Has anyone tuned the memory to achieve throughput over 300 Gbps on 8 channels? I've seen results close to 380 Gb/s, but so far I've only managed to get a stable 303 Gb/s at 6800 Mhz. Perhaps this is influenced by the number of cores; I don't have many, only 20.
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
right.
AIDA64 cache and mem. benchmark use all cores and adds each result up to one,
you don't know how much bendwidth one thread gets. if you tweak to max. multi bandwidth you may have bad per core bandwidth.
The memory read speed with AIDA64 on SPR-WS is especially tied to core count, the write speed is mostly independent considering the core count floor is 16 on the overclockable 8 channel chips.
All that being said AIDA64 isn't exactly the best benchmark to show a platform's true memory performance; for example I'm only getting 1414.7GB/s max memory read speed under AIDA64 on my fastest system, but MLC is showing 1834.0GB/s read speed if I enable AVX512 to be used for the test.
Last edited:
Thx for the clarification about core QTY both of you 
@twin_savage
I see you've been working on this topic for a while, regarding memory settings, could you share your timings? Also, specify which voltages need to be increased to increase the memory frequency and decrease the timings, as the ASUS BIOS has specific names for them that differ from the usual ones.
And a bonus question: does anyone know if utilities that display memory timings?, such as Asrock TimingConfigurator and Asus MemTweakIt, work in Windows 10/11? I've tried the latest versions of these utilities, and they either don't work or provide incorrect information. Are there any other versions that do?
@twin_savage
I see you've been working on this topic for a while, regarding memory settings, could you share your timings? Also, specify which voltages need to be increased to increase the memory frequency and decrease the timings, as the ASUS BIOS has specific names for them that differ from the usual ones.
And a bonus question: does anyone know if utilities that display memory timings?, such as Asrock TimingConfigurator and Asus MemTweakIt, work in Windows 10/11? I've tried the latest versions of these utilities, and they either don't work or provide incorrect information. Are there any other versions that do?
let me prefaces this by saying I don't think I'm particularly good at overclocking. Also my memory was hard-to-OC 64GB dual rank x4 organized sticks:
The only memory specific voltages I touched were VDD, VDDQ and "VCCD HV 1.14v IN /MC Voltage", all running at 1.35v.
The easiest way to get write speed up was adjusting the bclk since the mesh caps out at 2700MHz normally.
The really inflated memory performance numbers I mentioned in my last post were for a sapphire rapids HBM machine which I didn't think you'd be that interested in since it isn't applicable to W790, but here it is anyways:
The only memory specific voltages I touched were VDD, VDDQ and "VCCD HV 1.14v IN /MC Voltage", all running at 1.35v.
The easiest way to get write speed up was adjusting the bclk since the mesh caps out at 2700MHz normally.
I haven't found any good ones I'd recommend.
The really inflated memory performance numbers I mentioned in my last post were for a sapphire rapids HBM machine which I didn't think you'd be that interested in since it isn't applicable to W790, but here it is anyways:
I noticed that when the processor is idle, its consumption is at 150 watts, which is clearly too much for 3535X. The BIOS was configured to default settings. I tested this on two freshly installed Windows 10 systems, with no unnecessary background processes running. I encountered a similar situation on an LGA3647, and by changing the BIOS settings, I reduced the consumption to 50 watts. Could you tell me which settings might be responsible?
50W is very good, you can't have less thought.
reduce mesh clock gives low performance, imagen you have 4 tiles and 10 EMIB drawing power.
what i use is:
(BIOS defaults)
Code:
use high perf.power plan.
but min.processor state = 25% (default is 100%)
cmd window
- to unhide Processor idle demote threshold
powercfg -attributes SUB_PROCESSOR 4b92d758-5a24-4851-a470-815d78aee119 -ATTRIB_HIDE
- to unhide Processor idle promote threshold
powercfg -attributes SUB_PROCESSOR 7b224883-b3cc-4d79-819f-8374152cbe7c -ATTRIB_HIDE
now you have 2 more options in power options "processor power management"
default is 40/60
set to 95/98@RolloZ170
But 100W at complete idle still bothers me, especially when the 3175X has 8 more cores, and its consumption is half that. I'll try to find the necessary settings...
@All
In general, what is the minimum power consumption of a processor, as reported by owners of these processors in this thread?
Yah, but it was about my LGA3647 3175X (I will attach a screenshot from hwinfo below).
OMG! I didn't even realize the Mesh consumes so much! I lowered it from x27 to Auto and got -55W at idle. (By the way, the CMD commands didn't change anything, consumption didn't change, so the cache was the reason.)
But 100W at complete idle still bothers me, especially when the 3175X has 8 more cores, and its consumption is half that. I'll try to find the necessary settings...
@All
In general, what is the minimum power consumption of a processor, as reported by owners of these processors in this thread?
Attachments
it changes responsibility too, anti lag.
my w5-3425 wanted min. 76W package power.
you can have less with 24xx/25xx models (single chip, no EMIBs)
my w5-3435x will do ~82 watts as reported by hwinfo64 at idle in Windows if and only if i set the power plan to "Power saver"
Hi everyone in 2026, what's your temperature difference between cores under full CPU load? I have 20 cores, and one core has a 20°C temperature difference. I tried changing the thermal paste and cooling system, but this did not change the situation, apparently this is a feature or this core (14-th) is not located in a very good place.
what CPU exactly or a riddle ?
can be bad soldering, if your CPU is ES this is not sorted out at test.
if not soldered CPU (maybe who knows) dried out TIM.
bad soldering, Intel should replace it.
edit:
3535x is XCC 4 tiles each 5 cores model, 4 cores at almost same worse place.
Hi everyone!
I'm building a deep learning rig on the ASUS Pro WS W790E-SAGE SE and I have a question regarding the dual PSU setup.
I have two different PSUs:
Super Flower Leadex III 1300W (Main: for MB/CPU/4080)
Fractal Design ion 850W unit (Secondary: purely for 3x MI50 accelerator cards via PCIe cables)
The manual explicitly states: "ensure that both PSUs need to be of the same brand and model".
My understanding is that connecting different PSUs to ATX_PWR(1) and ATX_PWR(2) might bridge the 12V/5V rails, causing cross-loading issues if voltages aren't identical.
The question is:
Does anyone know if the board isolates the power rails between the two ATX headers and only syncs the PS_ON signal? Or does it electrically merge them?
Is it safe to plug both mismatched PSUs into the motherboard headers, or should I ignore ATX_PWR(2) and use an external Add2PSU adapter to sync them safely?
I'm building a deep learning rig on the ASUS Pro WS W790E-SAGE SE and I have a question regarding the dual PSU setup.
I have two different PSUs:
Super Flower Leadex III 1300W (Main: for MB/CPU/4080)
Fractal Design ion 850W unit (Secondary: purely for 3x MI50 accelerator cards via PCIe cables)
The manual explicitly states: "ensure that both PSUs need to be of the same brand and model".
My understanding is that connecting different PSUs to ATX_PWR(1) and ATX_PWR(2) might bridge the 12V/5V rails, causing cross-loading issues if voltages aren't identical.
The question is:
Does anyone know if the board isolates the power rails between the two ATX headers and only syncs the PS_ON signal? Or does it electrically merge them?
Is it safe to plug both mismatched PSUs into the motherboard headers, or should I ignore ATX_PWR(2) and use an external Add2PSU adapter to sync them safely?
Sadly I don't know if anyone can answer that. Why not spend the $170 and buy a matching Leadex III?
When i had the W790E-SAGE SE for a few weeks (in the return period) i have tried with 2 non-matching psu, it sort of worked but i had start up glitches, that the security fuse of the apartment blew out when i started the computer about 20% of computer start ups. I returned one of the psus it was a amazon returns Corsair. and then decided to run the set up QYFS +2xRTX 3090 and A4000 Ada throttling their power with a single 1300W psu, and reducing the cpu short time power limit so 400W cpu+2x270W rtx+60W a4000 +rest. That worked fine. I am not sure how you will cool your MI50s but reducing power limits might be a good idea any way. For long compute runs those last 20% power is not worth everything for me. In a hurry on short runs i temporary increase power sometimes. When you do deep learning, it is rare to have full load on the cpu and gpu simultaneously. My cpu always maxes out around 60% when the gpu are doing the work. Long story short why not try and set max 400W cpu 3x250W gpu and use just one psu?
Last edited: