even slightly surpassing NVIDIA’s Turing architecture. And the next generation of RDNA2 architecture will no doubt be completely ahead of NVIDIA. As for whether we can defeat NVIDIA’s next-generation ampere architecture, time will give us the answer!
In addition to game cards, AMD has also disclosed some professional card-related information. RDNA architecture uses more transistors on every flow processor to improve IPC performance. But for professional cards, the emphasis is on floating-point performance, so the older generation of GCN architecture is actually more suitable for professional computing.
That is to say, in the future the professional card will be separated from the game card, RDNA architecture for game cards, and then the professional card will be using CDNA architecture (gcN architecture upgrade).
The new CDNA architecture will support AMD’s future Infinity connected architecture, an evolutionary version of AMD’s Infinity Fabric bus they are currently using in the CPU, which is not just a bus used to interconnect inside the CPU, but also to connect GPUs.
Infinity Architecture is AMD’s interconnect solution to heterogeneous data consistency, which can improve the interconnection bandwidth, reduce heterogeneous communication delay, and so on, while also solving the problem of heterogeneous data inconsistencies and reducing programming difficulty.
The first generation of 7nm process professional card Radeon Pro Vega is based on the GCN architecture, with 4096 flow processors; Most likely the upcoming Radeon Instinct Mi100, with 8,192 stream processors and 32GB of HBM2e memory.
CDNA3 will use Infinity Fabric 3.0, while the process will be upgraded to 5nm and will be released by 2022, tentatively set as the Radeon Instinct Mi150. This is a bit like the Radeon Instinct GPU used in the 20 billion floating-point operation of El Capitan Supercomputing.