Why there haven’t been any RAM tests from me/us for a while, why we won’t have any gaming benchmarks in our RAM tests and how we want to test RAM kits for their real differences in the future, you’ll learn in the following piece. In addition, this should serve as a basic article, which will be linked accordingly in upcoming reviews and kept up to date.
No more game benchmarks!
A (relatively) new CPU generation – a new benchmark collection for DDR5 tests. Because every game reacts differently to RAM settings and we already had a big time commitment with the few game tests in our previous RAM kit tests, we will henceforth do without game benchmarks in every test of memory kits. Instead, we will limit ourselves to a few, meaningful, synthetic benchmarks.
We will only solve the transfer task of inferring a game title from the synthetic benchmarks with 5 titles today as an example. If necessary, others might be added in the future, but here, of course, the community is also in demand. Because even if we test 50 games, the 51st will behave differently again and the 49th won’t even be relevant tomorrow. Thus the only remaining option is the abstraction via synthetic benchmarks.
A particularly good result in latency-sensitive benchmarks like PyPrime 2.0 then suggests high performance in latency-sensitive titles like CSGO or Assetto Corsa Competitizone, for example. Conversely, if a particularly good result is achieved in bandwidth-sensitive benchmarks, so inclined games will also profit here. Thus, we can focus on the real differences between RAM kits in our tests and increase the cadence of the tests at the same time. But I’m getting ahead of myself – how do we actually test?
“XMP/JEDEC” vs. “OC”
For better classification, we will present the benchmark results separately in two categories:
- XMP/JEDEC – with the performance as the RAM kit was designed by the manufacturer
- OC – with maximum performance that can be achieved by manual tuning and the potential of the particular kit
Thus, we can continue to serve several target groups of readers – on the one hand, those who simply run their RAM kit according to the specification, and on the other hand, the hardcore tinkerers who are always looking for the last FPS. In addition, we achieve fairness when comparing different kits, since they are officially designed by the manufacturers only for operation in XMP mode anyway, and on the other hand, because the silicon lottery still has a certain influence when overclocking a kit.
Test hardware and methodology
An Intel Core i9-13900KF CPU with an Asus Maximus Z790 Hero motherboard with BIOS 0813 serves as the test platform. The CPU is statically overclocked to 5.5 GHz on the P-cores and 5.0 GHz on the cache. The E-cores also remain active, since the scheduler problems in current Windows 11 versions have been fixed, and run at 4.5 GHz. XMP profiles are enabled with the “XMP II” option in the BIOS, as this also loads subtimings such as tRFC2 and tWR as specified by the kit manufacturer.
Our Windows 11 image is a commercially available Pro 64 bit with version 22H2. Windows Defender Anti-Virus was deactivated, since it accesses a relatively large amount of RAM quite frequently and thus affects the reproducibility of the results. The Nvidia driver used is version 528.24 and MSI Afterburner was used to raise the performance and temperature limit of our MSI Suprim X RTX 4090 graphics card to the maximum. The performance data in the games is recorded with Nvidia Frameview, based on the open-source software Presentmon. By the way, Resizeable Bar is enabled by default in the BIOS and driver; whether it is used is up to the respective game title or associated driver profile.
Of course, we have not forgotten the AMD Ryzen user base, but as a test platform for fast kits and new dies, the current CPUs with a RAM clock limit at 6400 Mbps are rather unsuitable. The test results up to this clock limit can be transferred 1:1 for the most part and the dies and ICs in AMD EXPO kits are the same as in Intel XMP. So we take the liberty of completely skipping the tests on AMD Ryzen (at least currently) and simply refer to the results with the corresponding clock rate instead.
Reference Configs
DDR4 will remain with us as a reference config even with Raptor Lake, on the one hand with a conservative XMP mode that should work with most motherboards and CPUs, but also with a higher overclocking in Gear 1, which is only possible since the improved IMC of Raptor Lake CPUs.
We also drill down the comparison configs for DDR5. Samsung 16 Gbit B-die, Hynix 16 Gbit A-die, M-Die and Micron 16 Gbit RevA are represented here in order to be able to map real reference points for the entire DDR5 market at this point in time. More details about the individual kits and configurations can be found on the next pages.
Game Settings
When it comes to the graphics settings of our few gaming tests, we will continue to not sacrifice resolution or visual quality just to artificially put more load on the CPU and RAM. If you use AAA games at 720p with low details as RAM benchmarks, you might as well use synthetic benchmarks, the realism is similarly low.
Instead, we use the maximum graphics settings in our tests, with maximum ray-tracing settings if available and supported, but without DLSS. Because even in the latest version, DLSS still causes graphical artifacts that are a no-go for many enthusiasts and gamers. The resolutions are still 2560 x 1440 (1440p) and 1920 x 1080 (1080p). The CPU and RAM limits in our tests are therefore not artificially forced and should be very representative for actual gaming with high-end hardware in 2023. So if you’re still looking for more FPS with an i9-13900K and an RTX 4090 GPU, this should be a good reference point for RAM tuning.
- 1 - Introduction and concept
- 2 - DDR4 Samsung 8 Gbit B Die (4S8B)
- 3 - DDR5 SK Hynix 16 Gbit A Die (5H16A)
- 4 - DDR5 SK Hynix 16 Gbit M Die (5H16M)
- 5 - DDR5 Samsung 16 Gbit B Die (5S16B)
- 6 - DDR5 Micron 16 Gbit Rev A (5M16A)
- 7 - Synthetics (1/2) – PyPrime 2.0 2b, y-cruncher 2.5b
- 8 - Synthetics (2/2) – Geekbench 3, AIDA64
- 9 - Gaming (1/3) – Assetto Corsa Competizione, Marvel's Spider-Man Remastered
- 10 - Gaming (2/3) – Dead Space (2023), Cyberpunk 2077
- 11 - Gaming (3/3) – Shadow of the Tomb Raider
- 12 - Summary and conclusion
27 Antworten
Kommentar
Lade neue Kommentare
Urgestein
Urgestein
Veteran
1
Veteran
Urgestein
Urgestein
Urgestein
Veteran
Veteran
Veteran
Mitglied
Mitglied
Neuling
Veteran
Urgestein
Neuling
Neuling
Alle Kommentare lesen unter igor´sLAB Community →