AMD's 3D V-Cache technology uses blocks of SRAM stacked on top of the CPU logic die, where the CPU cores reside. With this, it allows the processor to access enormous cache reserves for applications, in order to load them much faster without having to send instructions to RAM. If it were a storage device, AMD's 3D V-Cache would have read speeds of 182 GB/s and writing speeds of 175 GB/s.
Know better how a cpu works with AMD 3D V-Cache
Through an experiment, it is possible to use this additional level 3 cache, which we know as L3M cache; like a RAM disk. In that experiment, the L3 SRAM behaves similarly to a storage drive.
AMD 3D V-Cache read and write speeds have been benchmarked
Big disclaimer here, says Tech Power Up; is that this is only possible by exposing the L3 to the CrystalDiskMark benchmark, and that there are no real-world applications that can do this in a way that CrystalDiskMark, a tool that we have used in storage analysis. According to Twitter user GPUsAreMagic, the steps to replicate this procedure consist of getting an AMD Ryzen CPU with 3D V-Cache, installing OSFMount and creating a FAT32 formatted RAM disk, and running the CrystalDiskMark benchmark. The values should be set to SEQ 256 KB, Queue Depth 1, Threads 16, and data fill to 0s instead of “random«.
The results have surprised because, although the size of the SRAM L3 is tiny compared to the weight of current applications and games, it is very fast and accessible to the CPU. This is what makes compensate you by loading data locally before moving the data to system RAM.
find out about how does RAM work with our article
With a pilot plant AMD Ryzen 7 5800X3D, L3 cache speeds as a RAM disk exceed 182 GB/s in reading and 175 GB/s in writing. In another test, shared by Albert Thomas, when converting the cache into a RAM disk of the AMD Ryzen 7800X3D, which scores a little less with more than 178 GB/s read speed and more than 163 GB/s write speed. The CrystalDiskMark benchmark only performed these tests on small allocations ranging between 16 MiB and 32 MiB, which means that there are no real-world workloads that can use it yet, and it doesn't seem like it's going to be possible. But we must understand that they are designed for small and repetitive processes such as the commands of a video game, which are where an improvement in performance has been seen.
