AMD Ryzen 7 2700X and Ryzen 5 2600X in review

AMD's return to the CPU sector last year was reflected in a seemingly endless launch of new processors that could put real pressure on Intel in almost all areas of the desktop PC market. This forced Intel to make the most drastic adjustments to its desktop portfolio in the last decade. Nevertheless, AMD continues to manage to continuously gain market share.

To Infinity Fabric and back

AMD's first-generation processors showed higher memory latency than expected, affecting performance in memory-sensitive applications. AMD claims to have reduced memory latency by 11%, along with several significant reductions in L1, L2, and L3 cache latency for 2000 series processors. Before that, AMD's own statement as a table:

AMD's Measurement Results	L1 latency	L2 latency	L3 latency	Memory latency
Latency improvements	13%	34%	16%	11%

We start with the memory and infinity fabric subsystems and then move on to the IPC tests.

With Sandra from SiSoftware, we measured cache and memory latency with three different access patterns, which gives us more granularity than a single test. Sequential access patterns are almost completely preloaded into the TLB, so the sequential test is a good measure of prefetcher performance.

The in-page random test measures random access within the same memory page. It also measures TLB performance and represents the best random performance (this is the measurement used by providers for official datasheets). The full sample test provides a mix of TLB hits and errors with a high probability of errors, so it quantifies worst-case latency.

	L1	L2	L3	Memory
Area	2KB – 32KB	512KB – 8MB	512KB – 8MB	8MB – 1GB

Regardless of the memory access pattern, the smallest blocks of data fit into the L1 cache. And the larger the data, the larger the caches, as we described in the table above.

We tested both the Ryzen 7 1800X and the Ryzen 7 2700X on the same X470 motherboard to see the improvements first-hand. We include test results with the Ryzen 7 2700X for DDR4-2933 for the standard configuration, DDR4-3466 for overclocked configuration and DDR4-2666 for normalization with the Ryzen 7 1800X.

To put it simply, we see big improvements in memory latency across the board. With normalized DDR4-2667 memory frequencies and timings, the Ryzen 7 2700X achieves impressive gains over the Ryzen 7 1800X, regardless of data access pattern. The improvements to the 2700X are 11.49% for full random numbers, 6.64% for in-page and 9.35% for sequential access pattern.

These numbers are impressive, but the Infinity Fabric really picks up speed as we increase the 2700X's memory to DDR4-2933. The fabric connects the IMC and the cores, so we see even bigger improvements of 18% in the Full Random Test, 13.4% for a Full Random Access Pattern and 12.9% in the sequence test.

AMD has not accurately quantified the steps to improve memory latency, but we suspect that the company has significantly improved both the Infinity Fabric and the built-in storage controller itself.

% increase to 1800X	L1	L2	L3
In-Page	11.11%	51.72%	26.38%
Full-Random	11.11%	53.5%	25.64%
Sequential	11.11%	13.3%	13.3%

We've measured large cache latency shortenings that are even more than amD's forecasts, even though the company may use different access patterns for testing. In both cases, the results we have presented in the table are impressive. We also see a significant increase in cache bandwidth on a broad front. Supplying the cores with lower latency and higher throughput is a win-win for all parties involved. Intel's S-Series processors still have a great single-core L1 bandwidth advantage, but AMD's L2 cache is measurably faster in both single and multi-threaded testing. AMD even has better L2 and L3 cache latency than Intel's processors in several tests.

The complex Zen+ design merges two four-core CCX with the Infinity Fabric, a fastener that also processes IMC, Northbridge, and PCIe traffic. Therefore, the latency of this component is a critical factor in ensuring that the observed memory latency gains can actually be transferred to the cores.

SiSoftware Sandra's Processor Multi-Core Efficiency helps us demonstrate the performance of the Infinity Fabric. We use the multi-threaded metric with the setting "best pair match". The utility measures the ping times between threads to quantify the latency of the fabric in any configuration. We've lowered these metrics down to latency averages for the different paths to be more understandable.

AMD reduced the latency of the Ryzen 7 2700X by 11.8% and critical cross-CCX latency by 8.3%. We also found that the Ryzen 7 2700X has significantly improved fabric bandwidth. Intel's ringbus and mesh, on the other hand, still have the advantage of lower overall latency from core to core.

Instructions Per Clock (IPC)

Due to time constraints, we have conducted a number of IPC-oriented benchmarks. It's possible that further optimizations or a larger number of workloads could lead to different results, but we'll find out in the next few days. For the following tests, we set a static 3 GHz clock frequency.

Our single-core Cinebench test shows a 1.61% IPC improvement for the Ryzen 7 2700X, and while AMD has improved, Intel still has a distinct IPC throughput advantage. The switch to the multi-threaded Cinebench test shows an improvement of 2.6%.

Intel's Core i9 uses two 256-bit AVX FMA units per core that work in parallel, while Ryzen's Zen architecture distributes 256-bit AVX operations across two FMA units per core. This difference gives the Skylake X processor a sovereign guidance in the y-cruncher test, a single- and multi-threaded program that Pi calculates using AVX commands. We see a 3.9% increase in the 2700X's multi-threaded Y cruncher results compared to the Ryzen 7 1800X, but the gains in single-threaded AVX performance are marginal.

We also see similar results in our single-core cryptography tests, although the Ryzen 7 2700X has a big lead over the 1800X in multi-threaded AES-256 ECB encryption. The Zen architecture includes two aes cryptographic accelerators for each core, so it's not surprising that the new Ryzens dominate the Intel S series models in the AES-256-ECB tests.

Pages: