AMD Zen 3 Ryzen Deep Dive Review: 5950X, 5900X, 5800X and 5600X Tested
by Dr. Ian Cutress on November 5, 2020 9:01 AM ESTSection by Andrei Frumusanu
CPU MT Performance: SPEC 2017
Whilst single-threaded performance of Zen3 seems to be an absolute win for the new Ryzen 5000 series CPUs, multi-threaded performance is also the other important aspect of a design. Generally, what comes into play much more for multi-threaded performance is the power efficiency of the design. As none of the current x86 consumer or enterprise parts are able to actually run all their cores at maximum frequency for peak performance due to platform power limitations, any resulting performance boost we might see between generations with a similar power cap will be due to power and energy efficiency gains between the designs.
For AMD, we’re limiting the detailed comparisons here to the 3950X and the 5950X which both have a PPT of 142W, which means that’s the maximum peak power for the platform, and observed 120-125W sustained figures in actual workloads. We’re also throwing in a 10900K for context, but given the very different core count numbers it doesn’t serve an exact apples-to-apples comparison.
Starting off with SPECint2006, the performance uplifts for the new Ryzen 5000 series here doesn’t seem to be all that significant in most tests, with uplifts well below those of the single-thread benchmarks.
Most of the tests are showing a 10% performance uplift, with the more memory heavy test showing no improvement. Some of the minor uplifts such as a 5% boost in 502.gcc seem quite disappointing and showcase that the new platform isn’t all that big of a boost for productivity workloads.
In SPECfp2017, 510.parest and 527.cam4 stand out as the two workloads with the biggest improvements, with the rest of the workloads all either having sub-5% improvements, or even just flat or slower performance than the 3950X.
** Preliminary figures to be re-tested due to 2x16GB SR vs 4x8GB SR memory DIMM configuration, scores will improve.
Overall, the new Ryzen 5000 series are showcasing a quite conservative 8-9% performance improvement over their predecessor generation SKUs. The 12-core variant here is showing a little bigger boost of 10-13%.
In this regard, the new chips seem to have missed the mark in terms of AMD’s claims for 12% better performance per watt for the new 5950X, and 26% better performance per watt for the 5900X – their marketing should have done better in explaining those figures, or just be more conservative with their numbers. This also doesn’t bode too well for the eventual EPYC-based Zen3 Milan chips.
Facebook
Twitter
RSS
339 Comments
View All Comments
TheinsanegamerN - Tuesday, November 10, 2020 - link
There is no x590 chipset coming. X570 is ryzen 5000s chipset.There's also this miracle fo technology, if you have a micro atx or full atx board, you can put in ADD IN CARDS. Amazing, right? So even if your board does not natively support 2.5G LAN you can add it for a low price, because 2.5G cards are relatively cheap.
TheinsanegamerN - Tuesday, November 10, 2020 - link
the x570 aorus master and msi x570 unify also have 2.5G lan. And surely there will be newer models next year with newer features and names, gotta keep the model churn going!alhopper - Sunday, November 8, 2020 - link
Ian and Andrei - 1,000 Thank Yous for this awesome article and you fine technical journalism. You guys did amazing work and we (the community) are fortunate to be the benefactors.Thanks again and keep up the Good Work (TM).
Rekaputra - Sunday, November 8, 2020 - link
Wow this article it so comprehensive. Glad i always check anandtech for my reference in computing. I wonder how it stack againt threadripper on database or excel compute workload. I know these are desktop proc. But there is possibility use it for mini workstation for office stuff like accounting and development RDBMS as it is cheaper.SkyBill40 - Sunday, November 8, 2020 - link
Once some availability comes back into play... my old and trusty FX 8350 is going to be retired. I've been waiting to rebuild for a long time now and the wait has clearly paid off regardless of how the is the end of the line for AM4 or well Ryzen 4 does next year. I could wait... but nah.jcromano - Friday, November 13, 2020 - link
I'm in a similar boat. I'm still running an i5-2500k from early 2011 (coming up on ten years, yikes), and I'll build a new rig, probably 5600X, when the processors become available. I fret a bit over whether I should wait for the next socket to arrive before taking the plunge, but given the infrequency with which I upgrade, I think it's likely that the next socket would also be obsolete by the time it mattered.evilpaul666 - Sunday, November 8, 2020 - link
I'd love to see some PS3 emulation testing added.abufrejoval - Monday, November 9, 2020 - link
Control flow integrity (or enforcement) seem to be in, and that was for me a major criterion for getting one (5800X scheduled to arrive tomorrow).But what about SEV or per-VM-encryption? From the hints I see this seems enabled in Intel's Tiger Lake and I guess the hardware would be there on all Zen 3 chiplets, but is AMD going to enable it for "consumer" platforms?
With 8 or more cores around, there is plenty of reasons why people would want to run a couple of VMs on pretty much anything, from a notebook to a home entertainment/control system, even a gaming rig. And some of those VMs we'd rather have secure from phishing and trojans, right?
Keeping this an EPIC-only or Pro-only feature would be a real mistake IMHO.
BTW ordered ECC DDR4-3200 to go with it, because this box will run 24x7 and pushes a Xeon E3-1276 v3 into cold backup.
lmcd - Monday, November 9, 2020 - link
Starting to feel like the platform is way too constrained just for the sake of all 6 APUs AMD has released (all with mediocre graphics and most with mediocre CPUs, no less). I hope AMD bifuricates and comes up with an in-between platform that supports ~32-40 CPU PCIe lanes and drops APUs. If APUs can't be on-time with everything else there's so little point.29a - Monday, November 9, 2020 - link
"Firstly, because we need an AI benchmark, and a bad one is still better than not having one at all."Can't say I agree with that.