Cinebench R23 efficiency race

[Count von Schwalbe]

Here you goView attachment 257695


Nope, only det0x is comparing apples to oranges, since he picked the score of 8 GC cores at 65w and compared it to 16 zen 3 cores at 50w. Of course afterwards I did the same, to point exactly what you just described, the cherrypicking.


I only wanted to see if core for core zen 3 or GC is more efficient, and the results speak for themselves I think, 8 GC cores slam the 8 zen 3 cores both in performance and efficiency. There is no wattage point that zen 3 do better than 8 GC cores. Heck, im not even sure Zen 4 can do better at this point.

Looking at the result above, it seems that P-cores are about 10% more efficient than Zen 3, at 50w at least. Interesting! I would love to know how that compares across various wattages!

Also, det0x did give the oranges above his apples comparison

 

[fevgatos]

Looking at the result above, it seems that P-cores are about 10% more efficient than Zen 3, at 50w at least. Interesting! I would love to know how that compares across various wattages!

Nah, the difference is around 20%, I'm just too bored to go into manual tuning right now

But sure, let's go with 10% for now until I decide to take this more seriously, cause right now im just letting the motherboard decide wtf it's doing

 

[Count von Schwalbe]

Hey, another thought! Could you do a run with E-cores only? I have been doubtful that they are actually more efficient than the more powerful P-cores.

Nah, the difference is around 20%, I'm just too bored to go into manual tuning right now

But sure, let's go with 10% for now until I decide to take this more seriously, cause right now im just letting the motherboard decide wtf it's doing

Yeah, just strictly based on the numbers above. Numbers someone like me could replicate if I got one.

 

[fevgatos]

Hey, another thought! Could you do a run with E-cores only? I have been doubtful that they are actually more efficient than the more powerful P-cores.

E cores are not more efficient than P cores unless you drop them down to like 2-3 watts per core. Also they don't seem to like undervolting at all, 1.05v is the least I can run mine regardless of the actually clock frequency. P cores >>> e cores in efficiency. Actually, P cores >>> everything in efficiency right now.

You know what I'd find interesting? Performance normalized comparisons. Let's say, how many watts do 8 zen 3 cores need to reach 17k in CBR23. Probably 150+ watts. To match 8 P cores at 65

 

[Det0x]

When I run Alderlake with all 16 cores enabled I get
15215 points / 35watts = 434 points per watt.

Very nice, please share screenshot

Here you goView attachment 257695

14981 points / 50 watt = 299points per watt
But you said you could reach this ~performance at 35watt above, what gives ?

E Cores are not more efficient than P cores unless you drop them down to like 2-3 watts per core. Also they don't seem to like undervolting at all, 1.05v is the least I can run mine regardless of the actually clock frequency. P cores >>> e cores in efficiency

This is actually the powerlevels Zen3 are made for.. First and foremost Zen3 is a server+mobile product, with sadly desktop being a afterthought.

Highest performance Epyc's have 280 PPT powerlimit, use ~50-100watt (depending on motherboard) for interconects/IO-die which leaves 230w for the cores in bestcase
230 watt / 64 cores = ~3,6watt per core under full load
And that's why Intel cant compete in servers.. Efficiency is the key word here

how many watts do 8 zen 3 cores need to reach 17k in CBR23. Probably 150+ watts. To match 8 P cores at 65

Let me counter with, what cooling do you need to reach 33-34k in Cinebench ?

Like i said earlier, in the end it all boils down to GC physical size, they are so big that intel could only put 8(10) of those on a consumer cpu(die) and keep the price in check at the same time. (10P cores would score lower than 8p+8E in full multithreaded benchmarks)

But that is no reason to handicap desktop Zen3 with a artificial limit for 8 cores maximum in this efficiency comparison when we both have to 5900x and 5950x as normal desktop consumer cpus for sale today (well other than influence the result of course)

 

[Count von Schwalbe]

But that is no reason to handicap desktop Zen3 with a artificial limit for 8 cores maximum in this efficiency comparison when we both have to 5900x and 5950x as normal desktop consumer cpus for sale today (well other than influence the result of course)

An 8 core limit makes sense, higher core-count products can disable but the reverse does not apply.

 

[fevgatos]

But that is no reason to handicap desktop Zen3

It is when you want to figure out for example if a 16P core would be more efficient than a 5950x, right?

Here is the screenshot

Let me counter with, what cooling do you need to reach 33-34k in Cinebench ?

That's not a counter. Of course the 5950x gets a higher score when balls to the wall oced, since the E cores have a hard cap at 4 ghz (and no HT), meaning the 8 GC cores are supposed to get a score of 24k to score that 33-34 you are asking for. Im not even sure a 12900k can get that score, if it can you are looking at a chiller and probably 350-400w - maybe more.

 

[Det0x]

An 8 core limit makes sense, higher core-count products can disable but the reverse does not apply.

Yes, but on the other hand, 8 cores @ 50watt or 65watt forces Zen3 to run well outside maximum efficiency range.

Lets just say IO die use 10w, then we have:
40watt / 8 cores = 5watt per core
55watt / 8 cores = 6.875watt per core

That is a big difference from 40watt / 16 cores = 2.5watt per core

Here is the screenshot
View attachment 257712

Why did you stop with benchmate ? Just alittle strange you get ~same score at 50w and 35w (?)

 

[fevgatos]

Yes, but on the other hand, 8 cores @ 50watt or 65watt forces Zen3 to run well outside maximum efficiency range.

Lets just say IO die use 10w, then we have:
40watt / 8 cores = 5watt per core
55watt / 8 cores = 6.875watt per core

That is a big difference from 40watt / 16 cores = 2.5watt per core

We can then run 8 cores at 20 watts if you want to

Why did you stop with benchmate ? Just alittle strange you get ~same score at 50w and 35w (?)

I didn't stop, this is from an old run. Not strange at all, the score in 50w is with 8GC cores, the one in 35w is with all 16 cores (e cores on).

 

[gffermari]

Thanks fevgatos and the others for the numbers.
But the performance normalised scenario is not right. It’s like asking how much fuel a fiat panda consumes at 180km/h and comparing it to a big sedan that can cruise comfortably at the same speed.

The normalisation on the wattage is correct and obviously since alder lake is newer than zen 3 is better. But isn’t it bad that they can’t fit 16 of them in a die while AMD can fit as many as they want?

 

[fevgatos]

Thanks fevgatos and the others for the numbers.
But the performance normalised scenario is not right. It’s like asking how much fuel a fiat panda consumes at 180km/h and comparing it to a big sedan that can cruise comfortably at the same speed.

The normalisation on the wattage is correct and obviously since alder lake is newer than zen 3 is better. But isn’t it bad that they can’t fit 16 of them in a die while AMD can fit as many as they want?

Of course they can fit 16 of them in a die. The problem isn't whether they can fit them, it's whether the consumer is ready to pay for what that would cost. Also - you are ignoring the fact that - at the same time - ALD vastly outperforms zen 3 in ST. If the P cores weren't as wide, they wouldn't have that high ST performance. So it's kind of a balancing act between how big you want your cores to be and how many you can fit. That's exactly where E cores come in, offering performance / die space.

Yeah, just strictly based on the numbers above. Numbers someone like me could replicate if I got one.

Im not sure you could replicate the zen 3 numbers, that's manually tuned to no end. Which is fine of course, if I wasn't a lazy ***k I would do so as well, but im just saying the numbers are not entirely representative.

 

[Det0x]

Im not sure you could replicate the zen 3 numbers, that's manually tuned to no end. Which is fine of course, if I wasn't a lazy ***k I would do so as well, but im just saying the numbers are not entirely representative.

Yes all my Zen3 numbers above are manually tuned with static OC on all the different set powerlevels, with a binned 5950x.

This is the lowest vcore (powerusage) possible to set in bios with a static OC


Anyway I was fiddling around in the bios trying to find where my missing 13 watt are going: Cores + SOC = 25watt, not 38watt

VRM included in PPT?
x570s chipset included ?

In the end i could not find any answers, but i found some other settings that pretty much leaves this comparison moot..
I'm sure you have also found these settings in your Unify-X bios fevgatos.. Now i see why its bad to use the time finetuning a static OC when i can produce whatever fantasy numbers i want

"17237 points @ 35watt"

 

[Arc1t3ct]

Anyway I was fiddling around in the bios trying to find where my missing 13 watt are going: Cores + SOC = 25watt, not 38watt
View attachment 257727
VRM included in PPT?
x570s chipset included ?

In the end i could not find any answers, but i found some other settings that pretty much leaves this comparison moot..
I'm sure you have also found these settings in your Unify-X bios fevgatos.. Now i see why its bad to use the time finetuning a static OC when i can produce whatever fantasy numbers i want
View attachment 257732
"17237 points @ 35watt"

What trick is this???

 

[fevgatos]

Yes all my Zen3 numbers above are manually tuned with static OC on all the different set powerlevels, with a binned 5950x.

This is the lowest vcore (powerusage) possible to set in bios with a static OC
View attachment 257728

Anyway I was fiddling around in the bios trying to find where my missing 13 watt are going: Cores + SOC = 25watt, not 38watt
View attachment 257727
VRM included in PPT?
x570s chipset included ?

In the end i could not find any answers, but i found some other settings that pretty much leaves this comparison moot..
I'm sure you have also found these settings in your Unify-X bios fevgatos.. Now i see why its bad to use the time finetuning a static OC when i can produce whatever fantasy numbers i want
View attachment 257732
"17237 points @ 35watt"

I have no idea what settings you are talking about. All im doing is going into bios -- settings limit to whatever wattage im testing and undervolting. That's it.

What trick is this???

Im assumming he is messing with ac / dc LL to make the CPU missreport wattage. If that's not it, I have no clue what he is doing. But it's easy to tell it's fake juts by looking at hwinfo

 

[Det0x]

What trick is this???

First you downvote, then you ask me question ? Na its fine, ill share.

All motherboards have telemetry offset values, people can pretty much fake whatever powerusage you want.
And that's why different motherboards reads different temperatures and power etc.. Cant even compare different chipsets from the name vender, nevermind trying to compare numbers from something like Gigabyte vs Asus

And that's sadly why i said this comparison is moot
(people cant be trusted not to not use a "small" 10% offset to get ahead)

In screen above, its easy to make it read from 13w to 1850w maximum.

 

[fevgatos]

First you downvote, then you ask me question ? Na its fine, ill share.

All motherboards have telemetry offset values, people can pretty much fake whatever powerusage you want.
And that's why different motherboards reads different temperatures and power etc.. Cant even compare different chipsets from the name vender, nevermind trying to compare numbers from something like Gigabyte vs Asus

And that's sadly why i said this comparison is moot
View attachment 257744
In screen above, its easy to make it read from 13w to 1850w maximum.

And that's not true. Only AMD cpus have that. And it's pretty obvious from your hwinfo screenshot that something is messed with btw. Even your temperatures alone are a red flag, 78c on double rads @ 35w?

The power report deviation at the bottom of yourhwinfo screen is what you messed with. That feature does not exist on Intel, it does on AMD cause their CPUs do not directly report power consumed like Intel do. Here you go, from the actual Stilt himself

IMPORTANT - Explaining the AMD Ryzen "Power Reporting Deviation" -metric in HWiNFO

Ryzen CPUs for AM4 platform rely on external, motherboard sourced telemetry to determine their power consumption. The voltage, current and power telemetry is provided to the processor by the motherboard VRM controller through the AMD SVI2 interface. This information is consumed by the processors...

www.hwinfo.com

 

[Det0x]

And that's not true.

What is not true ?

And it's pretty obvious from your hwinfo screenshot that something is messed with btw. Even your temperatures alone are a red flag, 78c on double rads @ 35w?

I used those settings to highlight the problem, so all could see its wrong.

But i could used a small 10-20% offset at something like 35w or 50w and oneone would be the wiser without me telling its fantasy numbers

The power report deviation at the bottom of yourhwinfo screen is what you messed with.

Power report deviation is normal for Zen3 in Hwinfo with 100% stock settings, that dont tell you anything..

That feature does not exist on Intel, it does on AMD cause their CPUs do not directly report power consumed like Intel do. Here you go, from the actual Stilt himself

All (intel also) motherboard have telemetry offset's, like i said above, that's why different boards give different numbers in 100% in the same situation. (if you have them hidden in your bios / understand how to use them is a other story)

 

[Deleted member 185158]

First you downvote, then you ask me question ? Na its fine, ill share.

All motherboards have telemetry offset values, people can pretty much fake whatever powerusage you want.
And that's why different motherboards reads different temperatures and power etc.. Cant even compare different chipsets from the name vender, nevermind trying to compare numbers from something like Gigabyte vs Asus

And that's sadly why i said this comparison is moot
(people cant be trusted not to not use a "small" 10% offset to get ahead)
View attachment 257744
In screen above, its easy to make it read from 13w to 1850w maximum.

!

Pardon my noticing, but it looks like a 32 thread Cpu is only running 16 threads?? See CBR23 reports 8c/16t.

 

[fevgatos]

What is not true ?

I used those settings to highlight the problem, so all could see its wrong.

But i could used a small 10-20% offset at something like 35w or 50w and oneone would be the wiser without me telling its fantasy numbers

Power report deviation is normal for Zen3 in Hwinfo with 100% stock settings, that dont tell you anything..


All (intel also) motherboard have telemetry offset's, like i said above, that's why different boards give different numbers in 100% in the same situation. (if you have them hidden in your bios / understand how to use them is a other story)

It's not true that Intel has the same telemetry options, cause Intel CPUs report their consumption directly. The only way you can mess with power reporting with Intel is by messing with AC DC LL, in which case a HWinfo screenshot also makes it obvious that it's been tinkered with. On Intel CPU's VID and Vcore need to match underload for the CPU to properly report wattage. Look at the screenshot, im running CPUZ - VID and Vcore match, therefore power reporting is accurate.

 

[freeagent]

!

Pardon my noticing, but it looks like a 32 thread Cpu is only running 16 threads?? See CBR23 reports 8c/16t.

Its probably the only way to get the wattage low and still push a decent clock.. But I am just guessing

 

[Deleted member 185158]

Its probably the only way to get the wattage low and still push a decent clock.. But I am just guessing

Yes, yes indeed. Per core epeen, I like to disable SMT for that. He's still reducing per core performance with it enabled half core count.

 

[Count von Schwalbe]

!

Pardon my noticing, but it looks like a 32 thread Cpu is only running 16 threads?? See CBR23 reports 8c/16t.

we were comparing 8 cores Z3 vs ADL P-cores.

 

[Det0x]

Yes, yes indeed. Per core epeen, I like to disable SMT for that. He's still reducing per core performance with it enabled half core count.

??
Please read from the start of this thread

 

[tabascosauz]

Anyway I was fiddling around in the bios trying to find where my missing 13 watt are going: Cores + SOC = 25watt, not 38watt

Should just be normal, no? Ever since they moved to chiplets in 2019, there's always a few watts loss outside of the Cores+SOC number. 1CCD has less losses, 2CCD is about 13-14 watts, but difference is always there. Might even be a few watts on monolithic APU, I'll have to check. iirc because outside of the main VDDCR rails being measured the minor stuff is only a guesstimate in telemetry

Knowing that, it's thus hard to see how there can be a valid comparison since there's a big difference in uncore power between Intel and AMD

 

[Deleted member 185158]

??
Please read from the start of this thread

we were comparing 8 cores Z3 vs ADL P-cores.

Oh I'm sorry guys!! When I clicked it, I was sent to page 2. (facepalm)

Only have 12 threads at 65w. I suppose I would also have to do power reduction..... to participate....

But how do I prove the 50w requirement for this test?? Should I use HWinfo in the screen shot then?