We decided against using gaming benchmarks today because data like 1% lows or frame time variances cannot be reliably extrapolated with the described approach. Of course, it’s still safe to assume that in a CPU- and RAM-limited gaming scenario, such as Cyberpunk 2077 at 1080p, there would be a comparable performance gain as in the synthetic benchmarks.
So now let’s look at the results. Colored light pink are the bars of the 4750G CPU (Z2 + new IO), dark pink are the bars of the 5950X CPU (Z3), and shaded from both colors is the theoretical hybrid of both, based on Zen 3 compute chiplets and a new IO chiplet (Z3 + new IO).
In the AIDA64 read test, the gains due to the higher clock rates made possible by the new IO die are immediately noticeable. The memory controller of the new IO-Die can operate up to 2267 MHz and thus DDR4-4533 in 1:1 mode. Furthermore, the gains in 1:2 mode become smaller, but even DDR4-5200 can still be operated stably for daily use.
In the write test, we again see large theoretical gains when running the higher clock rates in 1:1 mode, with almost no performance gain at all in 2:1 mode. Also noticeable here is that the Zen 2 CPU with new IO die is generally faster than the Zen 3 CPU at the same clock speeds. In addition to an actually more efficient memory controller, the monolithic design or tighter auto-timings could also are potential causes here.
The copy test shows similar gains, with the exception of the DDR4-5200 setting in 2:1 mode, which actually loses performance. The 1:1 operation also proves itself here, with which even more than 60000 MB/s are achieved with DDR4-4533.
In the latency test, we clearly see a general disadvantage with the Zen 2 CPU due to the older architecture of the compute chips. So here the newer Zen 3 CPU already has a significant advantage and if you combine this with the higher possible clock rates of the new IO-Die, you get impressively low latencies of just over 50 ns for Ryzen. Again, the exception is the DDR4-5200 clock rate in 2:1 operation, which provides for much higher latencies.
If we look at Geekbench 3 Multi-Core Memory Performance as the last of our synthetic memory benchmarks, we can also consider this as a general indicator of the combined performance of CPU and RAM. The first thing that catches the eye here is the generation leap of Zen 3 compared to Zen 2, mainly due to the 4 times larger cache with 32 MB. Effectively, Zen 3 can handle many instructions exclusively in the cache, so data streams need to pass through the comparatively slow RAM less frequently, which saves time and thus improves performance. Nevertheless, the new IO chiplet can contribute an additional 4-7% in performance gain with Zen 3. Tomorrow’s high-end config with a theoretical “Ryzen 9 6950X” and dual-rank DDR4-4533 in 1:1 operation could actually scratch the 10000 mark, mind you stable for daily use and with relatively simple XMP timings.
Conclusion
Of course, the comparison of 4750G and 5950X CPU is not perfect. Besides the different design of the two CPUs, there are also other variables, such as different AGESA parts per CPU generation in the same BIOS, non-transparent “Auto” BIOS settings, etc., that could falsify the performance predictions. Still, the CPUs almost couldn’t be more similar, so it’s rarely been so supposedly easy to infer the performance of an upcoming generation.
AMD recently made a big leap in performance with Zen 3 and Ryzen 5000, outshining the competition with their compute performance and efficiency. Accordingly, it would stand to reason that the compute chiplets would be mostly retained for the next generation, with the now somewhat aging IO die getting an upgrade instead. So what such a “Zen 3+” CPU might look like is not that far-fetched.
As always, all that’s left for us to do is wait, drink tea and eat speculative biscuits until we really know what Ryzen 6000 and a potential “Zen 3+” architecture will look like. Still, sometimes it’s worth looking backwards or sideways to infer what’s ahead. And if we are honest, such theoretical speculations are also a bit of fun for the community and for us.
Working with the B550 Unify-X motherboard to collect performance data was very enjoyable. Rarely have CPU and RAM on a Ryzen platform been so effortlessly tuned to such high clock speeds and tight timings as tested in today’s article. The only thing that would have been desirable is a “Safe Boot” button to avoid having to reset the entire BIOS after some settings didn’t work. Thanks again to MSI for providing the B550 Unify-X.
The motherboard was kindly provided to us by MSI, without any oblication for publication. Furthermore no influence was excercised over the tests and their results.
24 Antworten
Kommentar
Lade neue Kommentare
Mitglied
Urgestein
Mitglied
Urgestein
Urgestein
Urgestein
Urgestein
Urgestein
Urgestein
Veteran
Urgestein
Urgestein
Neuling
Urgestein
Urgestein
Urgestein
Urgestein
Mitglied
Alle Kommentare lesen unter igor´sLAB Community →