BIOS 2602 – Asus Crosshair X670E Gene
Now we start with the Ryzen 7 9800X3D. This is a normal retail CPU – not an engineering sample or anything like that – and also the silicon quality is also not great. X3D CPUs with one CCD, such as the predecessors 5800X3D or 7800X3D, are known to be an excellent choice for gaming, but until now they were not really overclockable. The 9800X3D changes this and for the first time AMD also allows a static clock to be set for this X3D CPU. This is exactly where I noticed the strange behavior, but one thing at a time.
First of all, the CPU simply runs with its standard settings as far as the core clock is concerned. FCLK and RAM are fixed as described and identical for all subsequent tests. So you would expect the same performance for all tests with the same core clock, right?

First of all, the 9800X3D CPU “stock” is allowed to deliver a performance baseline. The CPU ultimately averages out at a clock rate of approx. 4.9 GHz on all cores, at least according to the effective clock in the HWInfo. No limits such as PPT, TDC, EDC or TJMax are reached and the CPU needs approx. 18.5 seconds for y-cruncher 1b. So far so good.
Now we do the whole thing again statically at 4.9 GHz (1.25 V LLC5). Theoretically, the performance should be more or less identical. The effective clock is also around 4.9 GHz as in the stock test. But suddenly the performance drops to over 21 seconds in y-cruncher 1b. We only changed how the clock is reached – statically, instead of dynamically via the standard boost curve – and as a result the performance drops by 14%. This is completely confusing.

No limits are reached either. The boost limits are already overridden with a static clock and the core temperatures are also completely fine at a maximum of 60 °C. So there is actually no reason why the performance would drop like this.
Perhaps the CPU actually boosted higher than 4.9 GHz in the “stock” test and ran much faster…? a 14% performance difference would have to mean 14% more hidden clock speed, i.e. 5.58 GHz, but be that as it may. How do we test this? We can do it the other way around and cap the boost clock. With the “Fmax” setting (Max CPU Boost Clock Override in the Asus BIOS), the boost clock can be effectively lowered so that the highest possible clock rate is reached by all cores. With an Fmax offset of -1000 (MHz), the CPU then clocks to approx. 4.2 GHz on all cores. We have effectively turned the standard boost into a statically capped all-core underclock. With this setting, the CPU completes the y-cruncher in approx. 19.2 seconds.

Sure, that’s a bit slower than completely stock – we underclocked the CPU after all – but that’s only about 3.5% performance loss. Conversely, the CPU is now a good 10% faster at 4.2 GHz with Fmax than with truly static OC at 4.9 GHz! How can a static 4.9 GHz on all cores be 10% slower than 4.2 GHz in boost on all cores?
Now we can have some fun and use ECLK2 – the base clock that only controls the CPU cores – to boost the clock from the previous config so that we get back to 4.9 GHz and have a 1:1 comparison of the dynamic clock at 4.9 GHz and the static clock at 4.9 GHz. Of course, we now have to check the voltage manually, as the boost algorithm only requests voltage for 4.2 GHz. 1.25V at LLC5 as with the static OC is now present.

And lo and behold, we are back at approx. 18.5 seconds in the y-cruncher. It is therefore not due to a hidden boost of individual cores that the CPU is so much faster with dynamically clocked cores than with statically clocked cores.
To summarize the tests up to this point: As soon as the CPU is statically overclocked, i.e. “All Core Clock” is set in the BIOS, approx. 14% performance is lost in the y-cruncher. That’s massive, but how come?
- 1 - Test setup and Ryzen 9 7950X as baseline
- 2 - BIOS 2602 – Asus Crosshair X670E Gene
- 3 - Clock Streching or AVX throttling and workarounds (LN2 mode)
- 4 - The "Fix" BIOS 0029 – Asus Crosshair X670E Gene
- 5 - ASRock B650I Lightning WiFi for comparison
- 6 - BIOSes 2604 and 2506 – Asus Crosshair X670E Gene
- 7 - X3D-Cache confuses y-cruncher (developer)
- 8 - Cinebench R23 with "AVX-Light" as reference
- 9 - Summary and many unanswered questions
19 Antworten
Kommentar
Lade neue Kommentare
Urgestein
Urgestein
Urgestein
Mitglied
Veteran
Veteran
Urgestein
Urgestein
Mitglied
Urgestein
Veteran
Urgestein
Urgestein
Veteran
Urgestein
Urgestein
Veteran
Mitglied
Veteran
Alle Kommentare lesen unter igor´sLAB Community →