Reports putting AMD's new architecture's performance boost much higher than expected are piling up
The day when AMD releases processors with the new Zen 5 core, allegedly the biggest upgrade since the first Zen, is closing in. Their performance remains quite unclear though – on the one hand AMD’s materials talk about a 10–15% or a little bit more performance increase per 1 MHz, but at the same time there are rumors talking about 30% or now even 40% performance increase. So what to believe and what to watch out for?
As for the first more conservative figure, it comes from pretty good sources. It’s reported in an AMD document, which admittedly was shown by the youtuber Moore’s Law is Dead, but the architectural details have since been officially confirmed and in general these slides seem to be real, so we wouldn’t doubt the authenticity too much. This document states that Zen 5 is supposed to have a 10–15+% higher IPC (i.e. performance at the same clock speed) compared to Zen 4. That plus sign indicates that this is a conservative number and reality may be even better, for example 18%. The company simply keeps some margin in its prediction in case something goes wrong with the core, impairing final performance.
- Read more: Leaked AMD documents show Zen 5 and 6 IPC and architecture details
- Read more: AMD confirms Zen 5 details: 6 ALUs, full-performance AVX-512
But even if we take that margin into account, it still doesn’t fit the second set of much more optimistic reports and rumours that talk about performance and/or IPC going up by, say, even 30%. For example, just recently, the leaker going as Kepler_L2 stated (on the AnandTech forum) that performance core for core is supposed to go up by more than 40% for Zen 5, compared to Zen 4. That’s a lot different from the “10–15+%”.
It is probably safe to say that these high numbers are probably not referring to “IPC” (i.e., an increase in performance at the architecture level in isolation from clock speeds), and probably not even an increase in single-threaded performance as we usually define it, i.e., in typical consumer client software running on a single thread. In fact, Zen 5 will probably go much (if at all) higher in the maximum boost clock speeds compared to processors with Zen 4 cores, which top out are between 5.0–5.7 GHz.
Why +40% performance in multithreaded tests does not mean 40% improvement in a single thread
So is it even possible for the performance to be 40% higher “per core” in some cases? Probably yes, but you have to take into account that this number could be based on server processors. Specifically, it’s supposed to be a performance increase per core in the SPEC CPU 2017 benchmark, more accurately probably SPECint, that is, the integer part (not floating-point). Importantly, this is a multi-core “rate” benchmark, where the total score is then divided by the number of operating threads or cores, which then yields the aforementioned performance per core. Despite the fact that we are talking about performance per core (or core-for-core), it is a score from a multi-threaded performance test.
This may be the reason why the performance increase is reportedly so large, although the IPC should only increase by a little over 15%. Also, this is probably a figure for an Epyc server processor with Zen 5 (the one codenamed Turin) compared to an Epyc 9004 Genoa, not a performance increase on a Ryzen CPU. There are several things that can explain getting a 40% performance increase despite the core having a less than 20% single-thread IPC improvement.
The first thing that probably comes to mind is that Zen 5 will run at higher clock speeds under load than Zen 4, which could probably happen in multi-threaded tasks (as opposed to single-threaded tasks) on server processors because their operating clock speeds are not near the maximums possible. However, according to the leaker this is not the reason, or not the main reason, the clock speeds will probably be more or less the same. Either way, though, this factor won’t really help strictly performance of the desktop and laptop Ryzen 9000 processors in single-threaded programs (or games), because Zen 5 probably won’t go much higher in clocks than Zen 4 in those.
Better scaling, not directly better core performance?
The main factor here is probably that we are talking about multi-threaded performance, albeit divided per core. This means that the so-called single to multi-thread scaling is be better, which doesn’t mean that you will also measure better performance when you use a single thread program.
Zen 5 has a significantly wider core with six ALUs (+50%), but this likely means that more compute resources are left unused in typical single-threaded processing. SMT capitalizes on these unused resources by processing two threads simultaneously. This means that Zen 5 could have a better performance gain when using SMT – for example, if on Zen 4, SMT typically adds 10–15% of multi-threaded performance, then on Zen 5 it could be 15–25% (the numbers are made up just for illustration). But the important point here is that this gain from SMD only shows up in multi-threaded tasks, not in single-threaded programs. Thus, it may explain while you only see 15% better IPC in a single-threaded programs, but overall a processor with the same number of cores and threads will give 40% better performance.
Epyc will have the advantage of faster memory, Ryzen will not
Another such thing is memory bandwidth, which usually doesn’t limit performance in a single-threaded program, but can bottleneck multithreaded tasks quite a bit. And recent leaks show that the server version of Zen 5 will apparently introduce officiall support for DDR5-6000 memory, which means 25% better memory bandwidth versus Zen 4 (Epyc 9004 only supports DDR5-4800) and thus this gives potential for improved multithreaded performance.
But this factor probably won’t help desktop and laptop Ryzen 9000 with Zen 5 cores, which already use such fast memory via XMP profiles, and desktop Zen 5 probably won’t improve memory support much (maybe from DDR5-5200 to DDR5-5600, at best?). Either way this probably won’t affect the performance of single-threaded applications.
It should be said that in general, better performance scaling from single to multi-threaded like in the SPECint rate test may be due to various improvements that will only be applied to the server version of the processors, and not to the desktop Ryzens. The latter certainly use the same IO die, interconnect logic and memory controller as the Zen 4 (while this is not yet certain for the server version). But that larger SMT yield could also occur in Ryzen.
AVX-512?
It is also possible that the SPEC 2017 test benefits a lot from the AVX-512 implementation on the Zen 5 core. It will have fully 512-bit wide pipelines executing these instructions, so using them may increase performance in some subset of some algorithms by up to 100%. And that can inflate the performance increase in SPECint up to that 40%, even if in most programs we’re only going to be talking about that 15–20%. But at least those AVX-512 gains should be seen on desktop (and laptop) processors as well, in SPEC 2017. That is, if they’re not only made possible by the memory bandwidth improvements of the server processors…
So as you can see, those reports of 30–40% performance improvement per core, or even such a high IPC improvement (when measured in multi-threaded applications!) may well be true. But at the same time, that may not and likely will not carry over into the single-threaded performance, which is what gets most focus in PC processor reviews (i.e. in tests like Geekbench (ST), Cinebench ST, CPU-Z), or into game performance tests. These areas may well see a performance improvement of just those 15% plus maybe a little bit more.
So keep a cool head when looking forward to Ryzen 9000 and don’t have exaggerated expectations. These often lead to disappointment (as those who expected extreme jumps in IPC from the Apple M3 and the A17 Pro processors know).
Source: VideoCardz, AnandTech (forum)
English translation and edit by Jozef Dudáš
⠀
- Contents
- Reports putting AMD's new architecture's performance boost much higher than expected are piling up
