AMD RDNA2 memory latency better than NVIDIA Ampere
AMD’s last level cache called Infinity Cache appears to show a clear advantage over NVIDIA Ampere, a work-in-progress GPU memory latency test from Chips and Cheese appears to indicate.
The tech site has developed a simple pointer chasing benchmark written in OpenCL which measures the latency in GPU memory cache. The test reveals that despite featuring a third-level cache that the competitor’s GPU lack, AMD RDNA2 based Radeon RX 6800 XT has the same level of latency as NVIDIA Ampere with pointer chasing benchmark at 1GB. Anything between 64KB and 1GB shows the advantage of AMD RDNA L3 Cache called Infinity Cache with the latency being lower.
As noted by Chips and Cheese, the latency between L1 and L2 caches on NVIDIA GA102 GPU takes over 100 ns compared to 66ns on L2 to L0 cache on AMD Navi 21 GPU. The author of the test thinks that this might explain the excellent performance of AMD RDNA2 GPUs at lower resolutions.
AMD RDNA2 vs NVIDIA Ampere memory latency, Source: Chips and Cheese
GPU manufacturers are noticing the importance of large and efficient memory caches. NVIDIA GA100 (Ampere) L2 cache has grown to 40MB which is seven times larger than GV100 (Volta). AMD pioneered the use of High Bandwidth Memory in GPU chips, now used exclusively for its Instinct series featuring CDNA architecture. At the same time, the company has developed Infinity Cache for its Radeon RX (RDNA) gaming architecture.
The site has also compared older AMD and NVIDIA architectures, where a noticeable improvement is observed for the most modern architectures.
Older GPU architectures memory latency, Source: Chips and Cheese