Zen 5: I think we have to have a talk

Sorry for the click-bait title…

TL;DR: Zen 5 shines in small form factor PCs; if you are building a Ryzen system which will be cooling or power constrained and you have a choice between Zen 4 and Zen 5, go Zen 5. You will have higher performance, and less heat.

I recently picked up a MS-A1 from Minisforum. I wanted to get a little machine for my desk which I could use for various tasks, either hosting VM’s, playing games (No AAA, just things like World of Warships and the like), Linux, Windows, etc.

In reading the reviews on the Zen 5 launch, and the 9000 series, it has been almost universally panned. Small performance gains for games over Zen 4. When listening to Wendell, Linus and others, we hear that it is more optimized for server use, and the gaming side is a bit of a disappointment.

After some testing this week, while the criticism on Zen 5 for games is likely well true, I’m not sure that analysis holds true for small form factors, or with less than optimal cooling.

As background, I have a few AMD Threadripper systems (5955wx) and they have been plenty powerful. I also have Sapphire Rapids based W Xeons, and they have served me well too. I primarily need the PCIe lanes, so have been willing to live with slightly less performant cores than client. Given the issues with Intel 13 and 14th gen, it’s probably been a good choice to make.

I decided to try out the MS-A1 as it was upgradable and I wanted to get into AMD client processors. I chose the barebones model. Interestingly, what arrived this week was not the same was what Tech Youtubers reviewed in earlier this year. I received a 1.1 unit that had a different bottom layout. The SSD fan, now forms part of the backing plate for the CPU retention mechanism, and it cannot use the U.2 adapter. In fact, mine did not come with one. The layout is slightly different as well from what I can tell from pictures.

For the build, I bought 3 different CPUs: a 8700G, 7900x and a 9900x. Going for broke, based on the fact the system “looked” different than the pictures, I wanted to see if 9000 series would boot. Minisforum says its compatible and needs a BIOS update.
Yet they do not mention how to get that update. The 1.1 version I received does in fact support Ryzen 9000 series. No BIOS updates required. Upon checking my BIOS version is 1.06, and the latest version for download is 1.03. I have DDR5 5200 SODIMM, which the MS-A1 locks to 4800.

I decided the torture the MS-A1 and do some bench testing with it before I put it into regular use. I wanted to see how the 9900x copes with being in small spaces, and what the thermals bring. Being a Threadripper/Xeon person, hot, loud, slow comes to mind with modern CPU’s and low-end cooling solutions. I was expecting the machine to be loud and hot during the testing, since pinning the MS-A1 at 100% with 12 Zen 5 cores for over 24 hours is likely no small feat to keep cool.

This is where I discovered that all is not well with the Zen 5 reviews. While yes, they do show less performance at times, the overall performance of the 9900X in the small form factor has been impressive to say the least.

Let’s start with temps.


This image is taken from the machine running for a little over 5 hours under a Phoronix test workload. The temps stayed consistently in the 80-81C range, and the fan noise was audible but not excessive. The clock stayed typically above the 4.4 base, and most times sat at 4.5Ghz. When the system had a chance to drop load during tests, and cool a bit, the boost clock would shoot up again to 5.6Ghz and settle back down to the 4.4-4.5Ghz level. The back of the unit after 5+ hours had a lot of warm air coming out, but it was very not as scorching hot as I thought it would be.

I became curious. What if we start to compare results to previous “ideal” benchmark runs. These would be full systems, with better memory, better cooling. Surely the 100W TDP of the MS-A1 holds the 9900X back, even if it’s a 120W part.

From that above test, we start to see this pattern emerge, where when limited to 100W TDP in the MS-A1, the 9900X is ~10-15% slower than its fully cooled desktop counterpart, with higher performance memory. What gets more interesting is how it compares to the 7950X.


We see pretty consistently that the 100W TDP constrained 9900X still manages to outperform the 7950X in some cases. As we can see from other benchmarks, this is not a fluke.

In comparisons with a fully unlimited 9900X, the 100W TDP version is not substantially slower in the MS-A1.

Cpuperf-241004, AMD Ryzen 5 9600X / Ryzen 7 9700X / Ryzen 9 9900X / Ryzen 9 9950X Linux Benchmarks Performance Comparison - OpenBenchmarking.org

When we boil down the numbers, if we consider the 15% drop off that we get from 120 to 100 TDP in the MS-A1, this shows the thermal power performance of Zen 5 scaling quite well. If I had used the 7900X with its 170W TDP, it’s quite likely the performance in the MS-A1 would be quite poor overall. Its possible that a 7900X in the MS-A1 would perform at only 60-70% of what it would be in a larger well cooled chassis.

If we compare the 7900X and the 9900X , where the 7900X is in a full desktop, with faster memory and better cooling, against our small form factor MS-A1 9900X, you can see from the benchmarks that the MS-A1 9900X wins. It’s cooler and uses less power as well.

R9-7900x Benchmarks - OpenBenchmarking.org

I was a little surprised by the results. I thought for sure that the 9900X in the MS-A1 would be slower than a similar Zen 4 7900X, based on the reviews and what people have been posting. This testing shows that the use case for the CPU matters. In “ideal” conditions, a 5% bump from a Zen 5 to Zen 4 is not that impressive. This becomes a substantial 20%+ win for Zen 5 when we power and cooling constrain Zen 4.

Bottomline. If you are building a small form factor machine or you will be power/cooling constrained, and have a choice between Zen 4 or Zen 5, pick Zen 5. You will get higher overall performance than Zen 4. Oh, and for the 8700G, 0/10 do not recommend!

6 Likes

APU 8700G essentially is a laptop chip. It can run as low as 15W. If you are really constrained with power, you have to go with 8700G. 8000 is on 4nm node which is the same as 9000 desktop parts.

Check out Phoronix’s workload breakouts. My own benchmarks for the workloads I code are much the same for 7950X versus 9900X. AMD’s probably not happy about gaming reviews diverting attention from Zen 5’s other abilities or formalizing a 9900X price drop three days after launch but, eh, works for me.

The cTDP spec’s 45-65 W, though 35’s also an option? The corresponding laptop parts are the 8840U (15-30 W) and 8840HS (20-30 W) if that’s what you meant.

Not sure if the 8700G’s higher power is just the packaging shift to AM5 or if there’s other changes. Zen 3 U parts were on different design rules from H and desktop but I’m not having luck finding Zen 4 data that’d indicate on that.

1 Like

Hi,

You won’t believe how happy I am to read this post here.
This is exactly the configuration that I was thinking about. (9900x + MS-A1 + 96 GB RAM + 4 SSDs maybe a 5th small one instead of the wifi card).
One question that I still have: will all 4 SSDs work? Some CPUs will not have the PCIe lanes needed according to the Minisforum webpage and they don’t really write anything about the 9000 series.

Thanks
B

Yes everything should work. The 7000/9000 series CPUs have all lanes enabled; only the 8000 series have limitations (some more than others.

1 Like

Thanks, you helped me choose my own Christmas present:D
B

1 Like

NAScompares did a recent video in the MS-A1 with all drives populated,

NAScompares MS-A1

They did this with the 8700G if I recall. Performance on all 4 drives being used at once was not great. They recommended the MS-01 if you need multidrive speed.

I’ve not run mine with more than 2 drives and did not notice slow drive to drive copies. I chalked it up to how NAScompares does their testing.

Otherwise rock solid with good temps with 24 hour stress testing on my end.

HTH

1 Like

Maybe one last question:
What kind of thermal paste did you use when installing the CPU?

Thanks

Just used what came on the heatsink. Had access to MX6 and Kryonaut but wanted to see how stock worked.

HTH

+1. If it’s a decent cooler the stock paste’s likely also decent, though YMMV with Noctua’s intermittent quality control and such. The 9900X I use is under TF8 because I wanted to try it but the stock TF7’d be fine.

So, first of all, wow…
I am amazed that such a page exists :smiley:

Thanks for the insights for both if you.
I’ll now read this inside and out

Cheers

1 Like

FWIW, it’s not as rigorous but the other major paste test set I know is at Tom’s. But, y’know, 9900X + TF7 + Phantom Sprit 120’s tough to beat in an airflow case.

Your findings here are almost certainly not unique to Zen 5. Chip power/heat scales as roughly the 3rd power of clock frequency, so pretty much any CPU can trade away an almost-imperceptible amount of performance for a large power reduction. So when you compare across architectural generations, the tech advantage can easily outweigh the effect of power.

Another thing is that many (most) workloads are single-threaded or light enough on CPU that power limits aren’t hit. When power limits aren’t hit, they don’t matter. In that case, only architecture, clock speed, and memory speed make a difference.

Also, I noticed that first PTS benchmark set you linked has different CPU governor parameters (EPP=balance_performance for the full-power desktop, EPP=performance for the compact). That probably makes a significant difference to light tasks that don’t hit the power limit, so they are not comparable. Even aside from the different compiler versions and kernels.

But luckily, because there was Much Ado about the 9700X’s low stock power limit, Michael at Phoronix ran a huge test comparing 88W PPT to 142W PPT. Be sure to check the CPU power consumption for individual benchmarks to see which ones are riding the limit and which are not.

power limits determine clock speeds at high load. And a server runs multiple things and 20x single-threaded processes can add up to considerable load. And AM4/5 boost behavior on 2 or 4 core basically hits PPT limits. That’s a game server or two, or some janky process in a VM, or a maintenance/update job.

The background noise and chatter distributing over all cores isn’t any load to speak of.

With 40+ VMs and containers, I see load spikes on single cores regularly. And changing PPT directly translates to the fan noise I’m hearing. Power limits are very useful. Because my Ryzen goes crazy on power boosting to limits if 2,4 or 6 cores want attention. Full 12 core load is much more manageable as the heat gets more distributed across the CCDs.

All my Zen CPUs run on 65W TDP aka 88W PPT nowadays. Sweet spot for a lot of stuff