AMD Ryzen 3000 Undervolting Offset vs. Override Vcore Voltage
video description
Date: 2020-05-06
Related videos
Comments and reviews: 10
Rory
Hold on. I tested this and benched. I explained how and why in the video. The boost and core variance in the algorithms, works in a fashion. Now the temp and voltage are microcoded, if the temp is lower the clock will rise. Buy reducing voltage at the core, this lower temps and boosts are higher and bench results. I did not push as low as 1v. I did lock voltage at 1. 225 and 1. 32 and 1. 35 as well as 1. 37. But I let the other compensation run to allow compensation and also not. On total lock, there was no boost. (As you would expect) however due to science or ohm's law, by adding current, you could get a higher speed with lower temps. On the video I did this is shown at 0: 32 showing a boost clock of 5. 1ghz There should be some bench results in the video. When I did the video it was more for the lower temps and expected performance. It just so happened to jump up more than expected. Testing was performed over a week and serveral benches were ran to check the results. However. Due to the performance variable being as it's software and algorithmic, the results varied constantly. CB R15 varied by 200 points on the same clock speeds and settings. In monitoring all bench and performance results, all hardware with energy saving and boost stepping would not consistently boost or peak clocks at the same times. This resulted in a significant difference in results, by up to 4/700 points. The only thing that actually remained a constant was temperatures. (Actually was was required. However as you have said, all chips and configurations are different and you do get silicon lottery. Therefore one processors result can and do differ significantly in some cases. Therefore this result may not transfer to all CPUs or motherboards. However a key point is that the voltages on the processors is too high and is not required, this is not saying setting a low core is the right solution. But the processors need to run cooler. And it appears the stock software does not manage voltage well at all. But this maybe a glitch in mine. Hope that puts some light on the subject.
reply
Hold on. I tested this and benched. I explained how and why in the video. The boost and core variance in the algorithms, works in a fashion. Now the temp and voltage are microcoded, if the temp is lower the clock will rise. Buy reducing voltage at the core, this lower temps and boosts are higher and bench results. I did not push as low as 1v. I did lock voltage at 1. 225 and 1. 32 and 1. 35 as well as 1. 37. But I let the other compensation run to allow compensation and also not. On total lock, there was no boost. (As you would expect) however due to science or ohm's law, by adding current, you could get a higher speed with lower temps. On the video I did this is shown at 0: 32 showing a boost clock of 5. 1ghz There should be some bench results in the video. When I did the video it was more for the lower temps and expected performance. It just so happened to jump up more than expected. Testing was performed over a week and serveral benches were ran to check the results. However. Due to the performance variable being as it's software and algorithmic, the results varied constantly. CB R15 varied by 200 points on the same clock speeds and settings. In monitoring all bench and performance results, all hardware with energy saving and boost stepping would not consistently boost or peak clocks at the same times. This resulted in a significant difference in results, by up to 4/700 points. The only thing that actually remained a constant was temperatures. (Actually was was required. However as you have said, all chips and configurations are different and you do get silicon lottery. Therefore one processors result can and do differ significantly in some cases. Therefore this result may not transfer to all CPUs or motherboards. However a key point is that the voltages on the processors is too high and is not required, this is not saying setting a low core is the right solution. But the processors need to run cooler. And it appears the stock software does not manage voltage well at all. But this maybe a glitch in mine. Hope that puts some light on the subject.
reply
Today
So decreasing it down to 1 volt lowered heat output rather dramatically, and also lowered performance as well. Though, did power efficiency improve or worsen? Taking the CB Mark for CineBench R20 nT score difference, where the 1 volt undervolted one performed 80. 8-82. 5 % of stock. And the corresponding graph for the temperature Tdie, where we can see a reduction from about 60 C down to about 30-40 C, and if we say that the ambient temperature were a constant 25C, then we could say that the power dissipation reduced to only about 30%, though extremely rough calculation based on the idea that a heatsink has a linear delta-T/watt figure through its operational range, this would mean that the CPU's power efficiency is about 2. 7 times higher. (I would say that seems unrealistic) So an actual power consumption comparison plot would have been far more interesting, then the Tdie plot in my opinion. (Since Tdie will greatly vary with what heat sink one has, the power consumption would mostly be effected by clock speed and feature utilization on the die itself. (leakage currents are though temperature dependent) In the end, how much did the power efficiency actually improve? If it improved at all. Could be useful to know if running these in servers or the like.
reply
So decreasing it down to 1 volt lowered heat output rather dramatically, and also lowered performance as well. Though, did power efficiency improve or worsen? Taking the CB Mark for CineBench R20 nT score difference, where the 1 volt undervolted one performed 80. 8-82. 5 % of stock. And the corresponding graph for the temperature Tdie, where we can see a reduction from about 60 C down to about 30-40 C, and if we say that the ambient temperature were a constant 25C, then we could say that the power dissipation reduced to only about 30%, though extremely rough calculation based on the idea that a heatsink has a linear delta-T/watt figure through its operational range, this would mean that the CPU's power efficiency is about 2. 7 times higher. (I would say that seems unrealistic) So an actual power consumption comparison plot would have been far more interesting, then the Tdie plot in my opinion. (Since Tdie will greatly vary with what heat sink one has, the power consumption would mostly be effected by clock speed and feature utilization on the die itself. (leakage currents are though temperature dependent) In the end, how much did the power efficiency actually improve? If it improved at all. Could be useful to know if running these in servers or the like.
reply
Tech
I have a request: could you do some tests on underclocking? This sounds strange but there is an actuall good reason to know how low the power consumption can go before failing. The explanation for this request is simple but a long story so i try to make it as short as possible. Companies using the Ryzen zen 2 generation can now safe alot of money on cpu, motherboard and graphics card. Because the technology used allowes for turning on/off seperate cores or 1 or more ccd(s, companies can now regulate workloads with one cpu over multiple users at the same time. For example: 1 person works on administrative task do not need more then 1 core assigned to them at a low clockspeed while another person can use the same cpu at the same time using other core(s)/ccd(s) at different speeds via 1 graphics card. You can now conclude all the other combinations possible by yourself i hope(this is what makes this explanation short; ) ) This means that instead of buying multiple intel systems (very very expensive) they can now invest in just 1 system for a multi user envirnoment for a much cheaper price. (so yes i definatly think that once companies realizes the opportunity before them it will start rolling) I hope you get the point to this and perhaps honor my request.
reply
I have a request: could you do some tests on underclocking? This sounds strange but there is an actuall good reason to know how low the power consumption can go before failing. The explanation for this request is simple but a long story so i try to make it as short as possible. Companies using the Ryzen zen 2 generation can now safe alot of money on cpu, motherboard and graphics card. Because the technology used allowes for turning on/off seperate cores or 1 or more ccd(s, companies can now regulate workloads with one cpu over multiple users at the same time. For example: 1 person works on administrative task do not need more then 1 core assigned to them at a low clockspeed while another person can use the same cpu at the same time using other core(s)/ccd(s) at different speeds via 1 graphics card. You can now conclude all the other combinations possible by yourself i hope(this is what makes this explanation short; ) ) This means that instead of buying multiple intel systems (very very expensive) they can now invest in just 1 system for a multi user envirnoment for a much cheaper price. (so yes i definatly think that once companies realizes the opportunity before them it will start rolling) I hope you get the point to this and perhaps honor my request.
reply
Vicky
some intresting behaviour I have noticed on the x570 creation is that auto settings are almost always better, just running as much set to auto seems to yeald the best performance in my testing, trying to tighten up the memory from 3600 CL16 b-die to a 3600CL15 with much tigher timings across the board, the memory reported to be doing better then stock xmp when it was benched but cpu performance was slightly worse then just set XMP, messing around with PBO and the other oc features trying to set them was just worse then auto and the best OC settings was just set volts, set XMP, set multiplyer done this yealds the best perfomance. if MSI gives me Bclk to playwith on the creation I may be able to squeeze a little bit of performance, the all core wall of my chip being some where between 4. 475ghz and 4. 5ghz, 4. 5ghz seems to need more then 1. 4v and thats more then I am willing to give expecially because its pritty much good to go at 1. 375v and 1. 385v for ensured stablity at 4. 475ghz but yes what ever its setting for auto settings seems to be whats best, somthing rather supprising considering on previous boards thats almost never the case XD
reply
some intresting behaviour I have noticed on the x570 creation is that auto settings are almost always better, just running as much set to auto seems to yeald the best performance in my testing, trying to tighten up the memory from 3600 CL16 b-die to a 3600CL15 with much tigher timings across the board, the memory reported to be doing better then stock xmp when it was benched but cpu performance was slightly worse then just set XMP, messing around with PBO and the other oc features trying to set them was just worse then auto and the best OC settings was just set volts, set XMP, set multiplyer done this yealds the best perfomance. if MSI gives me Bclk to playwith on the creation I may be able to squeeze a little bit of performance, the all core wall of my chip being some where between 4. 475ghz and 4. 5ghz, 4. 5ghz seems to need more then 1. 4v and thats more then I am willing to give expecially because its pritty much good to go at 1. 375v and 1. 385v for ensured stablity at 4. 475ghz but yes what ever its setting for auto settings seems to be whats best, somthing rather supprising considering on previous boards thats almost never the case XD
reply
BulkChomp
The high voltage scare is almost the exact same thing we had when Ryzen 1000 launched. People were noticing the momentary 1. 5 voltage spikes created by the precision boost function, and then you had AMD representatives assuring people it's functioning as intended. With all the thousands of blissfully ignorant people running those things at stock, I've yet to see any news of masses of 1st gen Ryzens bought in 2017 dying. I guess we'll see in few more years if the engineers were wrong in making them behave like that. As for Ryzen 3000, the chips produced on TSMC's 7nm are really small with densely packed transistors, which invariably leads to much higher heat density and therefore higher operating temperatures. If you're really worried about high core temperatures, then you can simply manually lower the PPT limit in BIOS. This will degrade performance, but unlike Vcore offsets (which also lowers performance in most cases, it won't have any unintended negative side effects on the CPU's VID values, which means it's safer overall. If you're not having any issues then I think it's best to just let it do its thing.
reply
The high voltage scare is almost the exact same thing we had when Ryzen 1000 launched. People were noticing the momentary 1. 5 voltage spikes created by the precision boost function, and then you had AMD representatives assuring people it's functioning as intended. With all the thousands of blissfully ignorant people running those things at stock, I've yet to see any news of masses of 1st gen Ryzens bought in 2017 dying. I guess we'll see in few more years if the engineers were wrong in making them behave like that. As for Ryzen 3000, the chips produced on TSMC's 7nm are really small with densely packed transistors, which invariably leads to much higher heat density and therefore higher operating temperatures. If you're really worried about high core temperatures, then you can simply manually lower the PPT limit in BIOS. This will degrade performance, but unlike Vcore offsets (which also lowers performance in most cases, it won't have any unintended negative side effects on the CPU's VID values, which means it's safer overall. If you're not having any issues then I think it's best to just let it do its thing.
reply
Krullexneo
Been playing with my 3700x since getting it, I've found 2 things that have helped performance, thermals and voltage with my ASUS X470 PRO. Simply setting the vcore to an offset of -100mv kept things quieter, cooler and performance was almost identical, another way was I had it running at 4. 2Ghz 1. 352v, I've noticed at stock it will run at 1. 331-1. 362v during an all core 4. 2-4. 3 boost, I've never seen it actually hit 4. 4 yet, I just can't stand those voltage spikes while idle which cause the fans to ramp up like crazy for no reason, even with a very strict fan curve so I think I will be keeping it running at 4. 2 1. 352v, I know there was a post saying Ryzen 3000 series shouldn't go further than 1. 325v for all core but I honestly don't believe that at all seeing as it is doing higher than that at stock when all core boosting, also cores can still sleep and slow down even with a manual overclock which is very interesting, even though HWiNFO doesn't display that, Ryzen Master does show lower core clocks and sleeping cores even with a manual overclock set
reply
Been playing with my 3700x since getting it, I've found 2 things that have helped performance, thermals and voltage with my ASUS X470 PRO. Simply setting the vcore to an offset of -100mv kept things quieter, cooler and performance was almost identical, another way was I had it running at 4. 2Ghz 1. 352v, I've noticed at stock it will run at 1. 331-1. 362v during an all core 4. 2-4. 3 boost, I've never seen it actually hit 4. 4 yet, I just can't stand those voltage spikes while idle which cause the fans to ramp up like crazy for no reason, even with a very strict fan curve so I think I will be keeping it running at 4. 2 1. 352v, I know there was a post saying Ryzen 3000 series shouldn't go further than 1. 325v for all core but I honestly don't believe that at all seeing as it is doing higher than that at stock when all core boosting, also cores can still sleep and slow down even with a manual overclock which is very interesting, even though HWiNFO doesn't display that, Ryzen Master does show lower core clocks and sleeping cores even with a manual overclock set
reply
Optimatr
Best balance I've been able to find for my 3700x is setting a negative offset of 0. 09 while also enabling PBO with 200mhz. As I understand it, the PBO setting kind of counters the negative offset, but not the extent that it ever really over volts like it did stock. My out of the box vcore was damn near 1. 48 under load on an Asus x470 gaming-F, and temps were in the mid 80's in cinebench r20. Now my temps don't exceed 76 in the same test, and my vcore under load hovers between 1. 34-1. 39. My cinebench score actually slightly improved, and my multi core results are consistently between 4, 865-4, 900+, while my single core is 495-499. I think this slight improvement in results and thermals mirrors your results on your 3900x with negative 0. 1 offset. Idle temps also dropped from mid 40's and 50's to low to mid 30's now. Gaming temps in really cpu demanding games like BFV are in the low 60's, and most games stay in the 50s.
reply
Best balance I've been able to find for my 3700x is setting a negative offset of 0. 09 while also enabling PBO with 200mhz. As I understand it, the PBO setting kind of counters the negative offset, but not the extent that it ever really over volts like it did stock. My out of the box vcore was damn near 1. 48 under load on an Asus x470 gaming-F, and temps were in the mid 80's in cinebench r20. Now my temps don't exceed 76 in the same test, and my vcore under load hovers between 1. 34-1. 39. My cinebench score actually slightly improved, and my multi core results are consistently between 4, 865-4, 900+, while my single core is 495-499. I think this slight improvement in results and thermals mirrors your results on your 3900x with negative 0. 1 offset. Idle temps also dropped from mid 40's and 50's to low to mid 30's now. Gaming temps in really cpu demanding games like BFV are in the low 60's, and most games stay in the 50s.
reply
liren
Machine: Msi B450x Mortar Max, R5 3600, AIO CPU cooler. readings all from ryzen master. [ ryzen master manual mode, all connected 4. 2GHz 1. 1v, CB Marks 3582, FSB 1866, cpu temp max around 58c, cpu power around 45w ] [ ryzen master manual mode, all connected 4. 3GHz 1. 15v, CB Marks 3666, FSB 1866, cpu temp max around 65c, cpu power around 50w ] [ ryzen master PBO mode, voltage excluded, CB Marks 3312, FSB 1866, cpu temp max around 74c, cpu power around 78w ] (cpu frequency goes around 3. 9GHz automatically) i'm really not good at playing with ryzen, i must have something set wrong since with all bios settings at default, one click PBO setting always give around only 3. 9GHz and temp goes through up the roof. but now manually set a fixed cpu frequency and a lower voltage still gives better results.
reply
Machine: Msi B450x Mortar Max, R5 3600, AIO CPU cooler. readings all from ryzen master. [ ryzen master manual mode, all connected 4. 2GHz 1. 1v, CB Marks 3582, FSB 1866, cpu temp max around 58c, cpu power around 45w ] [ ryzen master manual mode, all connected 4. 3GHz 1. 15v, CB Marks 3666, FSB 1866, cpu temp max around 65c, cpu power around 50w ] [ ryzen master PBO mode, voltage excluded, CB Marks 3312, FSB 1866, cpu temp max around 74c, cpu power around 78w ] (cpu frequency goes around 3. 9GHz automatically) i'm really not good at playing with ryzen, i must have something set wrong since with all bios settings at default, one click PBO setting always give around only 3. 9GHz and temp goes through up the roof. but now manually set a fixed cpu frequency and a lower voltage still gives better results.
reply
robert
Depends on what you want outta your system. The bragging rights of working next to a jet engine ( 140W cpu power) or just as good in performance and saves on power bills each month ( 100W cpu power? my 1800x is undervolted to 1. 16v and below. I pick the setting based on Cinebench score after hours of testing, a spreadsheet to keep track of various P-state voltages & Cinebench #s. I used to let mine go up to 1. 37-1. 39 v (with P-state overclocking. the fans will go into banshee mode constantly (3x 120mm Industrial Noctua, 2000 rpm max, those fans will get loud close to max rpm. Now with the undervolting of 1800x, the fans are nearly silent and in everyday use, I can't tell the performance differences! Oh, and my hearing is back too. Noctua D15 + 140mm Noctua PWM fan Asus Crosshair VI Hero Wifi
reply
Depends on what you want outta your system. The bragging rights of working next to a jet engine ( 140W cpu power) or just as good in performance and saves on power bills each month ( 100W cpu power? my 1800x is undervolted to 1. 16v and below. I pick the setting based on Cinebench score after hours of testing, a spreadsheet to keep track of various P-state voltages & Cinebench #s. I used to let mine go up to 1. 37-1. 39 v (with P-state overclocking. the fans will go into banshee mode constantly (3x 120mm Industrial Noctua, 2000 rpm max, those fans will get loud close to max rpm. Now with the undervolting of 1800x, the fans are nearly silent and in everyday use, I can't tell the performance differences! Oh, and my hearing is back too. Noctua D15 + 140mm Noctua PWM fan Asus Crosshair VI Hero Wifi
reply
Ryan
The issue I ran into with my crosshair vi x370 and the 3800x was hwinfo reports the clock it's trying to hit and not what the CPU was actually hitting. Core VID also shows the voltage it want's and not what it's actually getting. If it's not getting very close voltage input to what the VID is indicating, it's not boosting at that speed. The lower the voltage is from what the VID is reporting, the further you are actually away from that clock speed. If you aren't using auto voltage PBO is all but worthless. If you are using auto voltage the thing is running 80c+ full load if you try to squeak out any performance increase. Run the thing at 1. 33-1. 35v, run cinebench r20 a few times, keep upping clockspeed until the benchmark crashes, go back in and drop it 50mhz, then wait for Ryzen 3.
reply
The issue I ran into with my crosshair vi x370 and the 3800x was hwinfo reports the clock it's trying to hit and not what the CPU was actually hitting. Core VID also shows the voltage it want's and not what it's actually getting. If it's not getting very close voltage input to what the VID is indicating, it's not boosting at that speed. The lower the voltage is from what the VID is reporting, the further you are actually away from that clock speed. If you aren't using auto voltage PBO is all but worthless. If you are using auto voltage the thing is running 80c+ full load if you try to squeak out any performance increase. Run the thing at 1. 33-1. 35v, run cinebench r20 a few times, keep upping clockspeed until the benchmark crashes, go back in and drop it 50mhz, then wait for Ryzen 3.
reply
Add a review, comment
Other channel videos
