What to expect from Ryzen 7000 / Zen 4? + 15% or + 40% performance?

… And it probably won’t stop right away. Some users do not admit that, just like each generation consists of a number of models, the release of each generation precedes the creation of a number of differently configured samples. And when there are a number of samples, different tests can be performed on different samples. Thus, when AMD reported that it achieved> 15% higher performance in a single core in one test (CineBench) than one Zen 3, these users automatically combined the result with a demonstration of up to 5.52GHz clock frequency in the game load and assume that CineBench took place on the same set, the same sample and the same clock settings as the game demo. And so that when CineBench ran at 5.52 GHz (and the current Ryzen 9 5950X reaches 4.9 GHz), the extra 15% is divided into 13%, which can be clocked, and 2%, which can increase the IPC.

Not surprisingly, someone comes to the conclusion with a logical construction that builds on unsubstantiated assumptions (the clock frequency of the sample on which CineBench ran was not actually published by AMD), but it is quite striking that anyone is able to believe that AMD two For years, she worked on an architecture aimed at increasing the IPC by 2%. Or that chipsets Zen 4which, due to their dimensions, provide space for ~ 50% increase of transistors compared to Zen 3should have consumed these 50% in addition to the 2% IPC.

Although the transistors’ budget can be used for elements other than increasing the IPC, it is clear that Zen 4 will not achieve a 2% increase in IPC – either on the basis of a simple improbability of setting such a target or on the basis of the use of erroneous logic.

Data on intergenerational performance shifts in the range of 25-40% have appeared for a long time. These numbers appear from a time when it was still assumed that there would be no significant shift in clock frequencies and from Zen 4 around 5.2 GHz was expected. This mathematically resulted in a 25% increase in IPC (1.06 [frekvence] × 1,25 [IPC] = 1,33 [nárůst výkonu]).

At present, it is clear that the increase in clock frequencies will be significantly higher. AMD se Zen 4 apparently decided to invest significantly more in design optimization to achieve higher beats. We know that AMD has a sample of 5.5 GHz. Furthermore, according to statements in “post-computer” interviews, we know that this value was relatively easy to achieve on the samples. However, in combination with leaking data on the existence of samples reaching 5.6 to 5.8 GHz (however, it is not written anywhere that it is not a system cooled by liquid nitrogen, we simply do not know), it is quite possible that the final form of the top model will target slightly higher than at 5.5 GHz – for example at 5.6 GHz.

So if we continue to believe that goal Zen 4 should be 25-40% power over Zen 3, then with a 14.3% increase in rates, 9 to 22.5% remain on the IPC. On average about 16%.

Then we have online resources. For example, the web Angstronomics claims to know specific data. However, it does not use the abbreviation IPC (Instructions Per Clock), but PPC (Performance Per Clock). The reason may be that under IPC, many people imagine a given and unchanging characteristic that is inherent in the processor core, and it does not matter whether the core is one or sixteen.

In fact, as the number of cores and the performance of the cores themselves increase, they are increasingly confronted with data transfer limits (whether at the kernel level or between the kernel and memory and whether in terms of latency or memory throughput). If only one core is active, it has a cache structure and a memory bus for itself. The more and more powerful cores there are in the processor, the more these resources are dragged on, the more resources are divided between the higher computing power and the more the power of the whole processor is far from a simple mathematical multiple of one core power (at a given frequency ) × number of cores.

IPC Zen 3

According to Angstronomics, AMD Zen 4 decided to deal with PPC rather than PPC in arithmetically difficult tasks (high kernel load, low data traffic) in situations where the performance of existing architectures is limited by data traffic. It would make sense – see the performance gains it gave in the game and server workload Zen 3 V-cache implementation.

This source speaks of an average 7% increase in PPC in a single-core load and an average of 10% increase in PPC in a multi-core load, with the increase in PPC (“IPC”) expected to be significantly higher in data-sensitive situations.

If we believe these values, then the intergenerational shift of single-core power could reach about 22% and the intergenerational shift of multi-core power around 26%. However, it should be noted that although the Angstronomics article tries to pretend to provide a lot of exclusive information, most of it has already appeared in the tweets of some leakers, in discussions on the ChipHell website, etc., so question marks hang over the true origin of this information.

Another source may (or may not) be a statement from Robert Hallock of AMD, which users repeatedly point out in discussions and which some sites see as an official confirmation. However, it must be remembered that Hallock spoke this sentence somehow between speeches in one of the interviews, where journalists tried to push him to make a statement about the “15%”. I first quote the sentence in the original version, so that no one draws conclusions from any nuances arising from the translation:

In translation: “16 cores is the maximum for the AM5 release, but we are achieving 40% + extra performance.” Although some see this as confirming a 40% intergenerational performance shift (more than a 20% increase in IPC), Hallock did not explicitly say Zen 3“Or” against Ryzen 9 5950X “. It can simply be a link to a demo with Blender.

At Computex, Lisa Su demonstrated that Zen 4 it only takes 31% less time to process a job in Blender than Alder Lakein other words (1 / 0.69 =) a 45% higher 16-core performance Zen 4 sorry Alder Lake. It was these “more than forty percent” that Hallock could talk about.

What is certain is that until some solid performance data is released (and AMD has nowhere to rush, there is plenty of time to be released and none of the previous Zen did not publish more performance data in such advance before release), it is not clear which of the data is closer to the truth.

While there was some consensus on Computex on the likely intergenerational performance shift, there is now a user camp that will not allow for a 15% shift compared to Zen 3 as well as a camp that expects nothing less than 40% on top.

–