Zen 2 Instructions Per Cycle (IPC)
It's important to remember that IPC can vary by workload, so dissimilar tasks may yield different outcomes. We set a static 3 GHz clock rate and dialed memory to the respective processors' supported frequency for the following tests:
All of the processors in this test come with eight cores and sixteen threads, so we left the core allocations unaltered and the Ryzen 7 1800X serves as our baseline comparison. These apples-to-apples tests expose a significant boost moving from the first-gen to third-gen Ryzen's, with single-threaded IPC improvements ranging from 5% to 18% for most applications. AVX-heavy workloads receive the biggest benefit with the AVX y-cruncher benchmark exposing a 70% and 163% increase in single- and multi-threaded workloads, respectively. That comes as a side benefit of AMD's move to a 256-bit floating point unit instead of splitting AVX workloads into two instructions fed across two cycles. Performance improvements are more muted with cryptographics workloads, which only experience a ~5% uplift.
AMD used SPECviewperf 13 to measure its IPC improvements, and it's clear it's results are more than plausible, if not a conservative estimate. Removing our results from y-cruncher, which is clearly a standout, and the crytpographic workloads yields an average 13% IPC uplift, which isn't too far off from AMD's official claim of 15%. Throwing in the y-cruncher and cryptographic tests yields a whopping 18% in IPC uplift.
Power consumption measurements are always a bit tricky. But as long as your 12V supply (EPS) readings, motherboard power supply sensor values, and voltage transformer losses plausibly coincide, everything is fine. Therefore, we're using pure package power to avoid possible influences from our motherboard. Results from the PWM controller are very reliable if you take them as averages over a few minutes.
We first measured power consumption during an AIDA stress test, but we disabled AVX instructions. The Ryzen 7 3700X beats the pack with 90W of package power consumption under full load, trailed slightly but the Core i7-9700K. The Ryzen 9 3900X pulls 142W at stock settings, and that increases to 168W when we engage the new "Auto OC" and Precision Boost Overdrive features. That is 29W more than the stock Core i9-9900K during this test, but the 3900X comes with four more cores, so performance efficiency will be the real judge of power consumption.
We're moving away from using AVX-based stress tests for our CPU power testing, though we will continue to use them for their intended purpose of validating overclocks. AVX-based stress testing utilities essentially act as a power virus that fully saturates the processor in a way that it will rarely, if ever, be used by a real application. Those utilities are useful for testing power delivery subsystems on motherboards, or to generate intense thermal loads for case testing, but they don't provide a performance measurement that can be used to quantify efficiency.
The y-cruncher benchmark computes pi using a heavy multi-threaded AVX workload and also generates a performance measurement that we can use for efficiency metrics. We're also adding in HandBrake in x264 and x265 flavors. The latter uses a heavier distribution of AVX instructions than the former, but both transcoders are great for stressing the processor with a real-world workload. As we can see, the Ryzen 7 3700X sucks the least power of the group during the tests, but we should also bear in mind that it also has the lowest TDP rating.
As a sidenote, we tested Intel's Core i9-7920X as part of our test pool, but removed the power results due to incorrect reporting from the motherboard's sensor loop.
Combining the performance metrics recorded from the three applications, we can see that the Ryzen 7 3700X is an incredibly efficient processor given its level of performance, and the beefy Ryzen 9 3900X also impresses, though you could argue the overclocked configuration consumes a lot more power in exchange for a relatively tiny gain in performance. In either case, the 7nm process obviously confers a solid power-to-performance ratio.