<< Solved >> CPU advise: Intel Core i9 10900 or 10850K?

[e-troc]

MCE is only on K SKU

The 10900k has 100mhz more than 10850k, using more energy. Should be a higher binned chip. Retail intel CPUS have usually between -40 to -80 undervolt headroom. The onyl 10900k i had was -70. I hade some 10850k do only 45 and some go -80. 10900ES go -90 / - 110

10850K seems like a good choice then and I'll definitely try the settings you recommended. Thanks again!

 

[etorix]

I still don't really understand the practical differences between the 10900 and the 10850K, however.

No-one really understands… but the most likely explanation is that Intel struggled to bin enough dies at 10900K requierements, and created the slighly lower spec 10850K for what it could actually produce.

 

[c-o-pr]

I found this writeup on overclocking the non-K 10900:

Intel Core i9-10900 Review - Fail at Stock, Impressive when Unlocked - TechPowerup

Intel Core i9-10900 Review - Fail at Stock, Impressive when Unlocked

In our Core i9-10900 review we're taking a close look at what can be gained from unlocking the power limit of this 65 W processor. Results are impressive: up to 40% faster apps and performance that rivals the Core i9-10900K at much lower pricing, but heat output is increased, too.

www.techpowerup.com

Re WhenMusicAttacks's run downs on OC: I'm reading, but also I'm mystified at this way of talking:

I would value the 10850k slightly more because of what i wrote. I had a client that had performance issues while the cpu was just using 80w - enabling all core turbo to 5ghz moved the power on that same worload reach 100w, and solved the stuttering. Even though with Cinebench the scores were almost the same, with the stock config using 150w (undervolted) and doing 2400 and the MCE setting

This paragraph sounds like it says something technical, but what?

• Saying "My client's app stuttered at 80W" is a very peculiar statement. I'm reading between the lines to get the point, but this language leads to absurd thinking. At a technical level, describing app performance in terms of power disparities is a wild figure of speech. While we can observe that clocking up a CPU can improve code speed, and that clocking up has power implications, we can't jump from end-user thermal assessment in watts back to general app suitability to purpose. This way of thinking is like trying to assess African butterfly wing flaps effects on insect migration by taking measurements of the whorl of a Gulf hurricane.
• Benchmarks are never intended to be predictors of performance, they're intended to be a reference by which system changes can be compared. There's an huge gap between making a benchmark comparison and understanding a system balance, so saying that Cinebench stayed the same in the face of thermals is a meaningless point regarding app responsiveness.

For example, if the CPU is bumping up against its rev limiter this can cause oscillation due to hysteresis in feedback of the governor loop. This will manifest as media playback stuttering.

The problem and solution are not in sheer power, but getting the governor configuration to a place the system doesn't oscillate. Note that internally the causes and effects may be extremely complex, but get resolved with a simple config adjustment.

I suggest to everyone that if you can avoid thinking about voltage levels and boost clock over-current time windows, you want to avoid such parameters. How much time do you want to spend twiddling voltage to get a config that doesn't crash, and how transportable is your learning? If there's a custom config for every build, that's a big hazard. OTOH if there's a solid recipe you want it in form of lore, you want it in the board logic so you don't have to futz with it!

For production systems, what you want is for Intel and board maker to produce a stable arrangement at a level of performance against a reference you can test to match your workload. This lets you assess that a system is suitable for a use-case (client), and let's you confirm that config changes don't spoil the setup.

I'm not putting down hard-won experience of gaming PC mindset, but I am pointing out pitfalls of this mindset to understanding system perf. Gamers don't think systematically, they think tactically. They like an incremental advantage for conquest and tend to tolerate big losses in stability to win the fight — like losing all their data. Note that Boxers and football (rugby) players are never the picture of long-term heal. This analysis is my way of coping with the cognitive dissonance of this whole scene, which is largely about making Mac as confusing and unreliable as Windows PC!

To bring this back around, my experience with Asus ROG AI OC is that its trying to simplifies finding a top performing stable configuration without lots of crash testing. It allows hand tailoring of OC voltages and power windows for those who are crazy enough to think like this, but it also makes the tweaking seem passe by wrapping the config up with some simple settings.

There are still many ways to go wrong, because the settings are poorly documented and the OC features are very dynamic. Moreover, if your work depends a build which clocked to the edge of its thermal limit and you suffer through a literal heat wave on big job... But the Asus BIOS does greatly ease finding a sweet spot.

That's my story, an I ain't changin'

 

[WhenMusicAttacks]

I found this writeup on overclocking the non-K 10900:

Intel Core i9-10900 Review - Fail at Stock, Impressive when Unlocked - TechPowerup

Intel Core i9-10900 Review - Fail at Stock, Impressive when Unlocked

In our Core i9-10900 review we're taking a close look at what can be gained from unlocking the power limit of this 65 W processor. Results are impressive: up to 40% faster apps and performance that rivals the Core i9-10900K at much lower pricing, but heat output is increased, too.

www.techpowerup.com

Re WhenMusicAttacks's run downs on OC: I'm reading, but also I'm mystified at this way of talking:



This paragraph sounds like it says something technical, but what?

• Saying "My client's app stuttered at 80W" is a very peculiar statement. I'm reading between the lines to get the point, but this language leads to absurd thinking. At a technical level, describing app performance in terms of power disparities is a wild figure of speech. While we can observe that clocking up a CPU can improve code speed, and that clocking up has power implications, we can't jump from end-user thermal assessment in watts back to general app suitability to purpose. This way of thinking is like trying to assess African butterfly wing flaps effects on insect migration by taking measurements of the whorl of a Gulf hurricane.

What i think you don't understand is that the intel programmed stepping gets you into some strange limitations like if you need 5 core, you can get all 10 to 4.6ghz or just 5 at 4.6 and 3 at 1.2, it does not matter. With MCE enabled and power limit enforced, you can get 1 core at 5, 4 cores at 4.7, 3 cores at 1.2ghz and it will fit in the TDP, but this state would not fit into intel frequency scheme.

• Benchmarks are never intended to be predictors of performance, they're intended to be a reference by which system changes can be compared. There's an huge gap between making a benchmark comparison and understanding a system balance, so saying that Cinebench stayed the same in the face of thermals is a meaningless point regarding app responsiveness.

Cinebench is used to reproduce worst case scenario for thermals, where i can test the extreme condition for the CPU. Without power limit and MCE enabled, i would need to lower my undervolting, thus getting terrible efficiency. And i don't want my machiens to be like the idiot macbooks that will turbo to their maximum rated frequency and then throttle down because of thermal and vrm limits and keep bouncing up and down. I want my maximum power state to be held indefinitely. Wich is never a possibility with a 10900k with MCE on and power limit off because of the TIM

I suggest to everyone that if you can avoid thinking about voltage levels and boost clock over-current time windows, you want to avoid such parameters. How much time do you want to spend twiddling voltage to get a config that doesn't crash, and how transportable is your learning? If there's a custom config for every build, that's a big hazard. OTOH if there's a solid recipe you want it in form of lore, you want it in the board logic so you don't have to futz with it!

For production systems, what you want is for Intel and board maker to produce a stable arrangement at a level of performance against a reference you can test to match your workload. This lets you assess that a system is suitable for a use-case (client), and let's you confirm that config changes don't spoil the setup.

I'm not putting down hard-won experience of gaming PC mindset, but I am pointing out pitfalls of this mindset to understanding system perf. Gamers don't think systematically, they think tactically. They like an incremental advantage for conquest and tend to tolerate big losses in stability to win the fight — like losing all their data. Note that Boxers and football (rugby) players are never the picture of long-term heal. This analysis is my way of coping with the cognitive dissonance of this whole scene, which is largely about making Mac as confusing and unreliable as Windows PC!

My computers make up usually for 2/3 of the italian top 100 tracks, being producers, mix and mastering engineers, musicians, even cover art photographers. All the best selling albums of the year are made using these machines, that i tune individually with weeks of testing to get the best performance per watt so they can be quiet and long lasting.

We as a lab are dedicated to give our customers the best all around computers, also taking efficiency into account. They also actually like the idea of having ES cpus sometimes.

To bring this back around, my experience with Asus ROG AI OC is that its trying to simplifies finding a top performing stable configuration without lots of crash testing. It allows hand tailoring of OC voltages and power windows for those who are crazy enough to think like this, but it also makes the tweaking seem passe by wrapping the config up with some simple settings.

There are still many ways to go wrong, because the settings are poorly documented and the OC features are very dynamic. Moreover, if your work depends a build which clocked to the edge of its thermal limit and you suffer through a literal heat wave on big job... But the Asus BIOS does greatly ease finding a sweet spot.

That's my story, an I ain't changin'

My only eperiences with setting XMP on asus boards made me enever want to buy them again. Sadly, seems with Alder Lake they are the best to enable Avx512 on alder lake and i'll be forced to deal with them. I am just now testing z690m plus d4 for my first alder lake hackintosh. With XMP it put VCCIO to 1.4v , thanks asus.

 

[e-troc]

I opted for the 10900 on the end. Found a deal just below 400€, while the 10850K has been going back up to 450€ or more. The settings @WhenMusicAttacks outlined (MCE + manual power limit) are intriguing, however reading the review by TechPowerUp again convinced me that I’ll get good performance with minimal fuss out of the 10900. Either stock for quiet operation, or with power limit removed when I need more processing power.

 

[WhenMusicAttacks]

Key is undervolting. No need to unlock tdp that much for normal usage, just do a -60mv and see how it goes, use a good cooler and keep airflow on the vrm if you want to unlock later

 

[e-troc]

Key is undervolting. No need to unlock tdp that much for normal usage, just do a -60mv and see how it goes, use a good cooler and keep airflow on the vrm if you want to unlock later

Thanks, I'll definitely give that a try. I am only planning to unlock TDP if my DAW sessions could use some CPU headroom. One reason I am upgrading from the 8700K is that I've gotten used to working with a 64 sample buffer throughout my process but want to be able to add a few more tracks without having to render to audio.

 

[WhenMusicAttacks]

Thanks, I'll definitely give that a try. I am only planning to unlock TDP if my DAW sessions could use some CPU headroom. One reason I am upgrading from the 8700K is that I've gotten used to working with a 64 sample buffer throughout my process but want to be able to add a few more tracks without having to render to audio.

Was your 8700k properly tuned?

 

[e-troc]

Was your 8700k properly tuned?

Probably not completely optimized, however I successfully delidded it and found two alternative settings that worked really well. One undervolted for quiet operation, one overclocked for extra processing power. I've worked with the first setting 90% of the time since the overclock depended too much on room temperature.

 

[WhenMusicAttacks]

that's great thing to have - delidding
Improves OC and UV margins if done properly exp on cpu without solder