System Tests: Dual Channel vs 8-Channel

For testing the processor at full glory, head on over to our review of the AMD Threadripper Pro 3995WX, which shows all the data with respect to other CPUs when it is fully enabled with eight memory modules – we go more in depth in our testing in that review, and the results of those tests and more can be found in our benchmark database, Bench. For this review, we’re doing something slightly different.

Our review sample from Lenovo was shipped with two 16 GB modules of DDR4-3200 ECC, enabling only dual channel memory. The list price difference between a 2x16 GB arrangement and an 8x16 GB arrangement if you buy modules from the open market is $600-$1000 depending on where it is sourced. At the rates that Lenovo buys the memory, this is more likely to be nearer $200-$300. So why did Lenovo skimp on the memory for a system that is sold for nearer $20000? I have no idea. But the performance difference is so much more than the 1% price increase of filling it full of memory. This is something I want to report on, because it matters.

We’ve asked Lenovo in the past why they ship review systems with fewer memory modules than at least one module per channel. The response usually boils down to ‘that is how our customers buy them’ or ‘it doesn’t matter for benchmarks like SPECviewperf’. Of the other media who were sampled this system, we know at least two others were similarly sampled with only two memory modules, both in the US, whereas our European counterparts were shipped with all eight modules, likely down to how each region manages review samples. Luckily Kingston came in to save the day, and supplied sixteen of their 16 GB KSM32RD8/16ME DDR4-3200 RDIMMs for all of our server testing.

But we want to compare the difference between a dual channel system, and proper system performance. Rather than offer a dozen pages showing all our test results, we’ll list the separate the sections we test, and score how many benchmarks are affected by the memory difference for a given percentage, along with notable highlights.

Benchmark Results

8-Channel Benefits over 2-Channel
AnandTech # Tests
in Segment
# Tests
> 103%
of 2-Ch
# Tests
< 97%
of 2-Ch
Average
%
Highlight
Rendering 11 2 0 +1% Corona, +7.5%
Office 3 1 1 +2% Agisoft, +8%
Science 9 6 1 +53% yCruncher MT, +212%
Simulation 5 1 0 +10% DigiCortex, +47%
Encoding 8 6 0 +34% 7z Encode, +135%
Legacy 10 0 0 Equal -
Web 3 0 0 Equal -
OpenSSL 8 0 0 Equal -
SPEC2006 ST 18 3 0 +2% 470.lbm, +10%
SPEC2017 ST 22* 3 1 +1% 549.fotonik3d +13%
SPECviewperf 8 1 0 +1% Maya, +5%
Total 105 25 2 +15% -
*521.wrf_r doesn't run in our suite due to WSL limitations

Overall, we’re seeing +15% uplift with the 8-channel configuration. But that’s skewed by several super high results:

  • yCruncher MT: +212%
  • AES: +62%
  • DigiCortex: +47%
  • NAMD: +39%
  • 7-zip Overall: +31%
  • AIBench: +29%
  • WinRAR: +23%

If we move to an even tighter 1% margin, then more than 50% of our tests still fall within that +/- of 1% (55/105).

To the point of rendering programs and SPECviewperf, having eight channels of memory doesn't make a difference to performance. As we discussed in our Threadripper Pro 3995WX review, the key market for these processors is the rendering market. For most of our rendering tests, we didn’t see that much difference between 8-channel and 2-channel, with the most notable being the Corona renderer. Similarly if we look at SPECviewperf, overall results are comparable – 8-channel is +1% higher in SPECviewperf overall. This would tend to agree with Lenovo’s line of reasoning, that certain customers are more concerned about memory capacity and performance, than filling up the memory channels.

However, if the workload is encoding (+34% gain) or science (+53%) gain, then that memory upgrade is crucial to the performance of this system.

The full run-down of the side-by-side tests can be found in our benchmark database.

System Benchmarks: Power, Temps, Noise Conclusion: Do It For The Bits
Comments Locked

47 Comments

View All Comments

  • eastcoast_pete - Tuesday, February 16, 2021 - link

    My guess on why Lenovo US shipped this review unit with such limited (and limiting) memory (two channels out of eight?!) is this: If the results are (still) good, no harm done, if the results are disappointing, well then, it's because the workstation was hampered by too little memory in too few channels, and nobody who buys a 64 core setup would skimp that badly on memory. It's either this, or the people in charge of marketing these high-end workstations are otherwise selling $ 350 entry-level laptops. And that thought would give me pause when thinking of buying such pricey equipment as the P620. I'd like the people who sell them have at least some idea of what they are selling and to whom.
  • tyger11 - Tuesday, February 16, 2021 - link

    I'm waiting for the Zen 3 version of TR Pro for my video workstation.
  • Snowleopard3000 - Wednesday, February 17, 2021 - link

    Now only if they can fit all this in a 17 inch Clevo Laptop...
  • drajitshnew - Wednesday, February 17, 2021 - link

    I wonder why AMD epyc and TR--WS have 120-128 PCIe, when nobody implements it, ever -- perhaps an intel design where there are only 50 lanes with switches would be better from a silicon area perspective.
    Though I really think that AMD could have made a reference board -- like Intel used to do.
  • DanNeely - Wednesday, February 17, 2021 - link

    ATX/EATX boards aren't big enough to use all 128 lanes. You could fit 8x cards in a rackmount config. I suspect the max lane count is for high capacity flash storage boxes though. Even after subtracting off a few lanes for multi-gig networking you can cram about 30 x4 SSDs into an Epyc box without any lane contention.
  • drajitshnew - Wednesday, February 17, 2021 - link

    I believe there are alternative form factors that would allow that many lanes
  • drajitshnew - Wednesday, February 17, 2021 - link

    that is why I suggest that AMD might make a reference design
  • vegemeister - Wednesday, February 17, 2021 - link

    Aren't half the lanes used for talking to the other package in dual-socket systems?
  • phoenix_rizzen - Sunday, February 21, 2021 - link

    CPU 1 has 128 lanes.
    CPU 2 has 128 lanes.

    64 from each CPU are used to communicate between them. So 128 lanes for interconnect.

    64 from each CPU are available for connecting to peripherals. So 128 lanes.

    Whether a single socket or a dual socket, there's 128 lanes available for peripherals.
  • jabber - Wednesday, February 17, 2021 - link

    I'll pick one up in 5 years time for £400...

Log in

Don't have an account? Sign up now