Big Navi increases amps in GPU memory latency test
The Chips and Cheese outlet has released a new study that compares memory latency between the latest AMD and Nvidia graphics cards. The results paint RDNA 2 with lower memory latency than Ampere. But it should be noted that lower latency may not directly correlate with actual performance.
Similar to processors, modern graphics cards exploit a tiered cache hierarchy system. Ampere sticks to a more traditional setup with an L1 and L2 cache. RDNA 2 (Big Navi), on the other hand, has the L0, L1, L2, and Infinity Cache, which basically acts as an L3 cache. Both architectures take advantage of contrasting designs, so it’s interesting to see which wins. On paper, RDNA 2 has more levels to go on the way to memory.
To measure memory latency with graphics cards, Chips and Cheese used a methodology normally applied to processors. The post used pointer lookup benchmarks in OpenCL to gauge both cache and memory latency. The socket used the Radeon RX 6800 XT and GeForce RTX 3090 as test rabbits.
In general, RDNA 2 had lower memory latency compared to Amps at all levels. The results from Chip and Cheese confirmed that the Infinity Cache only added an additional 20ns penalty on an L2 hit. The memory latency of RDNA 2 was on the same order as that of Ampere, although it had to go through more levels of cache to reach its destination.
Traveling from L1 cache to L2 cache on Amps requires more than 100 ns. In comparison, RDNA 2’s L2 cache is only 66ns from its L0 cache, regardless of having an L1 cache between them.
Chip and Cheese concluded that it takes a lot of cycles to circulate the GA102 die, given its size (628mm²). The publication believes that RDNA 2’s low-latency L2 and L3 caches give it an advantage over Amps in small workloads. This could explain why RDNA 2 graphics cards excel in gaming at lower resolutions.