A look at the AMD Ryzen 9 3950X

#1 · 2019-11-15, 14:02

Quote from elmor on 2019-11-15, 14:02

Since it was first discovered the latest product line of AMD mainstream desktop processors would be configured with up to two CCD (Core Complex Dies) there has been anticipation for a 16-core model. With the release of the AMD Ryzen 9 3950X, the highest available core count in this segment has doubled in the past six months. In this piece we'll take a brief look at how it does in terms of performance, power, temperature and boost clocks.

Processor specifications

The AMD Ryzen 9 3950X is not a revolutionary processor by any means. What I mean by that is the changes are very much incremental compared to the already released Ryzen 9 3900X. It's still the same two CCD configuration, but with all available cores enabled. The advertised peak boost clock is increased from 4.6 GHz to 4.7 GHz. It can be argued that this is impressive by itself while keeping the 105W TDP spec. Realistically this is achieved by lower frequencies when all cores are used, possibly with the help of binned dies for better power efficiency.

Effective Clock

In this test I'll be using a new method for measuring boost frequencies, added in HWInfo 6.14 as "Effective Clock". It better measures boost frequencies through a per-thread CPU counter (APERF MSR). This counter increases with each clock cycle the CPU core is executing, meaning it does not increment when the clock is gated (turned off) and the core is idling in lower power states (C-states). If the CPU core clock is 4.0 GHz and is constantly executing the counter will measure 4.0 GHz. When executing half of the time, it will measure 2.0 GHz. If the core is clocked at 4.0 GHz half the time and 2.0 GHz the other half, it will measure 3.0 GHz. This allows for measuring a precise average frequency over a period of time to see how close to the advertised speeds you are in your specific workload.

HWInfo 6.14 Effective Clock

Test setup

AMD Ryzen 9 3950X

ROG Crosshair VIII Formula (BIOS 1005)

2x8GB Gskill FlareX 3200 CL14 @ 3200 CL14

Generic 120 mm tower air cooler

Windows 10 1903 + AMD Chipset Driver 1.9.27.1033

Test method

The benchmark was repeated three times and the presented values are the calculated average. In the single threaded tests, the thread affinity for the benchmark was manually set to the highest ranked core. HWInfo 6.14 was used to record the monitoring information during the run. The average frequency was measured using the "Effective Clock" item. The max frequency reading was measured using the "Core Clock" item. The average power was measured directly from the VRM controller and reports the CPU Core power only (excluding CPU SOC). The tests were repeated first at default settings and then overclocked to 4.35 GHz. Note that these are synthetic tests primarily to investigate clock frequencies, power consumption and temperatures.

Results and discussion

Cinebench R15 - Single Thread

The single threaded score is just as expected from the modest frequency bump, the results are just in line with 3900X or slightly above. By looking at the max and average frequencies we see that the average core frequency during this test is 4571 MHz and the max clock is recorded as 4625 MHz. Note that while temperatures seemed "fine" at just above 60 °C, this is using a tower air cooler and not high-end air or water cooling. As expected when overclocking using a fixed core frequency, single threaded performance suffers, though not by a significant amount. I'd speculate parts of why the difference is not larger is due to removed power limits and possibly disabled C-states.

Cinebench R15 - Multi Thread

Moving over to multi threaded tests shows that performance scales very well in this benchmark with the added cores. Over 3900 points is just insane on a stock mainstream desktop platform. It's even more impressive when considering the average CPU core power consumption is only 103 W and 3.8 GHz during an all core load with 16 cores is very respectable. The benefits of a CCD design and TSMC's 7nm process is really showing here. However when overclocking the CPU, the power consumption starts to rapidly increase. A 13% increase in the benchmark score requires a 76% increase in power consumption and temperatures start to get toasty.

Cinebench R20 - Multi Thread

Re-running the multi-threaded tests on the latest version of the benchmark tells very much the same story. Due to the heavier benchmark the average all core frequency drops slightly to 3762 MHz at default settings and the highest recorded temperature raises by 6 °C. The average power consumption stays the same which just means the stock power limit works as expected. When looking at the overclocked numbers we see even higher power consumption and temperatures. At 208 W the air tower cooler is really struggling to keep up and temperature reaches up to 94 °C.

Ryzen 9 3950X Cinebench R15 - Single benchmark

Since it was first discovered the latest product line of AMD mainstream desktop processors would be configured with up to two CCD (Core Complex Dies) there has been anticipation for a 16-core model. With the release of the AMD Ryzen 9 3950X, the highest available core count in this segment has doubled in the past six months. In this piece we'll take a brief look at how it does in terms of performance, power, temperature and boost clocks.

Processor specifications

The AMD Ryzen 9 3950X is not a revolutionary processor by any means. What I mean by that is the changes are very much incremental compared to the already released Ryzen 9 3900X. It's still the same two CCD configuration, but with all available cores enabled. The advertised peak boost clock is increased from 4.6 GHz to 4.7 GHz. It can be argued that this is impressive by itself while keeping the 105W TDP spec. Realistically this is achieved by lower frequencies when all cores are used, possibly with the help of binned dies for better power efficiency.

Effective Clock

In this test I'll be using a new method for measuring boost frequencies, added in HWInfo 6.14 as "Effective Clock". It better measures boost frequencies through a per-thread CPU counter (APERF MSR). This counter increases with each clock cycle the CPU core is executing, meaning it does not increment when the clock is gated (turned off) and the core is idling in lower power states (C-states). If the CPU core clock is 4.0 GHz and is constantly executing the counter will measure 4.0 GHz. When executing half of the time, it will measure 2.0 GHz. If the core is clocked at 4.0 GHz half the time and 2.0 GHz the other half, it will measure 3.0 GHz. This allows for measuring a precise average frequency over a period of time to see how close to the advertised speeds you are in your specific workload.

Test setup

AMD Ryzen 9 3950X
ROG Crosshair VIII Formula (BIOS 1005)
2x8GB Gskill FlareX 3200 CL14 @ 3200 CL14
Generic 120 mm tower air cooler
Windows 10 1903 + AMD Chipset Driver 1.9.27.1033

Test method

The benchmark was repeated three times and the presented values are the calculated average. In the single threaded tests, the thread affinity for the benchmark was manually set to the highest ranked core. HWInfo 6.14 was used to record the monitoring information during the run. The average frequency was measured using the "Effective Clock" item. The max frequency reading was measured using the "Core Clock" item. The average power was measured directly from the VRM controller and reports the CPU Core power only (excluding CPU SOC). The tests were repeated first at default settings and then overclocked to 4.35 GHz. Note that these are synthetic tests primarily to investigate clock frequencies, power consumption and temperatures.

Results and discussion

Cinebench R15 - Single Thread

The single threaded score is just as expected from the modest frequency bump, the results are just in line with 3900X or slightly above. By looking at the max and average frequencies we see that the average core frequency during this test is 4571 MHz and the max clock is recorded as 4625 MHz. Note that while temperatures seemed "fine" at just above 60 °C, this is using a tower air cooler and not high-end air or water cooling. As expected when overclocking using a fixed core frequency, single threaded performance suffers, though not by a significant amount. I'd speculate parts of why the difference is not larger is due to removed power limits and possibly disabled C-states.

Cinebench R15 - Multi Thread

Ryzen 9 3950X Cinebench R15 - Multi benchmark

Moving over to multi threaded tests shows that performance scales very well in this benchmark with the added cores. Over 3900 points is just insane on a stock mainstream desktop platform. It's even more impressive when considering the average CPU core power consumption is only 103 W and 3.8 GHz during an all core load with 16 cores is very respectable. The benefits of a CCD design and TSMC's 7nm process is really showing here. However when overclocking the CPU, the power consumption starts to rapidly increase. A 13% increase in the benchmark score requires a 76% increase in power consumption and temperatures start to get toasty.

Cinebench R20 - Multi Thread

Ryzen 9 3950X Cinebench R20 - Multi benchmark

Re-running the multi-threaded tests on the latest version of the benchmark tells very much the same story. Due to the heavier benchmark the average all core frequency drops slightly to 3762 MHz at default settings and the highest recorded temperature raises by 6 °C. The average power consumption stays the same which just means the stock power limit works as expected. When looking at the overclocked numbers we see even higher power consumption and temperatures. At 208 W the air tower cooler is really struggling to keep up and temperature reaches up to 94 °C.