Apple M2 vs. AMD Rembrandt vs. Intel Alder Lake Linux Benchmarks

[yump]

Ryzen tends to run slower with more dimmns because the memory clock goes down.

That doesn't make sense since the more you overclock the worse your single threaded performance gets. That should also get better. I'm actually kinda surprised since other 1920x systems have gotten much worse scores. But yea somethings wrong if overclocking is making single threaded performance worse. Either your temperature is too high and it lowers the clock, or you have a lot of memory errors and that could be hurting performance. Also possible you don't have enough voltage.

Maybe you got the tabs mixed up? At 3.9 GHz fixed, qarium's ST is 95.3% of stock and MT is 104%. At 4.0 GHz fixed, ST is 97.5% and MT is 105.4 %.

What I've read is that overclocking on Zen usually makes single thread performance worse, because stock can boost to very high frequencies when only a few cores are lit up, while "OC mode" uses the same fixed clock for all the cores all the time, which is harder to get stable. In order to have a secular improvement, you have to go through the "Precision Boost Overdrive" menu to uncap the power limits and undervolt the voltage-frequency curve.

Can't say for sure though, because I don't have any chips that new.

[drakonas777]

you will not be able to compare amd 5nm chips with the apple soc.... why? simple the amd 5nm cpus will have 6nm IO die... and then again you can claim it is not ARM ISA instead it is the 6nm node of the IO die.

i say this: the nm node does not matter at all all what does matter are the products out in the hand of the people.

I was writing specifically about mobile APUs, which in AMD's case will be monolithic and N4. Node does matter, because additional transistors density and efficiency and/or performance have meaningful contribution. Sometimes it's just one piece of the puzzle, like in M1/M2 case. Sometimes it's far more crucial, like in ZEN2 case. It depends. If uArch was all that matters, everyone would use old lithos and save huge amount of manufacturing costs. Unfortunately that's not the case.

I agree regarding products though. As I said there is a series of reasons why Apple M based products are great and there is a series of reasons why a big portion of x86 products are garbage. It is how it is.

[arQon]

With Macbook Air you are right, with Macbook Pro you are dead wrong, that entire demographic is mainly creatives/programmers.

Okay. Even ignoring that "creatives" generally only need the computing power of a box of crayons :P and Macs have *always* been fashion accessories for that crowd - how many Airs do you think sell to marketing depts etc, vs Pros sold into development teams? Serious question, not snarky.

Even if you count webmonkeys as "developers", and even with making allowances for mobile apps (which, obv, you absolutely want a Mac for if you're targeting IOS), once you step outside of the Web space ~99% of client software is Windows, and ~99% of server software is Linux. While there's nothing really stopping you doing the second set *on* a Mac (and probably close to half the server devs I know do indeed use Macs for that), I'm confident you have a massive selection bias going on here.

> Like if you go to a facebook/google campus/office you will see like 80%+ of programmers have a mac.

erm, yeah - I moved to Silicon Valley a LONG time ago, thanks. 80%+ of baristas have Macs too. It doesn't make them developers.
On the flip side, 100% of VPs/etc have Macs. 100% of Marketing. About 80% of Sales. None of them are developers either, by any stretch of the imagination.

> Even if you look at projects like home-brew (macos specific package manager) and see the ridiculous number of installs if has this is quite obvious

Great: in that case, we can get much closer to a more definitive answer, and end the rather futile back and forth. Define "ridiculous" for us, since the only download numbers I could find for homebrew were on the SF mirror (at 13 for the past week, which I'm pretty sure is not the number you meant), then take the number of Mn machines sold and divide it by that.
It's certainly not going to be perfect, but it'll at least be a pretty good ballpark, and accurate enough for at least one of us to be better informed, and quite possibly both of us.

[mdedetrich]

Okay. Even ignoring that "creatives" generally only need the computing power of a box of crayons :P and Macs have *always* been fashion accessories for that crowd - how many Airs do you think sell to marketing depts etc, vs Pros sold into development teams? Serious question, not snarky.

Even if you count webmonkeys as "developers", and even with making allowances for mobile apps (which, obv, you absolutely want a Mac for if you're targeting IOS), once you step outside of the Web space ~99% of client software is Windows, and ~99% of server software is Linux. While there's nothing really stopping you doing the second set *on* a Mac (and probably close to half the server devs I know do indeed use Macs for that), I'm confident you have a massive selection bias going on here.

There is no point of arguing about this because you doing the no true scotsman fallacy and claiming that "web developers" are not "real" programmers.

> Like if you go to a facebook/google campus/office you will see like 80%+ of programmers have a mac.

erm, yeah - I moved to Silicon Valley a LONG time ago, thanks. 80%+ of baristas have Macs too. It doesn't make them developers.
On the flip side, 100% of VPs/etc have Macs. 100% of Marketing. About 80% of Sales. None of them are developers either, by any stretch of the imagination.

You ignored the part where I mentioned "programmers/engineers", I don't care about the barristers.

> Even if you look at projects like home-brew (macos specific package manager) and see the ridiculous number of installs if has this is quite obvious

Great: in that case, we can get much closer to a more definitive answer, and end the rather futile back and forth. Define "ridiculous" for us, since the only download numbers I could find for homebrew were on the SF mirror (at 13 for the past week, which I'm pretty sure is not the number you meant), then take the number of Mn machines sold and divide it by that.
It's certainly not going to be perfect, but it'll at least be a pretty good ballpark, and accurate enough for at least one of us to be better informed, and quite possibly both of us.

If you check wikipedia you can say that even as far back as 2013, it was the most contributed project to on github (see https://en.wikipedia.org/wiki/Homebr...r)#cite_note-9) . In regards to other statistics such as installs (even on formula level) see https://formulae.brew.sh/analytics/

[qarium]

Ryzen tends to run slower with more dimmns because the memory clock goes down.

right with this amount of ram you can not go that high but i could OC it a little bit still even with errors pop up it would run fine thanks to ECC... you should unterstand that only uneducated people buy non-ecc ram...
educated people and even OC people buy ECC ram because thats the only feature what makes OC the ram save.

That doesn't make sense since the more you overclock the worse your single threaded performance gets. That should also get better. I'm actually kinda surprised since other 1920x systems have gotten much worse scores. But yea somethings wrong if overclocking is making single threaded performance worse. Either your temperature is too high and it lowers the clock, or you have a lot of memory errors and that could be hurting performance. Also possible you don't have enough voltage.

"I'm actually kinda surprised since other 1920x systems have gotten much worse scores"

your example is "Ubuntu 18.04.4 LTS" the 1920x is a NUMA system means you need more modern kernel to get the NUMA optimisation. the ubuntu 18.04 is to old it does not have the NUMA optimisations. it is same for the 2990WX it has 4 NUMA nodes and the 1920x has 2 numa nodes.
but you see you are surprised of the good result it is also water cooled. and also this kind of systems need 4 or 8 dimms because of 4 ram channels... if the 18.04 result is only with 1 or 2 ram dimms the result is bad. it really needs 4 or 8 dimms see "31.32 GB" maybe it does not have the 4-8 channel configuration


"But yea somethings wrong if overclocking is making single threaded performance worse"

no in my knowlege this is perfectly normal its because of the watt/thermal budget of 250watt TDP...
if you set the clock manuelly the turbo boost function who sets the clocks automatically is turned off.
and the cpu burns more energy because all cpus set to 4ghz instead of dynamic clock to save energy.
this makes it very easy to get higher multicore results but to get better singlecore result is hard to get maybe by overclock the ram or even the FSB...

"Either your temperature is too high and it lowers the clock"

thats all wrong because i disabled dynamic clock and set all cores to fixed 4ghz .

"Also possible you don't have enough voltage."

believe it or not but it goes to 4ghz with the default voltage i did increase it a little bit to get over 4ghz but honestly no luck with that.

overclock the ram and FSB would give me some more score points...

[qarium]

Maybe you got the tabs mixed up? At 3.9 GHz fixed, qarium's ST is 95.3% of stock and MT is 104%. At 4.0 GHz fixed, ST is 97.5% and MT is 105.4 %.

What I've read is that overclocking on Zen usually makes single thread performance worse, because stock can boost to very high frequencies when only a few cores are lit up, while "OC mode" uses the same fixed clock for all the cores all the time, which is harder to get stable. In order to have a secular improvement, you have to go through the "Precision Boost Overdrive" menu to uncap the power limits and undervolt the voltage-frequency curve.

Can't say for sure though, because I don't have any chips that new.

exactly... "I've read is that overclocking on Zen usually makes single thread performance worse"

thats true... this is because the 250watt TDP limit with a fixed clock to all cores are burned away fast.
if you let it at defaults the system will save TDP limit on other cores with low frequencies to boost 1-2 cores to the max.

"Maybe you got the tabs mixed up?"

of course he mixxed it up the result did go better and better but he did not read it correctly.

i think i can get better result by OC the FSB and oc the ram and oc the ram controller in the cpu...

[Dukenukemx]

Maybe you got the tabs mixed up? At 3.9 GHz fixed, qarium's ST is 95.3% of stock and MT is 104%. At 4.0 GHz fixed, ST is 97.5% and MT is 105.4 %.

What I've read is that overclocking on Zen usually makes single thread performance worse, because stock can boost to very high frequencies when only a few cores are lit up, while "OC mode" uses the same fixed clock for all the cores all the time, which is harder to get stable. In order to have a secular improvement, you have to go through the "Precision Boost Overdrive" menu to uncap the power limits and undervolt the voltage-frequency curve.

Can't say for sure though, because I don't have any chips that new.

That makes sense. I forgot overclocking them basically disables the boost clock.

**EDIT**
Actually now I think about it doesn't because he clocked it to 4Ghz which is the boost. Just to check I ran Geekbench on my Ryzen 7 1700 which normally runs at 3Ghz and boosts to 3.7, but I have it overclocked to 3.7Ghz all the time. This system does have 3Ghz ram but because I have 4 sticks, I can't clock it higher than 2600Mhz. Keep in my that I had to increase the voltage to the memory controller to keep that ram stable. This is my HTPC machine which has a lot of stuff running including Firefox and Opera which tons of tabs left open. Actually the results are so good that they nearly match my 2700X. But it is overclocked to basically it's base frequency like what qarium did but the single threaded performance wasn't hurt. It's also generationally the same as qarium's 1920x. Also 3.7Ghz vs qarium's 4.0Gz

https://browser.geekbench.com/v5/cpu/16612986

[qarium]

That makes sense. I forgot overclocking them basically disables the boost clock.

and be sure 55% higher multicore score is better than 6-8% higher singleclock score ...

also get the point about 4K gaming the GPU is the bottleneck and not the CPU and also not the RAM.

i know at 2K or 2,5K it is different of course you need faster CPU if you want good 2K gaming performance.

the higher the resolution the more the cpu do not count for example 5K is worst and 8K is the worst.

you can not notice 6-8% single core performance in practice it is in the range of statistic error...

but of course if you use gentoo or handbrake or Blender or whatever multicore task 55% higher score will benefit you a lot.

you i think you already learned a lot... for example if you want to OC ram you want ECC ram... or the point that if you want faster system you have to optimise the bottleneck and a 8% faster core in singlecore performance does not help you at all if the bottleneck at 4K gaming is the gpu...

i think you can learn a lot if you are at phoronix.com forum ;-)

[qarium]

**EDIT**
Actually now I think about it doesn't because he clocked it to 4Ghz which is the boost. Just to check I ran Geekbench on my Ryzen 7 1700 which normally runs at 3Ghz and boosts to 3.7, but I have it overclocked to 3.7Ghz all the time. This system does have 3Ghz ram but because I have 4 sticks, I can't clock it higher than 2600Mhz. Keep in my that I had to increase the voltage to the memory controller to keep that ram stable. This is my HTPC machine which has a lot of stuff running including Firefox and Opera which tons of tabs left open. Actually the results are so good that they nearly match my 2700X. But it is overclocked to basically it's base frequency like what qarium did but the single threaded performance wasn't hurt. It's also generationally the same as qarium's 1920x. Also 3.7Ghz vs qarium's 4.0Gz

i tell you the secret behind all of this...

the 2700x and 1700 both are normal cpus without "NUMA" on 1950/2950/1920 and so one you can choose in the bios to emulate a non-numa system or a numa system for a long time windows could not handle numa the right way because emulating non-numa was better but linux can handle it perfectly means numa is better for 2990wx you can only choose numa.

also we and you we talk about the top CLOCK speed like 3,7ghz and 4ghz and so one but i have friends who massivly are in the OC game (myself is not so much in the OC game)

and the top mhz number is not the relevant for performance. there is a clock number for the ram channel data bus... this is the relevant one (the only relevant one) in DDR ram it is the half of the speed means 4000mhz ram is 2000mhz... on 2600mhz it is only 1300mhz... you can have different multiplyer to than 1:2 ...

it is not the FSB of 100mhz of the mainboard also i do not talk about RAM speed lower ram clock can be faster if this internal ram channel data bus clocks higher.

for all these cpus 2700x and 1700 and also 1920x the best result is if you have something like 4000mhz ram and overclock the ram channel data bus to 2000mhz with multiplayer 1:2 ...

this is even faster than higher clocked ram because if your ram clocks higher and the databus clocks lower because of other multiplicator then the result is not better instead it is worst.

of course again if you do overclock ram you want ECC ram because you get error message if you clock to high and also you have higher possibility of working ram even if you have error message because ECC can fix it.

gain the point you wonder why your OC result did not hurt singlecore and my OC did hurt singlecore is the point that you have non-NUMA cpu... and i have numa cpu.

you have a cpu with dualchannel ram and no data transfers between chiplets no data transfer between numa nodes... my cpu has 4 ram channel on 2 different chiplets means there goes data from 1 chiplet to the other chiplet and no cpu chiplet has full access on all ram channels because 2 ram channels are on one chiplet and 2 ram channels are on the other chiplet

the 2990wx is even worst than this in design because then 2 chiplets have no ram channels at all and 2 chiplets with 4 ram channels need to feed data to these 2 non-ram-channel numa nodes...

the more you need to push data from one numa node to another and from one ramchannel chiplet to the other ram channel chiplet the more your singlecore performance suffers from this.

again if you OC the ramchannel data bus clock to 2ghz you get the maximum performance. even at lower top clocks...

[qarium]

overclocking

Its the infinity_fabric who does make the biggest different in OC...

https://en.wikichip.org/wiki/amd/infinity_fabric

and it does not mean higher RAM clocks in mhz means it runs faster... its because there are different multiplayers for ram... means some high mhz ram configs have low performance because the infinity fabric run slow.

4000mhz ram and 2000mhz infinity fabric gives you good result for example.

some newer systems can higher than this 2000mhz infinity fabric clocks and
some older systems can only handle lower clocks on the infinity fabric.

this even gives you higher result even if the max clock speed is lower... for example only 3.5ghz instead of 4ghz..

it is similar with games at 4K and the gpu bottleneck ... the infinity fabric is a bottleneck in AMD cpus.

this effects all ryzen and threadripper from 1000,2000,3000, and AMD did fix it with ryzen5000 modern 5000 ryzen can clock at insane high infinity fabric clocks means at full speed of DDR4-5333mhz ram...

at 1:2 multiplayer 5333mhz ram means: 2666mhz infinity fabric clock...

your 1700 ryzen is good if you can do something like 1600-1800mhz... my threadripper maybe can do 2000mhz...