UPDATE:
Short version:
Not all multi-threaded workloads are created equal so unless you got a water loop don't even bother with static OC and just run 3970X on Auto with power-limit enforced and don't go anything beyond +200MHz PBO.
Longer version:
A note regarding overclocking. I was able to find a real-world multi-threaded workload that pushes this CPU to the limit and doesn't play well with static overclock because power limit doesn't work with static overclock. Under my static overclock of 4.15GHz at 1.200Vcore synthetic CPU stressing workloads only result in power consumption of 260-270Wpackage but my real workload pushed the power consumption to 330+W when all 64 threads were loaded and CPU utilization was 99ish%. This memory and CPU intensive workload takes hours and actually crashed in the middle until I figured out that it was a thermal problem. The first thing I did was drop clocks to 4.0GHz and kept 1.200Vcore and it still crashed but it took longer to crash. Then I looked with hardware monitoring software and saw power over 300Wpackage. I tried synthetic CPU stressing workloads again for comparison and power draw was several tens of watts lower. I tried to force power limit in BIOS but it doesn't apply for static OC. I went to Auto settings in BIOS and haven't looked back. My 64-thread workload that broke static OC runs at 3.7GHz base on all cores with some cores boosting 3.725-3.750GHz on Auto while synthetic CPU stressing software runs 3.9-4.0GHz on Auto with power limit at 280W on Auto settings. So read the rest of the original review but ignore static OC section for CPU. The RAM OC still applies and boosts performance. I also ended up upgrading to NEMIX dual-rank 4x32GB DDR4-3200 22-22-22-52-1T 1.2V ECC UDIMMs that I recommended in original review that I run at DDR4-3600 1.35V 24-24-24-56-1T because latest version of my workload used 90+GB and caused SSD memory swapping and thrashing. I think I lost around 15-20% bandwidth/latency vs my tightly tuned dual-rank Samsung B-dies but if I were going to double those up from 64GB to 128GB there is no doubt I would have lost some of my tight timings 16-17-16-36-1T anyway (probably to 18-18-18-38 with 8x16GB dual-rank, based on experience) and therefore would have lost bandwidth/latency anyway. I went with better reliability as my run times are long.
ORIGINAL:
3970X is basically 2x tightly-coupled 3950Xs in one package from CPU cores configuration perspective but being an EPYC Rome derivative (has a wider IO chiplet) also comes with 4x the PCIe 4.0 lanes of 3950X.
I'm cooling mine using $90 be quiet! 250W TDP Dark Rock Pro 4 CPU air-cooler (a necessary monster in this case) and paired with $400 MSI TRX40 WIFI board and 4x DIMMs of G.SKILL Flare X 16GB (dual-rank) binned Samsung B-die DDR4-3200 14-14-14-34-1T 1.35V RAM for a total of 64GB RAM (expandable to 8x DIMMs if I need 128GB of fast RAM - though will likely need to re-tune memory OC with 2 DIMMs/channel). If you need more than 64-128GB RAM I recommend saving the hassles of memory testing/validation and just going with 32GB DDR4-3200 ECC UDIMMs (not RDIMMs, which are not supported) from NEMIX for up to a maximum of 256GB with 8x DIMMs. One note on Mclk and Uclk is that Uclk (uncore clock - L3 caches and infinity fabric) runs at Mclk (memory clock) speeds up to 1800MHz corresponding to DDR4-3600 so that becomes your memory tuning sweet spot for "free" performance gains on these Ryzen CPUs. Oh and don't forget to update your motherboard BIOS to the latest version before optimizing anything as results/stability may vary across different BIOS versions!
Please note that each CPU sample is a little different and your results may vary but it shouldn't deviated too much from my findings below if you are doing it right. Use my optimizations below as a starting point and you may do a little better or worse depending on your sample and your cooling setup. I played around with OC using both PBO method (+100MHz, +200MHz) and static all-core OC method and settled with a static all-core OC, let me explain why. PBO method gives you slightly higher low-thread boost clocks (maybe 3-4% higher, not more) but at the cost of higher automatic Vcore of 1.4+V at idle and low-thread workloads. The PBO method downside comes when you run all-core workloads and Vcore droops down from 1.4+V to 1.3+V with all-core clocks only peaking to 3.9-4GHz depending on a core and CPU package power consumption shoots up to 330W. You may be able to play with Load Line Compensation (LLC) settings to bring down the all-core workload Vcore droop even lower to lower the package power consumption but in my case I haven't seen cores boost above 4GHz on all-core workloads no matter what I did with PBO and LLC and package power consumption still peaked at 300+W so it must be a thermally-sensed sustained limit. With static OC I was able to find a sweet spot where I wanted it to keep package power consumption under all-core workloads closer to the rated 280W with clocks/Vcore optimizations. I was able to do 44x multiplier (44x 100MHz Bclk) for 4.4GHz@1.300Vcore with level 2 on LLC resulting in 1.300Vcore@330Wpackage on all-core workloads but package temperatures peaked just above 80C. Clearly 250W-rated CPU air-cooler wasn't going to work well for 300+W. I should note that these new CPUs do power management really well to internally clock/power-gate as needed based on resource utilization so there's no problem of drawing too much power at idle with a static OC like in the chips of the past. My final balanced CPU static OC is 41.5x multiplier (41.5x 100MHz Bclk) for 4.15GHz@1.200Vcore with level 2 on LLC resulting in 1.150Vcore@260Wpackage@71Cpackage on all-core workloads. Actually, I was stable at 42x multiplier for 4.2GHz@1.25Vcore which with level 2 LLC resulted in 1.20Vcore@280Wpackage during all-core workloads but I chose to save 20W instead of pushing extra 50MHz.
For RAM OC I settled on DDR4-3600 with 1:1 Mclk:Uclk with 16-17-16-36-1T at 1.45Vmem. Going above 1800MHz on Uclk (uncore) switched Mclk:Uclk to 2:1 ratio despite BIOS setting explicitly set to 1:1 so I didn't bother further and left Vsoc on Auto. Now for the importance of dual-rank RAM, I also have 8GB (single-rank) G.SKILL DDR4-3200 14-14-14-34-1T RAM kits with binned Samsung B-dies that I was able to run at DDR4-3600 14-16-16-16-36-1T at 1.45Vmem. Note that all listed settings are best-performant OC I was able to squeeze out from each type of RAM kit while passing 4 passes of MemTest86 with SMT on (all-core) and 48-hour uptime servicing both heavy CPU and GPU loads simultaneously without idling (I have 2x Radeon VIIs running at 1375MHz@825mVcore and 801MHz@850mVhbm2 that draw ~125W and a stock ITX SFF RTX 2070 in the case). Well the dual-rank 16GB DIMMs at stated OC settings still outperformed tighter-tuned OCed single-rank 8GB counterparts in aggregate RAM benchmarks by about 3-4GB/s aggregate RAM bandwidth and memory latencies were a toss up within a margin of error. For RAM, as a rule of thumb, with single-rank DDR4 DIMMs (usually 4/8GB DIMMs) you need at least 2 tighter timing stops all the way across the board to even try to match the dual-rank DIMMs in real performance/throughput (i.e. 14-14-14 single-rank vs 16-16-16 dual-rank at the same clocks), let alone surpass it. That's it on OC front. Find the best settings to keep the CPU within your cooling budget (important!) and keep the CPU package temperatures low enough (don't get scared by higher core hotspot temperatures that go to 80+C, look at package temperatures).
A little note on 3980X (48c/96t, surely the SKU is in the works for later) and 3990X (64c/128t). These are not well-balanced chips compared to their EPYC counterparts and start lose linear scaling above 40c/80t workloads in many real-world workloads reliant on RAM bandwidth to keep the CPU cores fed (outside of benchmarks that fit into caches) because EPYC counterparts have 8-channels of memory but these TRX40 parts only get 4-channel memory. You will also need Enterprise version of Windows to correctly detect/utilize beyond 32c/64t and NUMA CPU configurations.
| Processor | 4.5 GHz ryzen_threadripper_2990wx |
|---|
