Question Alder Lake - Official Thread

[Hulk]

With the release of Alder Lake less than a week away and the "Lakes" thread having turned into a nightmare to navigate I thought it might be a good time to start a discussion thread solely for Alder Lake.

 

[igor_kavinski]

First, that is one benchmark. Second, you are ignoring power consumption.

I'm not ignoring the power consumption. ADL will consume more power but it will also get the work done faster too, so the extra power consumption will be for shorter periods of time. This assumes that workloads are mixed. Sustained and prolonged highly parallel workloads make ADL the bad choice. But then, people who do that sort of stuff are not in the millions.

You are angry at Intel for their past greed. I'm angry at AMD for their current laziness. They are late with Zen 4. They got complacent after Zen 3 and didn't try to be first to market with a DDR5 platform. Competition would have been a lot better with AMD being their usual self.

 

[igor_kavinski]

Oh, and MSRP might be higher, but my last 2 5950x were $600 and $580. And they go for $650 everyplace.

Maybe in the US. Not the case where I live (UAE). 5950X costs $800+ here.

 

[Markfw]

Maybe in the US. Not the case where I live (UAE). 5950X costs $800+ here.

Worldwide, things differ. Lets drop this and stick to Alder lake in an Alder lake thread. They have good uses, but are not totally dominant right now. There are too many areas where they are behind. Default power settings for one. And as you said heavily 100% multi-threaded tasks. But its progress. But why they chose high power just to win a few benchmarks to me is a mistake on Intel's part.

 

[DrMrLordX]

That's 16 ADL threads smacking down 32 Zen 3 threads in MySQL. How is that NOT impressive?

The 11900k AND 11700k were already beating the 5950X. If you want MySQL performace, desktop Ryzen doesn't seem to be the way to go.

 

[szrpx]

Is that benchmark AVX-512 accelerated or something? Cypress Cove isn't better than Zen 3 IPC wise.

Also, Alderlake barely moves the needle vs Rocketlake in that benchmark, not an uplift worth mentioning.

 

[lobz]

You made a good joke! If you need another:

I could go on and on spamming this thread with dozens of times he called Intel or its chips jokes, crap, etc. But, I prefer just to quote himself when he says he hates Intel in multiple threads.

That was not a joke but nonetheless it was amusing to observe how desperately you were trying to use your obvious dishonesty, even searching for a post from literally 20 years before and STILL not realizing what that sentence means 😂

 

[Zucker2k]

That was not a joke but nonetheless it was amusing to observe how desperately you were trying to use your obvious dishonesty, even searching for a post from literally 20 years before and STILL not realizing what that sentence means 😂

Actually, he didn't have to go back 20 years. Thankfully, the person you're defending is quite consistent so here goes. This is from February 12 this year, just 15 days ago:

And people wonder why we hate Intel most of the time ! Charging extra for whats already in the chip. They care about nothing but Money. I hope AMD can change that eventually through competition. In servers they are the king, but too many are unconvinced that AMD is better (stupid managers, I know, I retired from a company that had data center managers like that).

This post is only referencing servers...

Inflammatory posting along with continued thread derailing
is not allowed.


esquared
Anandtech Forum Director

 

[JoeRambo]

Is that benchmark AVX-512 accelerated or something? Cypress Cove isn't better than Zen 3 IPC wise.

Also, Alderlake barely moves the needle vs Rocketlake in that benchmark, not an uplift worth mentioning.

It is benchmark that operates at the speed of memory subsystem. Highly parallel, but there is plenty of lock contention that in turn means plenty of inter thread traffic and ends up hurting AMD, since they need to communicate between CCX'es over memory and take memory latency and that stops workload from scaling.
In enterprise environment this is sometimes worked around by placing VMed or raw DB instances with core ( and/or NUMA ) controls, ensuring that workload stays within boundaries of one CCX. Drawback is of course limit of max TPS, but you can see it already in benchmark, that throwing a boatload of cores does not scale much beyond certain point.

ADL is penalized for different reasons, hybrid is not really a DB throughput forte once in lock contention regime.

Oh, and Mysql is obviuosly a bad benchmark to show scaling, not know to scale after a certain point. Just like plenty other real world workloads.

 

[Carfax83]

I personally find it amusing when people are not impressed by Alder Lake. I especially don't understand it when I see this from the TPU i9-12900K review:

Here's another one. The Intel Core i5 12400 (crippled with DDR4 btw) beating the 5800x at 65w in one of the most CPU intensive games of the modern era, which scales up to 6 threads so it's not single threaded either.

And I've seen some people say that Alder Lake only has a 7% IPC advantage over Zen 3?!

 

[esquared]

The topic of this thread is:

Alder Lake - Official Thread


Further off topic post and derailing will be met with warnings and infractions.


esquared
Anandtech Forum Director

 

[DrMrLordX]

And I've seen some people say that Alder Lake only has a 7% IPC advantage over Zen 3?!

The bench in question can't even saturate all the 5800x's cores so there's no real core advantage; regardless, the uncapped 12400 is only 10% faster than the 5800x while the capped 12400 is only 6.25% faster. Without knowing more about clocks/power draw it's impossible to draw many conclusions about IPC from this bench alone. My guess is the 12400 is sitting at around 4.2-4.4 GHz while the 5800x . . . God only knows where that thing's running.

 

[Carfax83]

The bench in question can't even saturate all the 5800x's cores so there's no real core advantage; regardless, the uncapped 12400 is only 10% faster than the 5800x while the capped 12400 is only 6.25% faster. Without knowing more about clocks/power draw it's impossible to draw many conclusions about IPC from this bench alone. My guess is the 12400 is sitting at around 4.2-4.4 GHz while the 5800x . . . God only knows where that thing's running.

But when you compare it to the 5600x, which also has 6 cores and a slightly higher boost clock than the 12400 still loses by 26% for minimums and 13% for the average. There is hardly any scaling from the 5600x to the 5800x and the 3600 to the 3700x, but when you look at the Zen 2 vs Zen 3 performance, the difference is huge. This demonstrates that the benchmark heavily favors IPC and cache for CPU performance.

Also, the YouTube video I posted showing the benchmark was running on a 5800x and it dipped under 60 a lot, though to be fair it was run at 1440p and on a slower GPU than what Techspot used.

 

[DrMrLordX]

But when you compare it to the 5600x, which also has 6 cores and a slightly higher boost clock than the 12400 still loses by 26% for minimums and 13% for the average.

Okay but the 5600X also loses to the 5800X, which at least as of 2020 was kind of abnormal in a lot of game benchmarks (in some game benches at Vermeer's release, the 5600x was the fastest of the lot). Look at the minimums, something's going on there when moving to the 5800X. And we still don't know what clockspeed the 5600X is running either so . . . what conclusions can we draw?

 

[Carfax83]

Okay but the 5600X also loses to the 5800X, which at least as of 2020 was kind of abnormal in a lot of game benchmarks (in some game benches at Vermeer's release, the 5600x was the fastest of the lot). Look at the minimums, something's going on there when moving to the 5800X. And we still don't know what clockspeed the 5600X is running either so . . . what conclusions can we draw?

It's not just the 5800x and the 5600x that shows unexpected abnormalities, the 10600K is also a bit faster than the 10700K despite having less cores, less cache and lower frequency. That said, I did some digging and in a Reddit AMA thread for the in house Schmetterling engine which is used in the game, the developer claims the engine can scale to 64 threads provided there's no GPU bottleneck.

Source

The only reviewer that mentioned CPU scaling was Dsogaming, and it appears the game might use up to 8 threads, but doesn't like hyperthreading beyond a certain point. AMD's intercore latency for the 5600x and 5800x should be less than Intel's Comet Lake CPUs due to a much larger L3 cache, which may explain why the 5800x gains slightly compared to the 5600x while the opposite is true for the 10700K and 10600K which has much less L3 cache than both Zen 3 and Golden Cove per core.

That's the only explanation I can think of for the results. It also shows why Alder Lake has much better scaling than the other CPUs, because of the varying amounts of L3 cache between the CPUs.

Also, this game loves fast memory and benefits greatly from DDR5 which shows that it accesses the main memory frequently, hence why it may also love large L3 caches:

 

[jpiniero]

https://www.tomshardware.com/news/intel-nukes-alder-lake-avx-512-now-fuses-it-off-in-silicon

Intel is fusing off AVX-512 for real this time.

 

[jpiniero]

Intel introduces 12th Gen Core vPRO processors and W680 workstation chipset for professionals - VideoCardz.com

Intel Introduces vPro Platform for Revolutionary Business Performance Intel introduces vPro platform with 12th Gen Intel Core processors for revolutionary business performance. What’s New: Today, Intel announced the latest Intel vPro platform, powered by 12th Gen Intel® Core™ – which delivers...

videocardz.com

They are, however going to fully support ECC on Alder Lake Core processors... provided you buy a W690 board. Is limited to high end processors only but does include K. Wonder why no Xeon E.

 

[SAAA]

https://www.tomshardware.com/news/intel-nukes-alder-lake-avx-512-now-fuses-it-off-in-silicon

Intel is fusing off AVX-512 for real this time.

I wonder what's their future strategy, supporting AVX512 so sparsely these past years... are they killing it entirely, reserved for servers only? Or maybe for newer chips (Arrow etc) so they can claim again that "feature"?

Surely it's a weird step back after making it mainstream on rocket lake and some mobile parts.

 

[jpiniero]

I wonder what's their future strategy, supporting AVX512 so sparsely these past years... are they killing it entirely, reserved for servers only? Or maybe for newer chips (Arrow etc) so they can claim again that "feature"?

Surely it's a weird step back after making it mainstream on rocket lake and some mobile parts.

Either the Atom cores will get AVX-512 support eventually or some transparent feature to kick processes back to the Big Cores if AVX-512 instructions are hit.

 

[Exist50]

Either the Atom cores will get AVX-512 support eventually or some transparent feature to kick processes back to the Big Cores if AVX-512 instructions are hit.

Or they'll just deprecate AVX512 outside of server chips for the foreseeable future. I think that's quite plausible.

 

[jpiniero]

Or they'll just deprecate AVX512 outside of server chips for the foreseeable future. I think that's quite plausible.

I doubt that given that Zen 4 supports AVX-512.

 

[igor_kavinski]

Also depends on AMD including AVX-512 in Zen 4. If they make it available from top to bottom, all 11th gen CPUs plus Zen 4 CPUs will have AVX-512, giving some incentive to application developers to target owners of these newer systems. AVX2 also gained widespread adoption when AMD jumped on the bandwagon.

12th gen owners will be left in the cold if AVX-512 catches on, but I'm sure Intel can just re-enable AVX-512 in these CPUs if the market dynamics demand it (I still don't believe newer CPUs have it fused off, despite what they claim. It's probably something in the newer microcode preventing the CPU from exposing AVX-512).

 

[Exist50]

I doubt that given that Zen 4 supports AVX-512.

AMD doesn't have the marketshare to drive ISA adoption by itself though. I don't think it's a good thing, but Intel killing AVX512 on client would de facto do the same for AMD.

 

[igor_kavinski]

AMD doesn't have the marketshare to drive ISA adoption by itself though. I don't think it's a good thing, but Intel killing AVX512 on client would de facto do the same for AMD.

If they are really serious, they should microcode nuke 11th gen's AVX-512 support too.

It is possible that this is Intel's strategy to nullify AMD's hard work on AVX-512 in Zen 4 but it could backfire if AMD comes up with a killer application to boost their AVX-512 support's visibility and usefulness.

 

[jpiniero]

AMD doesn't have the marketshare to drive ISA adoption by itself though. I don't think it's a good thing, but Intel killing AVX512 on client would de facto do the same for AMD.

It'd look bad in marketing. I suspect Intel's original original plan was to have the OS handle the kick feature by using the new Fast Read Processor ID instructions but soon realized that was a bad idea... and between that and 10 nm uncertainty didn't bother coming up with a real solution and left it to further processors.

There's still a chance there will be Xeon E processors released without small cores and with AVX-512 I guess.

 

[Exist50]

If they are really serious, they should microcode nuke 11th gen's AVX-512 support too.

They advertised it, so too late. But I struggle to find another justification for their efforts to disable it other than really not wanting it to be used.