After messing around with curve shaper for a couple hours, I had something very similar in mind, but thank you for all the data points and notes throughout the video. It has cleared up a lot. And knowing they stack is amazing as some high temp max frequency things can be changed from mid temp high frequency upon my already stable co-32
I'm much looking forward to your video about Zen 5 memory stability with tuning. Also please make a video about DDR8000 uclk=2000. A full analysis with benchmarks would be great because some overclockers are disagreeing about DDR8000, some are reporting significant performance increase in varied benchmarks while others are reporting a "very minor difference."
8000 with fclk=uclk is about the same performance as 6400 with fclk 2133 (3:2) My CPU requires a high SOC voltage for 6200 though and can't do 6400 1T at all. On the other hand, it does 8000 1T below stock SOC (1.05v).
Nice content In other words, now you can draw a voltage curve of almost any shape what you want, and even tie it to temperature. I think for most people it will be enough to simply set the offset to -30 at maximum load and get small overclock, or additionally limit the frequency to standard one to get undervolt on high loads and lower the temperatures
Interesting sneaky pick into 9950x overclocking guide to come. So, you could get close to 5.9Ghz all core in ccd1 and 5.6 ccd0 ? I'll be back here as soon as a new video is uploaded. Great work so far!
I feel like you are the only person who remembers about AMD marketing after launch. Lots of people reading marketing slides about OC technologies months in advanced and then never talk about it ever again after the product launches. I am still waiting for a full fledged tutorial on how TVB works from the big tech youtubers who promoted the feature by reading the slides before launch and then never heard from again.
Very interesting video,. Looking forward to further investigations into this. Could you also include a short tutorial on how you achieved getting all those data points, what softwares were used etc etc., Would love to do something similar to that on my 5900x.
I added a brief explanation in my blog post skatterbencher.com/2024/08/07/granite-ridge-overclocking-curve-shaper/ BIOS Settings Disable SMT (this lowers the operating temperature) Stress Test OCCT Memory Test > 16 threads (low stress on CPU but all cores to C0) Collecting Information HWiNFO > save to CSV (collect CPU frequency and CPU VID information) Shamino OC Tool > reduce Precision Boost Fmax at runtime in steps of 50MHz every 10 seconds Analysis Excel > Vstack frequency and VID of Core 0-7 (CCD0) & Core 8-15 (CCD1)
@@SkatterBencher thank you. The spread of voltage at each frequency you got on the curve, is this to be read as variation of voltage at a given frequency varies with temperature?
The different points are from the different cores within the CCD because they technically each have their own VFT curve. For simplicity, I average to build a single V/F curve.
I'm a little confused at 12:29, specifically the numbers of default and CO-15 at 5615MHz: How can there be a difference of 92mV (1333 - 1241)? I though the max mV offset of a single step from CO is 5mV? 15*5mV = 75mV. So how can there be a 92mV offset with a CO value of -15? EDIT: many other numbers also don't add up, sometimes there's less than 3mV offset for 1 CO step. How did you measure these?
Awesome in depth explanation! Can you explain how you got the initial vCore/freq curves for your CCDs? And more important: how do we check for stability for every curve point after applying curve optimizer/shaper?
HWiNFO to log voltage and frequency while running OCCT memory stress test. The memory test forces all cores to be active but the load is very low so there's no throttling. Checking stability is very tough because you need to try different parts of the curve. I usually first try the weakest part of the curve by running multiple types of workloads ranging from really light ones like OCCT memory, to more strenuous ones like gaming and geekbench, and also heavy workloads like Cinebench, and finally stress tests like OCCT CPU. The key is to find the workload that's causing problems first and use that as reference point for stability checking.
I think best CS will be +ve at mid freq and -ve at high freq so that when gaming or low thread task it boosts higher but in all core temp it maintains higher all core boost all the time , correct me if im wrong
I don't have a "best practice" yet, but my current view is that it's best to use Curve Shaper as a "fixer" in combination with either CO or ECLK. With CO, use +ve CS to fix parts of the VFT curve that's too much undervolted. With ECLK, use -ve CS to fix parts of the VFT curve that's not undervolted enough.
What's up with that hockey stick flattop on CCD1's default curve? It kind of looks like CCD1 would be operating with excessive margin under some conditions. If you use use curve shaper to put an extra ~200 MHz on CCD 1 in the "high" and "max" frequency bands, is it stable?
That's great thinking! I agree with the hypothesis and how to test it. I'll give it a try when I finally get around doing the 9950X OC guide. Note that Curve Shaper is only available per CPU. I haven't tested too thoroughly yet but it feels we should at least have a per CCD shaper since the V/F curves are so distinct.
CO_CSHF = -30 means we'd use -30 Curve Optimizer and -30 Curve Shaper (High Frequency). This will definitely be better than CO_CSHF = -30 ... if it can run! Remember, Curve Optimizer and Curve Shaper stack. So setting both to -30 will give you near -60 undervolt points around the "high frequency" shaper point. That wouldn't be stable (on this CPU).
I'm curious as to the amount of time it takes for the curve optimizer to run per core; I have a 7900X with 12 cores and wondered if this would take minutes or hours to run.
What's with the difference between the max frequencies of the 2 CCDs? 300 MHz is a significant difference, and is unacceptable unless there is a proper excuse for it such as the presence of 3D V-cache on the slower CCD or something. The max deviation I can live with is 150 MHz
This was already case with Ryzen 7000 series where CCD0 was significantly better than CCD1. We also see the same with Ryzen Threadripper, where only CCD0 has the high Fmax. I suppose the idea is that any CCD other than CCD0 serves as "core fillers" to increase core count.
@@AndyU96 no, this is a deliberate choice to simplify segmentation (always one good and one less good die packaged together, rather than having two SKUs with same core count such that former having two good dies and latter two less good dies)
Scalar gets you a little more voltage under load; however there's no way to override the voltage limits of PBO. You can go higher in voltage with manual OC.
@@TheBackyardChemist it does get over that limit but no longer at high frequencies, you need to apply negative offset in both PBO and VRM offsets if you want max boost per volt possible
Really after so many tools at this point by AMD, they could release a program that automates all this stuff so you can either get the most efficiency by your silicon, either the most performance, zero potential wasted.
They will never do this. Anything that may introduce any form of instability is an absolute nono. This is why everything is tuned so conservatively. This is because silicon degrades over time. Albeit very slowly, it does happen. By introducing such a tool to a user that is not aware of the consequences this may result in a crashing computer with a "defective" AMD CPU.
Urg I HATE the bios screen design for setting values for the curve optimiser. It should be a 2D grid of numbers with the axes of temperature and frequency . 0 means no adjustment. a negatrive number means a negative adustment. A positive number means a positive adjustment. Who the fuck thought 3 drop down menus was the UX solution of choice for this, two of which can be completely implied from the third value?
Seems most the reviewers of the 9000 series don't bother tweaking it and just say the chips suck at the default settings and normal PBO, when I think they are designed so people will utilize the new overclocking feature, im gonna get one just to see how far I can push it. Already got a 7950X3D to hit 5.9ghz stable, it's been like that since I get it last year using undervolting and +200 PBO with 65C cap on Air.
Benching with defaults firsts has been a staple for a long while now because not only enthusiasts buy these CPUs, in fact overclockers are a tiny market compared to the normal crowd.
@@darrk_matter that is precisely not representative of the situation of most people, though. most people won't bother to do more than enable expo and maybe enable pbo. if they want guides for enthusiasts, channels such as this are available. i grant that it can be irritating when people misrepresent the capabilities of certain platforms based on surface-level research or sensationalist media, but such is the nature of casual fandom
Tell me you don't understand CPU benchmarking in gaming, without telling me you don't understand CPU benchmarking in gaming. Protip - the point is to make the CPU the bottle neck, not the GPU.
Curve Shaper is a powerful tool that needs an MSI afterburner style GUI to make viable; the BIOS controls are horrendous. Great video, thank you!
You are one of the few if not the only that delves into this at the deep end and explains it very well.
honestly, even as an enthusiast, this is quite a lot of info to absorb
Not really. We been doing it on GPUs forever, just with better UIs using things like MSI Afterburner.
I appreciate you taking the time to do these deep dives. Very helpful.
Thank you for doing this. AMD OC is where it's going to shine in 9K series and fine-tune for individual use cases.
Wow, this video is wonderful, it explains the bios settings I don't understand AND graphs everything to make it make so much more sense.
Thanks for your hard work, this is really exciting
Really interesting deep dive, thanks! Great to see all the graphs and data you gathered.
After messing around with curve shaper for a couple hours, I had something very similar in mind, but thank you for all the data points and notes throughout the video. It has cleared up a lot. And knowing they stack is amazing as some high temp max frequency things can be changed from mid temp high frequency upon my already stable co-32
can't wait for the next overclocking guide. Excellent details and explanation!
I cant' wait to see what 9950X3D can do with overclocking. Got my 7950X3D async'd by your method and it's really amazing. Thanks Much
What mobo are you using with your 7950x3d?
Nice presentation on important observations and concepts. Thanks.
Looking forward to your next guides!
I'm much looking forward to your video about Zen 5 memory stability with tuning. Also please make a video about DDR8000 uclk=2000. A full analysis with benchmarks would be great because some overclockers are disagreeing about DDR8000, some are reporting significant performance increase in varied benchmarks while others are reporting a "very minor difference."
8000 with fclk=uclk is about the same performance as 6400 with fclk 2133 (3:2)
My CPU requires a high SOC voltage for 6200 though and can't do 6400 1T at all. On the other hand, it does 8000 1T below stock SOC (1.05v).
Nice content
In other words, now you can draw a voltage curve of almost any shape what you want, and even tie it to temperature. I think for most people it will be enough to simply set the offset to -30 at maximum load and get small overclock, or additionally limit the frequency to standard one to get undervolt on high loads and lower the temperatures
Interesting sneaky pick into 9950x overclocking guide to come.
So, you could get close to 5.9Ghz all core in ccd1 and 5.6 ccd0 ?
I'll be back here as soon as a new video is uploaded. Great work so far!
I feel like you are the only person who remembers about AMD marketing after launch. Lots of people reading marketing slides about OC technologies months in advanced and then never talk about it ever again after the product launches.
I am still waiting for a full fledged tutorial on how TVB works from the big tech youtubers who promoted the feature by reading the slides before launch and then never heard from again.
Great content... waiting for the next video!!
Thanks! I just finished writing an update to Curve Shaper to talk about temperature. Very excited to share that with everyone! :)
Very interesting video,. Looking forward to further investigations into this.
Could you also include a short tutorial on how you achieved getting all those data points, what softwares were used etc etc., Would love to do something similar to that on my 5900x.
I added a brief explanation in my blog post skatterbencher.com/2024/08/07/granite-ridge-overclocking-curve-shaper/
BIOS Settings
Disable SMT (this lowers the operating temperature)
Stress Test
OCCT Memory Test > 16 threads (low stress on CPU but all cores to C0)
Collecting Information
HWiNFO > save to CSV (collect CPU frequency and CPU VID information)
Shamino OC Tool > reduce Precision Boost Fmax at runtime in steps of 50MHz every 10 seconds
Analysis
Excel > Vstack frequency and VID of Core 0-7 (CCD0) & Core 8-15 (CCD1)
@@SkatterBencher thank you. The spread of voltage at each frequency you got on the curve, is this to be read as variation of voltage at a given frequency varies with temperature?
The different points are from the different cores within the CCD because they technically each have their own VFT curve. For simplicity, I average to build a single V/F curve.
I'm a little confused at 12:29, specifically the numbers of default and CO-15 at 5615MHz: How can there be a difference of 92mV (1333 - 1241)? I though the max mV offset of a single step from CO is 5mV? 15*5mV = 75mV. So how can there be a 92mV offset with a CO value of -15?
EDIT: many other numbers also don't add up, sometimes there's less than 3mV offset for 1 CO step. How did you measure these?
I detailed the process on my blog: skatterbencher.com/2024/08/07/granite-ridge-overclocking-curve-shaper/#The_Voltage-Frequency_Curve.
Awesome in depth explanation! Can you explain how you got the initial vCore/freq curves for your CCDs?
And more important: how do we check for stability for every curve point after applying curve optimizer/shaper?
HWiNFO to log voltage and frequency while running OCCT memory stress test. The memory test forces all cores to be active but the load is very low so there's no throttling.
Checking stability is very tough because you need to try different parts of the curve. I usually first try the weakest part of the curve by running multiple types of workloads ranging from really light ones like OCCT memory, to more strenuous ones like gaming and geekbench, and also heavy workloads like Cinebench, and finally stress tests like OCCT CPU. The key is to find the workload that's causing problems first and use that as reference point for stability checking.
@@SkatterBencher What do you think about corecycler?
Great software! I used it a lot in previous overclocking guides
@@SkatterBencher do you know a way to lock a specific point on the vid/freq curve for stability testing?
So... they fixed the stability at low frequencies that was impacted by too low voltages, and are achieving even higher clocks?
do you use curve optimizer + curve shaper or one or another?
I've been using Curve Optimizer as the main undervolt and Curve Shaper to patch up small instabilities
@@SkatterBencher should be good if you can start from -30 CO, mine can't go above 25, so im using only shaper
Like your previous videos, thank you. Great stuff. My question is that does it work with ryzen 7950x?
I tried Curve Optimizer on my 7950x and it had no effect.
@bgtubber I think it will work. Because there is no physical component, we need to control the curve shaper. It is just software development.
@@Hadi-zw9mb That doesn't mean that it will work. They can still block it when a non-9000 series CPU is detected on the motherboard.
@bgtubber it might happen. But I see what AMD did with FSR in their graphic cards and hope that they will do the same for curve shaper.
@@Hadi-zw9mb Yes, I hope so too. Fingers crossed.
Great stuff
I cannot wait for this week's drop!
Is it a bug if a BIOS shows the Curve Shaper menu with a Zen 4 CPU (7800X3D)?
It works with Zen 4 too, someone on reddit already show how Curve Shaper affect his 7950x
zen 6 now introducing v/f/t point cloud designer!
It's better to have the tools when you don't need it than to need the tools and not have it!
is that a joke or serious?
@@erkinalp joke lol
What's about "New memory overclocking on-the-fly" feature from the AMD slide?
I think best CS will be +ve at mid freq and -ve at high freq so that when gaming or low thread task it boosts higher but in all core temp it maintains higher all core boost all the time , correct me if im wrong
I don't have a "best practice" yet, but my current view is that it's best to use Curve Shaper as a "fixer" in combination with either CO or ECLK.
With CO, use +ve CS to fix parts of the VFT curve that's too much undervolted. With ECLK, use -ve CS to fix parts of the VFT curve that's not undervolted enough.
What's up with that hockey stick flattop on CCD1's default curve? It kind of looks like CCD1 would be operating with excessive margin under some conditions. If you use use curve shaper to put an extra ~200 MHz on CCD 1 in the "high" and "max" frequency bands, is it stable?
That's great thinking! I agree with the hypothesis and how to test it.
I'll give it a try when I finally get around doing the 9950X OC guide. Note that Curve Shaper is only available per CPU. I haven't tested too thoroughly yet but it feels we should at least have a per CCD shaper since the V/F curves are so distinct.
So what about CO_CSHF = -30? More frequency for less heat is my absolute preferred option.
CO_CSHF = -30 means we'd use -30 Curve Optimizer and -30 Curve Shaper (High Frequency). This will definitely be better than CO_CSHF = -30 ... if it can run!
Remember, Curve Optimizer and Curve Shaper stack. So setting both to -30 will give you near -60 undervolt points around the "high frequency" shaper point. That wouldn't be stable (on this CPU).
I'm curious as to the amount of time it takes for the curve optimizer to run per core; I have a 7900X with 12 cores and wondered if this would take minutes or hours to run.
In my experience, per-core curve optimizer tuning attempts takes multiple hours
Holy Cow!!!
Do CO -30 and Curve Shaper -30 stack to get even lower voltages?
I haven't able to try that specific case but it should stack.
Are the rumors still positive regarding the 9000X3D SKUs being unlocked this time around?
You have a 9950X??? Can you do a video on overclocking that cpu or you have to wait for review embargo on august 15?
well if AMD sent it to him, yes he has to wait
Gotdam that's alot of data!
What's with the difference between the max frequencies of the 2 CCDs? 300 MHz is a significant difference, and is unacceptable unless there is a proper excuse for it such as the presence of 3D V-cache on the slower CCD or something. The max deviation I can live with is 150 MHz
This was already case with Ryzen 7000 series where CCD0 was significantly better than CCD1. We also see the same with Ryzen Threadripper, where only CCD0 has the high Fmax. I suppose the idea is that any CCD other than CCD0 serves as "core fillers" to increase core count.
@@SkatterBencher Sounds more like a scam, I wasnt even aware of this
@@AndyU96 no, this is a deliberate choice to simplify segmentation (always one good and one less good die packaged together, rather than having two SKUs with same core count such that former having two good dies and latter two less good dies)
I'd love to have a script or a software that spits out that data
fr, how easy overclocking would be, with these graphs in real time
Could you make the V/F data available for download?
Should be no problem. Not sure how useful it would be though?
Did I miss it, or is there no way to increase the max voltage past 1.35V while using the curve?
Scalar gets you a little more voltage under load; however there's no way to override the voltage limits of PBO. You can go higher in voltage with manual OC.
@@SkatterBencher What happens if you set an offset in the VRM controller IC?
@@TheBackyardChemist it does get over that limit but no longer at high frequencies, you need to apply negative offset in both PBO and VRM offsets if you want max boost per volt possible
Really after so many tools at this point by AMD, they could release a program that automates all this stuff so you can either get the most efficiency by your silicon, either the most performance, zero potential wasted.
They will never do this. Anything that may introduce any form of instability is an absolute nono. This is why everything is tuned so conservatively. This is because silicon degrades over time. Albeit very slowly, it does happen. By introducing such a tool to a user that is not aware of the consequences this may result in a crashing computer with a "defective" AMD CPU.
AMD won’t do this because they don’t wanna do anything like Intel, and with good reason.
AMD provide you golden samples or are you always lucky?))) Waiting for Zen5 memory video.
Urg I HATE the bios screen design for setting values for the curve optimiser. It should be a 2D grid of numbers with the axes of temperature and frequency . 0 means no adjustment. a negatrive number means a negative adustment. A positive number means a positive adjustment. Who the fuck thought 3 drop down menus was the UX solution of choice for this, two of which can be completely implied from the third value?
Having separate enable, sign and magnitude inputs in the curve shaper config is horrifically bad UI design.
BIOS design fail
same problem in curve optimizer though
🥰
The GOAT
My ptsd is triggered 🥲
Seems most the reviewers of the 9000 series don't bother tweaking it and just say the chips suck at the default settings and normal PBO, when I think they are designed so people will utilize the new overclocking feature, im gonna get one just to see how far I can push it. Already got a 7950X3D to hit 5.9ghz stable, it's been like that since I get it last year using undervolting and +200 PBO with 65C cap on Air.
Benching with defaults firsts has been a staple for a long while now because not only enthusiasts buy these CPUs, in fact overclockers are a tiny market compared to the normal crowd.
I highly doubt your effective clock is hitting 5.9ghz
5.9 GHz?
Precisely. Doesn’t matter to their simps, though, they’ll parrot whatever their fav “techtubers” say despite lack of fair and adequate testing.
@@darrk_matter that is precisely not representative of the situation of most people, though. most people won't bother to do more than enable expo and maybe enable pbo. if they want guides for enthusiasts, channels such as this are available.
i grant that it can be irritating when people misrepresent the capabilities of certain platforms based on surface-level research or sensationalist media, but such is the nature of casual fandom
Why can't AMD just document these facts so we don't need to invent the wheel again
it isnt august 14th Aware
Another feature that would take away your precious sleep and time gaming..lol..
you only need to do this tuning once a yea
Yet another example of how badass Zen 5 truly is. Can’t wait to get my 9700x and cut it loose with PBO/CO 🔥
Worse for gaming than 7800x3d and costs more. Just wait for the x3d release. You only get significantly better performance for professional workloads
and then you realize you don't play at 720p low with a 4090 and you gain literally nothing from it.
Tell me you don't understand CPU benchmarking in gaming, without telling me you don't understand CPU benchmarking in gaming.
Protip - the point is to make the CPU the bottle neck, not the GPU.