I have just made my last hardware upgrade (at least for the next three years)!

[RjL190365]

Two-and-a-half months ago, I upgraded the guts of my longtime main PC from a 4th-Generation Intel i7-4790K quad-core CPU to an AMD Ryzen R7 3800X CPU. I have since added a 500 GB Samsung 970 EVO Plus m.2 NVMe SSD to that PC, but remained saddled with a GeForce GTX 1060 6 GB card that I have had for almost three years.

Today, I have just upgraded that - to a GeForce RTX 2060 SUPER 8 GB card. Before I installed it, I moved the GTX 1060 to my spare miniITX PC that still has a 7th-Generation Intel i7-7700 CPU. I then ran the "Standard" preset of the Puget Systems' Premiere Pro benchmark on that PC with both a GTX 1050 Ti I had in there before and the GTX 1060. The GTX 1060 made only a slight improvement in the overall score over the GTX 1050 Ti with that quad-core CPU, with overall scores of between 340 and 352 with either of those GPUs.

When I ran the Puget tests on my recently upgraded main PC (still with the GTX 1060), it scored a rather poor 518 using the "Standard" preset. And the primary reason for such a poor score compared to what the CPU is capable of is that the GTX 1060 itself is a weakling these days. The overall playback score with 4k/59.94p heavy GPU effects was less than half that of the overall playback score of the R7 3800X with 4k/59.94p heavy CPU effects. But after removing the GTX 1060 and installing the RTX 2060 SUPER, the overall "Standard" preset score jumped from 518 all the way to 600. The GPU's overall playback score with heavy GPU effects, although somewhat higher overall than my system's CPU playback score with heavy CPU effects, is now much more in line with what I had expected from my system - and it also showed just how important the CPU/GPU performance balance really is.

So there you have it. For Premiere Pro it is best to spend roughly equal amounts of money on both the CPU and the GPU {in the case of the R7-3700X or 3800X, the ideal GPU would be an RTX 2060, with or without the "SUPER" designation, or even eVGA's new $300-ish RTX 2060 KO (which is actually a cut-down version of the RTX 2070/2080 rather than a from-the-ground-up midrange GPU)} or a newer GPU whose performance is the equivalent to the comparable GPUs in existence at the time the CPUs were new.

Randall

[Kevin-Monahan]

Thanks for sharing your findings with your hardware updates, Randall. I believe what you suspected is going wrong with a lot of other systems out there was proven to be the case with your system, as well. You need a nicely matched CPU and GPU for optimal performace. Glad you've got that sweet, sweet GPU power now.

Regards,
Kevin

[RjL190365]

Thank you for the encouragement, Kevin. Just to prove the opposite, I could run my i7-7700 with the RTX 2060 SUPER, but why would I would want to? The Puget Systems' benchmark result from that combo would not have been sufficiently higher than with the GTX 1060 to justify the added cost because the CPU would then have significantly bottlenecked the GPU. The "Standard" preset score with that combo would have been closer to 360 to 370, rather than the 352 with the GTX 1060. That makes the RTX 2060 SUPER too expensive to be worth using with a 4-core/8-thread CPU. A better match for that CPU would be anywhere between a GTX 1650 SUPER and a GTX 1660 SUPER (among current Turing GPUs).

[RjL190365]

There is one more piece of advice that I will be giving in this discussion:

I do not recommend buying a high(er)-end system just for the sake of "future-proofing." That concept does not exist any more. You see, by the time that new software requires all the power of that (very) expensive CPU, it will also require hardware capabilities that are not present at all whatsoever in current hardware. So instead, I will recommend that one purchases a PC build that is sufficient for today's needs, and worry about the future later.

EDIT: I re-ran the benchmark with the 0.9b version of the PugetBench set on my system running the 14.0.4 version of Premiere Pro, with the RAM set to both DDR4-3200 and DDR4-2133. While my result with the memory set to DDR4-3200 is as expected (the Extended overall score of 637 and the Standard overall score of 611), the result with the RAM speed set to only DDR4-2133 showed a drop - all the way down to 596 for the Extended and 570 for the Standard. Since that time I saw another Ryzen 7 3800X system pop up with the ostensibly superior-performing RTX 2080 SUPER GPU but with that system's 64 GB of RAM running at only DDR4-2133 speed. That system scored better than mine did at that same RAM speed, but still underperformed my same system with the memory set at DDR4-3200. Why would one shoot himself in the foot by going with a nearly top-of-the-line GPU for a moderately high performance AMD Zen2-based Ryzen platform only to bottleneck that entire system with slow system RAM?

Randall

[ericwillcow]

Thank for sharing your findings.

[RjL190365]

And I have seen results from more systems over time. Including some systems that rely solely on the integrated on-CPU graphics. The results from those particular systems without discrete GPUs were very, very disappointing: Their overall Standard preset scores were extremely low, never reaching even 200, because in those setups the IGP never allocated sufficient RAM to continue utilizing MPE GPU acceleration. The IGP all too frequently got its entire allocated RAM depleted, resulting in the MPE renderer being slammed mercilessly into the software-only mode in those systems early on in the PugetBench testing. The AMD Ryzen G series APUs ended up running out of allocated RAM even earlier than the Intel HD or UHD Graphics did; as a result, the Ryzen 3 3200G ran almost the entire time in MPE software-only mode (the renderer says OpenCL in the results page, which was the setting that it was in before it got slammed into the software-only mode). No wonder why it averaged only a 36 in the PugetBench for Premiere Pro. The Intel HD/UHD/Iris Graphics-based PCs that lack a discrete GPU went a little further in the testing before they too got slammed into the software-only mode. They ended up spending well over half of their testing time in software-only mode.

The very same fate befell the i7-860 system, which scored an all-time worst (so far) 15. No ATi Radeon HD 5800 series discrete GPU had more than 1 GB of VRAM to begin with. That falls well below the minimum recommended amount of VRAM, which is 4 GB or more.

Based on these testing results, future releases of Premiere Pro should REQUIRE a discrete GPU with 4 GB or more VRAM just to even enable MPE GPU acceleration at all. Integrated iGPUs and discrete GPUs with an insufficient amount of VRAM should "permanently" lock Premiere Pro's renderer to the MPE software-only mode.

[RjL190365]

With the release of Premiere Pro 14.2 and 14.3, these original test results (in Premiere Pro 14.0.4) are now superceded. Newer Nvidia GPUs can now take advantage of the NVENC hardware H.264 and HEVC encoding in 14.2 and 14.3. 

With those results now in, my main PC now scored 734 on the PugetBench test (using the Standard preset), while my reserve i7-7700 PC (after a GPU upgrade to a GeForce GTX 1650 SUPER) now scored 445 (with QuickSync enabled). Still a significant upgrade for the AMD system I now have as my main PC. The Intel system's export score is dragged down by the PCH, which really isn't designed for more than a single disk (bandwidth-wise).

[RjL190365]

One final update in this discussion:

My new system wasn't entirely stable at even stock speeds with either the original 32 GB of DDR4-3200 RAM or the 64 GB DDR4-3600 RAM that I bought to replace it. Sometimes, it failed to POST, forcing a hard reset. Other times, it crashed in mid-run of the PugetSystems benchmark. And the latest BIOS update for that motherboard (an Asus PRIME X570-P) was the last straw: The system all too frequently shut down and rebooted even when it was idling. At this point I have concluded that it was a flaky motherboard all along.

As a result, I went out shopping for a replacement motherboard yesterday, and ended up purchasing an Asus ROG STRIX X570-F Gaming to replace the X570-P. And while I picked up the motherboard I also picked up a Noctua NH-U12S cooler.

So far, my rebuilt new build is very stable, even with the new BIOS version for that motherboard - and at DOCP settings (with my 64 GB DDR4-3600 kit that's currently in the system). But only time will tell if it remains stable in the long term.

And that's not to mention that I had originally purchased the X570-P on promising reviews from professional Web sites (relatively good-quality VRMs for the price). But good VRMs, as it turned out, are irrelevant if the rest of the motherboard cheaped out on other components, which is why it wasn't 100% stable at even stock settings. At the time I purcahed the X570-P, the only other X570 motherboard choice that I had that was within my budget then was an MSI X570-A Pro, which had (and still has) terrible VRMs (or more specifically, terrible VRM design) - and that would have caused my CPU to throttle down to far below stock speeds during longer computer sessions.

The moral of this story is: If you're going to buy a CPU with a TDP that's no greater than 65W, or if you're going to buy RAM that's strictly JEDEC with absolutely no XMP or DOCP profiles whatsoever, then these cheap motherboards are fine. But anything beyond that (this means most memory kits that are on the market right now), then you'll need a better motherboard than these sub-$200 models. There's a BIG difference in quality between a $150 motherboard and a $250 motherboard.

Postscript: I almost purchased an Asus Crosshair VIII Hero motherboard, which costs $100 USD more than the X570-F that I ultimately purchased. But spending almost $400 on a motherboard is overkill unless I wanted to run my 3800X at a constant 4.3 GHz on all cores. And the $700 motherboards are IMHO only worth it for extreme overclocks, especially with custom water cooling or liquid nitrogen cooling.

[RjL190365]

Just one final, final public service announcement:

I do not recommend the installation of more than 32 GB total of RAM in any mainstream desktop CPU platform with a dual-channel memory controller. I noticed this after running my usual PugetBench tests with both 32 GB and 64 GB. You see, running 64 GB of RAM (using two 32 GB DIMMs) at DDR4-3600 speed results in no better of an overall Extended and Standard scores, and sometimes worse overall scores, than 32 GB of RAM (via two 16 GB DIMMs) that's running at only DDR4-3200 speed. And it's not the primary timings that are the cause of this degraded performance. It's one of the secondary timings - the tRFC. And currently, 16 Gbit ICs have a much higher tRFC latency than 8 Gbit ICs - 550 ns versus 350 ns. No wonder why at stock JEDEC DDR4-2666 settings and timings, 32 GB of RAM performed significantly better in PugetBench than 64 GB of RAM. In fact, 64 GB of DDR4-2666 RAM performed no better than 32 GB of DDR4-2400 RAM because of the wide difference in the tRFC latencies.

And going 64 GB of RAM via four 16 GB DIMMs will likely result in the memory controller failing to sustain higher clock speeds.

So, the moral is: If one will be installing more than 32 GB of RAM in a desktop PC, then I am now recommending a HEDT (High-End Desktop) build that's substantially more expensive, even without any RAM installed. 32 GB DIMMs currently aren't ready for prime time yet.

As for 64 GB DDR4 DIMMs, fuhgeddaboudit. 32 Gbit ICs will have a whopping 1250 ns (1.25 us) worth of tRFC latency! As such, DDR5 system RAM (successor to DDR4) must be not only on the market, but also matured enough, to render 64 GB unbuffered non-ECC DIMMs viable as the very high throughput (in GT/s) would more than offset the significant increase in latency.

[RjL190365]

One final tweak to my system:

I discovered that I have cheaped out on my m.2 NVMe SSDs. While the 500 GB Samsung 970 EVO Plus SSD was fine for cache disk use, my other SSD (a 1 TB Inland-branded Phison/Toshiba-based SSD) was slower than I would have liked as a working media disk: It claimed 2800 MB/s in sequential reads - but only for the SLC-mapped turbo cache, which was smaller than that of many other SSDs. Once that cache fills up, the true TLC sequential write speed rears its ugly head - in this case, only about 1024 MB/s.

I replaced it with a new 1 TB Samsung 980 PRO m.2 PCIe 4.0 SSD. Unfortunately, unlike the 970 PRO predecessor, the Samsung PRO-series SSDs now also began to rely on TLC NAND with a Turbo Write cache. But even so, the 1 TB Samsung 980 PRO produced a sequential TLC write speed (the true write speed) of about 2.2 GB/s, which is a bit faster than the true TLC write-speed performance of the 1 TB Samsung 970 EVO Plus SSD (at 1.7 GB/s). (For the record, the true TLC write speed of my 500 GB 970 EVO Plus was around 900 MB/s.)

This faster NAND memory in the 980 PRO (in terms of the true TLC sequential write speed) pushed my PugetBench for Premiere Pro 0.95.1 scores from 730 Standard and 689 Extended with the old Inland SSD to 751 Standard and 710 Extended with the Samsung 980 PRO. Both the Standard and Extended export scores increased with the new drive. For some reason the Effects score bounced between 58 and 62 with the old drive.

I will re-run the tests a few more times just to show that the increase is for real.

[RjL190365]

Finally, after all that, I am now beginning to think that I had wasted money on all this upgrading of my mid-tower ATX PC. As it turned out, I made the wrong AMD CPU choice back in December 2019: The Ryzen 7 3800X wasn't worth even the extra $30 over the Ryzen 7 3700X even then. It hardly performed better than the 3700X in Premiere Pro while consuming significantly more power.

And all that, combined with the fact that my household was eating up significantly more electricity than my neighbors because they have appliances that also eat up the bill.

As a result of this, I rectified this by upgrading my mini-ITX breadbox to the 3700X because I simply did not want my smaller PC to go to waste. I have just completed moving most of the innards (mainly the GPU and RAM) of my former main tower PC to this mini-ITX breadbox, while saving the tower as a placeholder for a potential future upgrade to something that is more powerful than my current CPUs and GPUs.

So, my mini-ITX breadbox PC is now my main PC, the only significant difference apart from the "downgrade" to a 3700x is the choice of the AMD B550 chipset instead of the X570 chipset of my former main tower.

Lessons learned, especially with the COVID-19 pandemic still ongoing.

[RjL190365]

It is now nine months since I updated this discussion. The mini-ITX box is once again in storage, still with the Ryzen 7 3700X but now temporarily with the GeForce GTX 1050 Ti, while I upgraded my main box to a 12-core Ryzen 9 5900X (carrying over my GeForce RTX 2060 SUPER from my previous builds). An update of the OS from Windows 10 to Windows 11 did curb my now-main system's PugetBench performance scores a bit, such that it now just barely outperforms some of the better-scoring 8-core/16-thread Ryzen 7 5800X systems overall in the PugetBench database.

And today, with newer-gen discrete GPU prices still ridiculously high right now, and with energy costs going higher and higher for the foreseeable future (faster than my jurisdiction and employer could raise their wages), I am now tempted to scrap all of my desktop PCs completely (and sell off their constituent parts), and just go with an M1 MacBook Air as my sole computer system. Downgrading my main system's CPU would end up being a massive waste of money, IMHO.

However, I may be able to improve the energy efficiency of my main PC (or more specifically, attain a better performance-to-energy-consumption ratio) by dumping out my existing RGB fans and controller from my current case and just go with three 120 mm non-lighted or single-color-lighted PWM case fans. RGB fans and their required controllers consume CPU cycles (due to the software that they require) while degrading performance.

I also thought about dumping AMD for Intel once again, with the new 12th-Generation Alder Lake CPUs. However, I am currently ruling that out because this particular upgrade would not have improved my PC's performance-to-power-consumption ratio sufficiently for the price that I were to pay for the CPU and motherboard.

[RjL190365]

"It turned out that I do not need to downgrade or cross-grade my CPU platform after all. I ended up spending a lot of money ($300-ish!) on a motherboard whose VRM quality is underwhelming for the motherboard price. In fact, most cheaper AMD X570 motherboards have mediocre-at-best VRMs, which meant that the way the AMD platform is designed, if the motherboard's own devices couldn't handle it, the extra heat goes back into the CPU package! The CPU effectively throttled at a maximum temperature that's only 65 degrees Celsius due to this very narrow headroom in CPU package powers! It's no wonder why I get CineBench 23 multi-core scores that were subpar for this model of CPU. I get only 19,500-ish when it should have been in the lower 20,000's. This is all caused by the much narrower-than-expected headroom between the idle CPU package power and the maximum CPU package power.

To make matters worse, I had this same CPU running on a mini-ITX Gigabyte B550 motherboard, and the CPU performance was noticeably and consistently better.

As a result, my decision now is to just upgrade my main PC's motherboard, and donate my system's current motherboard to the Ryzen 5 5600X system that's currently under construction.

Hope this settles it."

It turned out after another round of testing that it wasn't my motherboard per se that's causing the relatively poor performance scores of my Ryzen 9 5900X system with the GeForce RTX 2060 SUPER - but rather the motherboard's default settings. The default package power limit of 142W happened to hamstring (or choke) my system's performance. In the case of my system's particular motherboard (an Asus ROG STRIX X570-F GAMING), the motherboard's default settings gave the CPU too much voltage during idle, resulting in that extremely narrow headroom between the idle power and maximum power. No wonder why its maximum all-core CPU turbo-boosted clock speed was a relatively low 4.0 GHz on all 12 cores: The default package limits throttled my CPU down at very cool core temperatures.

I re-tweaked the motherboard settings, and now my PC's Cinebench R23 multi-core score is back where it should have been. My PugetBench results for Premiere Pro will be forthcoming.

And to think that I needed a new PC or a motherboard upgrade would have ended up throwing even more and more upgraded hardware at the Premiere Pro performance problem. User error is mostly to blame here.

[RjL190365]

Two and a half years have passed since I started this featured discussion. It is now time to reconsider my upgrade options, and came to the conclusion that a CPU platform update and upgrade would currently give me more bang for the buck than simply upgrading the GPU or adding more RAM.

As a result, I have now decided to redo both my main PC and the PC that I am currently in the middle of building for my brother. The new CPU that I will be using is an Intel i7-12700K, and my brother's PC-in-progress will have a non-K i7-12700. I made the first of a series of purchases, the CPU for my own rebuild, and will resell my AMD parts. This will clear out most of my spare parts bin.

My mini-ITX reserve system will get the Ryzen 5 5600X that I currently have in my brother's would-be-build, and will remain equipped this way since I will not be needing it unless I need a third PC for my media editing work.

The brother's PC-to-be will be using DDR4 RAM and a B660 chipset motherboard. My system will be using a Z690 motherboard; however, I have not yet decided whether to switch to DDR5 RAM or stick with DDR4.

This will be the last post that I will be making in this discussion. I will post my progress in a new discussion.

Randall