Single page Print

Memory subsystem performance
To get a sense of where Ryzen's dual-channel memory architecture slots into the pantheon of bandwidth, we employed AIDA64's directed memory read, write, and copy tests.

For eight hungry cores and 16 threads in a single socket, Ryzen CPUs fall decidedly mid-pack in these synthetic benchmarks. Intel sees fit to give its high-end desktop processors quad-channel memory controllers for plenty of breathing room, and the Core i7-5960X and Core i7-6950X generally enjoy much faster memory transfers as a result in these tests. Considering that the memory multiplier settings on our Gigabyte motherboard are locked out above 3200 MT/s right now, one can't just shove DDR4-3866 into an X370 mobo and get around this issue, as one can with Z270 motherboards.

We'd usually test memory latency with AIDA64 and cache latencies with SiSoft Sandra at this point, but AMD warned us that neither utility performs correctly with Zen's caches or memory controller. As a result, we're holding off on reporting those numbers through independent testing.


Source: AMD

AMD did provide reviewers with its own internal measurements of cache bandwidth and latency data for Zen. We won't be diving deep into these numbers, but it is interesting to see how Ryzen chips' cache hierarchies stack up against their Broadwell-E nemesis.

Synthetic math performance with Y-Cruncher
Normally, this spot is where we'd share the performance results from the synthetic benchmarks built into the AIDA64 utility. Unfortunately, those benchmarks haven't been updated for Ryzen as we go to press, either. Instead, we turned to Y-Cruncher, a program that calculates pi out to arbitrary billions of digits. Not only does this program require a powerful CPU to run well, it also benefits from fast memory since it's working with a large pool of RAM to store its calculations.

Y-Cruncher comes with a handful of executables that are tuned for various x86 extensions. Y-Cruncher also offers a Bulldozer-optimized binary that can use that chip family's unique SIMD instructions. We used the newest version of the executable that ran on each chip without throwing warnings about ISA compatibility. We ran the program in its multithreaded mode and chose a 2,500,000,000-digit test size.

One thing is clear from our Y-Cruncher results right away: AVX2 SIMD support seems to help a lot. Ivy Bridge, Sandy Bridge, and Bulldozer don't have it, and they suffer accordingly. The Ryzen CPUs have AVX2 support, but their 256-bit AVX throughput is half that of the Haswell and newer chips because of the 128-bit width of their FP units. Despite their high core and thread counts, the Ryzen chips land smack between Haswell and Skylake here. The Intel Extreme Edition chips put their copious memory bandwidth and execution hardware to good use by leading the pack in number crunching.