Threadripper Fully Unleashed
With the release of a 64-core Threadripper, AMD is not holding back anything anymore. It’s as close to a full-fledged EPYC processor as you can get on the TRX4 platform, possibly with the exception of RDIMM support. If the product makes sense for consumers or not can be discussed, but for overclockers and enthusiasts it’s an incredible display.
- AMD Threadripper 3990X (64 cores, 4.3 GHz Boost, 280W TDP) / AMD Threadripper “3970X” ES (32 cores, 4.4 GHz Boost, 280W TDP)
- ASUS ROG Zenith II Extreme (BIOS 0042 + 0802)
- G.SKILL F4-4000C18Q-32GTZKW (4x8GB)
- Custom water cooling (Rajintek CWB-TR4 RBW + Bykski B-RD360-TK60)
- Windows 10 1903, AMD Chipset Driver 126.96.36.1993, Ryzen High Performance power plan
Any benchmark results are the average of three runs. HWInfo 6.14 was used to record the monitoring information during the run. The average frequency was measured using the “Effective Clock” item. The average power was measured directly from the VRM controller (see below) and reports the CPU Core power or SOC power only. Load-Line Calibration was set to Level 5 for the overclocked results. The Prime 95 numbers were recorded after 10 minutes of stress testing without any errors.
EVC2 Power Measurement
In order to read the output power estimation directly from the VRM controller, an EVC2 was connected directly to the SMBus on the motherboard. It turns out that the ASP1405I controller can only read values up to 511.5 and any value above is clipped. Since the output current stayed below this value, the higher output power this could still be calculated using Pout = Vout*Iout.
Results and Analysis
At default settings, what stands out is the 54% higher Cinebench R20 score at maintained power consumption and temperature. The CPU frequency is only decreased to 3149 MHz compared to 3648 MHz on the 32-core version. This highlights the benefit of linear power scaling through additional cores/transistors, compared to exponential scaling from frequency and increased voltages.
When it comes to overclocking, there are a few trade-offs. When the frequency is maxed out, performance is 68% higher, but at 65% higher power consumption. The temperature remains comparable due to the large heat-spreader and chiplet-design. There’s a slight drop in achievable frequency, but keep in mind it’s a comparison of single processor samples.
In Prime 95 the highest achievable frequencies were about 300 MHz lower than the 32-core version, and power consumption was 28-35% higher. In this case we’re probably close to the package & heat-sink limit due to the longer running time.
SOC Power Consumption
Going from four chiplets to eight does increase power consumption somewhat. Since the number of memory channels and PCI-E lanes remain the same, the increase is most likely due to the additional Infinity Fabric links.