PHVM_BR

https://www.3dmark.com/sn/13390291 https://www.3dmark.com/spy/63122281

Perhaps because tRCD/tRP = 40 is too tight for 7200MT/s.

A good settings for your memory, I believe without much difficulty, would be 7200 CL36 with: tRCD/tRP = 42 or 44 tRAS = 60 to 84 VDD = 1275 to 1325 VDDQ = 1225 to 1275 You can also set the tREFI to 16-32k. Once you find stability, you can adjust the subtimings by researching and understanding each step. AI can help!

Are all the timings from the XMP profile? Use these timings as a starting point to achieve a higher frequency. Start at 7200MT/s (216), for this frequency with the loose timings of the XMP profile it should be possible with VDD 1200-1250. This is the first time I've seen a VDDQ higher than VDD on these Arrow Lake HX systems. Usually, the VDDQ is set lower, for example, 1250/1200 for VDD/VDDQ. You probably don't need VDDQ = 1225 with VDD = 1150. Then adjust the timings, starting with the primary ones (tCL, tRCD/tRP, tRAS). If you tighten the primary timings too much at 7200 you will need to increase VDD/VDDQ. I would keep VPP at the default (1800). After you achieve a good setup with reasonable voltages to avoid overheating, start adjusting the sub-timings, which is the fine-tuning that makes a big difference in gaming performance.

Is it running within the parameters of the standard XMP profile? 6400MT/s CL38-40-40-80 with VDD/VDDQ at 1350?

I didn't take thermal throttling into account... These undervoltage numbers are perfectly stable for me, but under heavy, longer-duration loads, the clock speeds decrease due to temperatures, and the stable values for 5.2/4.6GHz become unstable depending on the drop. The solution is to reduce the undervoltage or ensure that the system does not suffer from thermal throttling.

My undervolt, stable in any situation, is: -55mV for the P-cores (limited to 5.2GHz) -70mV for the E-cores (4.6GHz) I decided to test the individual undervolt of each P-core and each of the 4 E-core clusters using XTU. I achieved a considerable gain in CPU efficiency with undervolting per core/cluster: P-core 0 = -75mV P-core 1 = -55mV P-core 2 = -55mV P-core 3 = -70mV P-core 4 = -75mV P-core 5 = -55mV P-core 6 = -75mV P-core 7 = -80mV E-cores 0 = -75mV E-cores 1 = -70mV E-cores 2 = -100mV E-cores 3 = -90mV Since XTU does not provide the V/F curve for this CPU, it is still necessary to compensate for the undervoltage at the first two points of the curve (800MHz and 2000MHz) in the BIOS to prevent instability in idle mode. Currently, on CBR23 it runs with all cores locked at 5.2/4.6GHz, without thermal throttling, scoring around 42500 and consuming ~195W.

BIOS Advanced (D2D / NGU) and XTU (NGU / Ring).

To take advantage of that 250W CPU limit, I'm afraid even liquid cooling won't be enough. Perhaps LM + liquid cooling would be possible for short/medium duration loads, but the heatsink has aluminum parts, and caution is needed in choosing the solution for its application...

Try Custom with maximum settings / fan boost and Overboost Extreme.

Time Spy is an older benchmark and doesn't scale well with Core Ultra CPUs. Furthermore, memory makes all the difference in CPU scores, and it seems that with the i9-13900HX you used 6400MT/s modules while with the 275HX you used 5600MT/s. With VBS disabled, stock memory (5600MHz CL46) and stock CPU, I reached up to ~17700. Everything remained the same, but with CPU undervolting and an IETS GT600, the CPU score increased to ~18500. And finally, with memory adjusted to 6400MT/s CL36, I now reach up to ~21900. I've seen scores of ~24k with faster memory. Regarding the GPU, its score with the 4090 is good, slightly below its maximum overclocking potential (up to ~24.5k), but its score with the 5090 is not so good... With a well-tuned overclock, it should be able to exceed 27k and perhaps even 28k with a good cooling pad

My settings were all tested in several tests over a few hours. Everything is 100% stable until then... Regarding the cache, initially I kept it at 3.8GHz with an undervolt of -80mV, but I realized during testing that the undervolt on the cache delivered practically no benefits (maybe 1 to 2W less at full load). Then I tried overclocking in XTU and discovered that for 4.0GHz it was enough to maintain the original voltage, without undervolting. With the ring at 3.8 GHz and everything else the same, I get about ~86 ns. I believe that with Hynix A-die single rank it's possible to achieve 7600MT/s CL36 and latency below 80ns. We'll see... I'm waiting for a reasonable price to buy 2x16gb Fury Impact 6400MT/s CL38 and test it. As for the BIOS update, I haven't updated to the XMG BIOS yet for fear of something going wrong, but I've seen several reports that it works.

I'm currently using the XMG CC, but it's identical to the Avell one. I haven't noticed any differences...

168 = 5600 192 = 6400 204 = 6800 ...

To enable the XMP profile in the BIOS, you need memory modules with XMP profiles. To manually adjust the memory, XMP memory modules are not necessary. Change the profile to manual and adjust the settings. Simply changing the target memory frequency is not effective.