AMD RDNA3 GPU architecture first introduced chiplet small chip design. This kind of architecture in AMD Ryzen, Xiao Long processor has been playing well, to the graphics card what is different?
As with processors, AMD's graphics cards are doing so because individual chips are getting bigger and more expensive, and they have to change their thinking. In order to solve the problems of communication bandwidth and latency between multiple chips, AMD is also struggling with conscience.
AMD RDNA3 includes a GCD graphics core and up to six MCD video memory and Cache cores, among which the key Infinity Cache is not placed in the GCD, but in the MCD.
The connection between the GPU and video memory is called the Infinity Link, and its transfer rate is a whopping 9.2Gbps.
What's the concept?This is more than 10 times the bandwidth of Infinity Fabric on AMD Ryzen and Skylon processors!
In addition,The peak bandwidth between GCD and MCD is as high as 5.3TB/s.
Small chip designs inevitably add latency. Data shows that if Navi 31 runs at the same frequency as Navi 21, the latency of video memory and unlimited cache access will increase.
AMD solved this problem by increasing the core frequency by 18% and the Infinity Fabric bus base frequency by as much as 43%. The video memory access latency is basically the same, and the unlimited cache access latency is reduced by 10%!
The RDNA2 architecture is already quite frequent, and RDNA3 goes one step further, as shown by the STA(Static Timing Analysis) toolThe increase is about 30 percentBut the actual operating frequency is not that much higher.
The RX 6950 XT has 2.1GHz gaming frequency and 2.3GHz acceleration frequency, and the RX 7900 XTX has 2.3GHz and 2.5GHz acceleration frequency, respectively.
Today, of course, the nominal GPU frequency is only a reference value, and the actual operation can achieve much higher.