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.

 

[insertcarehere]

When comparing the M1 Max and 12900k, it's difficult to set them to a common performance target without (as you observed) skewing efficiency results. You can't really get an M1 Max to burn 200W+ on CBR23, but you sure can do it with a 12900k!

Of course, the original review didn't compare a 12900k in its entirety to the M1 Max. They disabled 2 P-cores to simulate a 6+8 Alder Lake-P, which is scheduled to come out on notebooks that will (nominally) compete with the MacBook Pros with M1 Pros & Max's.

 

[DrMrLordX]

Because that is the wrong way to look at CB MT. CB MT is an exercise in perf with core scaling, that is fundamentally a question of how many cores you can cram into a socket of some number of tolerable watts. If you set an anchor as some perf number but compare two different die sizes, it doesn't say a whole lot about efficiency. The smaller die will typically lose badly because it is forced to run up the curve to hit that perf number.

Er, technically the M1 Max die is enormous, but it has a lot more in there than IceStorm and FireStorm cores. If you take only the cores + cache of M1 Max then it is going to be smaller than Alder Lake-S though.

That aside, how do you then compare the efficiency? There are only so many good benchmarks in common between the M1 Max and 12900k, and it'll be awhile til we see an Alder Lake with a die area comparable to that of any M1 chip (again, not counting the iGPU + fixed function hardware bloat). Thanks to the heterogeneous nature of both chips, if you restrict thread count then you have to decide which cores to use? Plus Alder Lake hasn't got separate voltage rails for its core clusters, which is problematic.

Of course, the original review didn't compare a 12900k in its entirety to the M1 Max. They disabled 2 P-cores to simulate a 6+8 Alder Lake-P, which is scheduled to come out on notebooks that will (nominally) compete with the MacBook Pros with M1 Pros & Max's.

See posts from dmens, his entire point was that comparing an M1 Max @ 30W to a hobbled 12900k (or really any other Alder Lake) @ 35W is dirty pool since you're pushing the M1 Max to a point on its v/f curve where it's going to suffer. At least it is if you're trying to make an actual comparison based on efficiency, instead of an arbitrary performance comparison at a point near 30-35W. Yes, eventually there will be a 6+8 config Alder Lake-P out there with a published TDP of 35W (or so), so eventually this matchup may really happen. But if you're trying to make any broad statements about the relative efficiencies of the designs, such a comparison won't be entirely accurate.

 

[dmens]

Er, technically the M1 Max die is enormous, but it has a lot more in there than IceStorm and FireStorm cores. If you take only the cores + cache of M1 Max then it is going to be smaller than Alder Lake-S though.

That aside, how do you then compare the efficiency? There are only so many good benchmarks in common between the M1 Max and 12900k, and it'll be awhile til we see an Alder Lake with a die area comparable to that of any M1 chip (again, not counting the iGPU + fixed function hardware bloat). Thanks to the heterogeneous nature of both chips, if you restrict thread count then you have to decide which cores to use? Plus Alder Lake hasn't got separate voltage rails for its core clusters, which is problematic.



See posts from dmens, his entire point was that comparing an M1 Max @ 30W to a hobbled 12900k (or really any other Alder Lake) @ 35W is dirty pool since you're pushing the M1 Max to a point on its v/f curve where it's going to suffer. At least it is if you're trying to make an actual comparison based on efficiency, instead of an arbitrary performance comparison at a point near 30-35W. Yes, eventually there will be a 6+8 config Alder Lake-P out there with a published TDP of 35W (or so), so eventually this matchup may really happen. But if you're trying to make any broad statements about the relative efficiencies of the designs, such a comparison won't be entirely accurate.

If you want to remove variables, run ST and generate the perf/power curve. Leave aside the fact that ADL has much bigger p-cores and it will not scale as well on core counts per socket, by running ST you can see the architectural intent on how much area and power engineers are willing to spend to execute a single thread.

 

[pakotlar]

Er, technically the M1 Max die is enormous, but it has a lot more in there than IceStorm and FireStorm cores. If you take only the cores + cache of M1 Max then it is going to be smaller than Alder Lake-S though.

That aside, how do you then compare the efficiency? There are only so many good benchmarks in common between the M1 Max and 12900k, and it'll be awhile til we see an Alder Lake with a die area comparable to that of any M1 chip (again, not counting the iGPU + fixed function hardware bloat). Thanks to the heterogeneous nature of both chips, if you restrict thread count then you have to decide which cores to use? Plus Alder Lake hasn't got separate voltage rails for its core clusters, which is problematic.



See posts from dmens, his entire point was that comparing an M1 Max @ 30W to a hobbled 12900k (or really any other Alder Lake) @ 35W is dirty pool since you're pushing the M1 Max to a point on its v/f curve where it's going to suffer. At least it is if you're trying to make an actual comparison based on efficiency, instead of an arbitrary performance comparison at a point near 30-35W. Yes, eventually there will be a 6+8 config Alder Lake-P out there with a published TDP of 35W (or so), so eventually this matchup may really happen. But if you're trying to make any broad statements about the relative efficiencies of the designs, such a comparison won't be entirely accurate.

M1 max appears to be designed to operate at 30W and far beyond, and Apples performance claims would be inaccurate without that power draw.

Apple's M1 Pro, M1 Max SoCs Investigated: New Performance and Efficiency Heights

www.anandtech.com

how is 30W “pushing” the chip far beyond it’s power efficiency curve, given that M1 vanilla operates at 15-20W and this is a much higher performance (but not higher efficiency) design? Where do you expect M1 Max’s sweet spot to be? 25W? It’s certainly not 15-20W.

 

[eek2121]

My 5800X can do over 1800 single and 12000 multi

Micro-Star International Co., Ltd. MS-7B79 - Geekbench

Benchmark results for a Micro-Star International Co., Ltd. MS-7B79 with an AMD Ryzen 7 5800X processor.

browser.geekbench.com

This is the fastest valid 5800X on Geekbench. Note that there are some faster results on Linux/macOS, but I skipped those.

Er, technically the M1 Max die is enormous, but it has a lot more in there than IceStorm and FireStorm cores. If you take only the cores + cache of M1 Max then it is going to be smaller than Alder Lake-S though.

That aside, how do you then compare the efficiency? There are only so many good benchmarks in common between the M1 Max and 12900k, and it'll be awhile til we see an Alder Lake with a die area comparable to that of any M1 chip (again, not counting the iGPU + fixed function hardware bloat). Thanks to the heterogeneous nature of both chips, if you restrict thread count then you have to decide which cores to use? Plus Alder Lake hasn't got separate voltage rails for its core clusters, which is problematic.



See posts from dmens, his entire point was that comparing an M1 Max @ 30W to a hobbled 12900k (or really any other Alder Lake) @ 35W is dirty pool since you're pushing the M1 Max to a point on its v/f curve where it's going to suffer. At least it is if you're trying to make an actual comparison based on efficiency, instead of an arbitrary performance comparison at a point near 30-35W. Yes, eventually there will be a 6+8 config Alder Lake-P out there with a published TDP of 35W (or so), so eventually this matchup may really happen. But if you're trying to make any broad statements about the relative efficiencies of the designs, such a comparison won't be entirely accurate.

Why you people seek to defend a multi-trillion dollar company with a monopoly on it’s own platform baffles me. Apple has great chips, but early indications (and not just the benchmark already shown) indicate ADL-P will be extremely competitive, even though Intel is on an inferior node.

We will see when Intel launches ADL-P along with it’s GPUs. Leaks show that their GPUs are much more efficient than NVIDIA. If Intel ends up beating Apple at perf/watt it will be very interesting. After all, why did Apple actually drop Intel if there are no benefits to end users? Oh that is right, there IS a benefit. Higher profit margins for Apple.

Keep in mind I am not anti Apple. I own stock with them, have an iPhone, do my work on a Mac, etc. I just get tired of people making bogus claims about the efficiency of their hardware.

 

[dmens]

Why you people seek to defend a multi-trillion dollar company with a monopoly on it’s own platform baffles me.

Your faith is impressive though. Sorta like a technological Endsieg. But hey, whatever keeps you going LOL.

 

[deathBOB]

This is the fastest valid 5800X on Geekbench. Note that there are some faster results on Linux/macOS, but I skipped those.



Why you people seek to defend a multi-trillion dollar company with a monopoly on it’s own platform baffles me. Apple has great chips, but early indications (and not just the benchmark already shown) indicate ADL-P will be extremely competitive, even though Intel is on an inferior node.

We will see when Intel launches ADL-P along with it’s GPUs. Leaks show that their GPUs are much more efficient than NVIDIA. If Intel ends up beating Apple at perf/watt it will be very interesting. After all, why did Apple actually drop Intel if there are no benefits to end users? Oh that is right, there IS a benefit. Higher profit margins for Apple.

Keep in mind I am not anti Apple. I own stock with them, have an iPhone, do my work on a Mac, etc. I just get tired of people making bogus claims about the efficiency of their hardware.

I too only use artisanal handmade CPUs

 

[uzzi38]

I don't think so. We have seen a 12400 leak running at 4.0 Ghz in AIDA64 FPU using 78W, it would probably need 70W in Cinebench MT and roughly 3.9 Ghz to stay within 65W. From 4.0 or 3.9 Ghz down to 3.4 Ghz is a different world. If it drops down to 3.4 Ghz in every test after 56s something is not right which isn't even a problem, it's an ES 12400 and the bios possibly isn't fully optimized for the upcoming 65W SKUs.

Yeah just the same way we saw a 12900K pulling 217W in AIDA FPU.

Reality: Power throttling in Blender at default 241W power limit.

 

[coercitiv]

IIRC only K CPU have infinite turbo, "vanilla" CPU will go back to base clock after the burst.

You can still set the locked parts to have infinite Tau, just that the recommended stock setting is to not do that.

Non-K parts not having infinite turbo is a myth, the CPU type has no influence on turbo settings: it's all decided by the motherboard, at UEFI level. I've seen both Z and B chipset boards with unlocked power & turbo timer settings for non-K CPUs at stock. Everything comes down to VRM quality and whatever the manufacturer decides.

Here's a very clear example from ASRock, see how VRM quality has a massive influence on max power (and may have further effect on Tau):



BFB stands for Base Frequency Boost.

 

[Geegeeoh]

It's the "intended" behavior from Intel.
The fact the MoBo can ignore that doesn't make it a myth.

 

[coercitiv]

It's the "intended" behavior from Intel.
The fact the MoBo can ignore that doesn't make it a myth.

It is a myth if only Intel intends and mobo makers do as they see fit.

 

[TheELF]

It's the "intended" behavior from Intel.
The fact the MoBo can ignore that doesn't make it a myth.

The intended behavior from intel is Pl2 for tau seconds and then Pl1, even for the unlocked K parts.

Intel® Core™ Ultra and Intel® Core™ Processors Technical Resources

Intel® Core™ Ultra and Intel® Core™ Processors technical resources list includes applications notes, datasheets, packing information, product briefs, and more.

www.intel.com

volume 1, page 76

 

[Geegeeoh]

The intended behavior from intel is Pl2 for tau seconds and then Pl1, even for the unlocked K parts.

Intel is also making these processors have ‘infinite Turbo time’, meaning that with the right cooling, users should expect these processors to run up to the Turbo power indefinitely during heavy workloads.

The Intel 12th Gen Core i9-12900K Review: Hybrid Performance Brings Hybrid Complexity

www.anandtech.com

That's what I understood...

 

[DrMrLordX]

If you want to remove variables, run ST and generate the perf/power curve. Leave aside the fact that ADL has much bigger p-cores and it will not scale as well on core counts per socket, by running ST you can see the architectural intent on how much area and power engineers are willing to spend to execute a single thread.

Turbo/boost settings for ST workloads are far more aggressive than in MT, though. Plus you aren't necessarily saturating interconnect (where applicable) and caches, if that matters.

M1 max appears to be designed to operate at 30W and far beyond, and Apples performance claims would be inaccurate without that power draw.

Alder Lake-S is designed to operate at 241W and beyond, and Intel's performance claims would be inaccurate without that power draw . . .

(also, you should look at how much of M1 Max's power spec is dedicated to the actual CPU cores, vs. other parts of the die).

Why you people seek to defend a multi-trillion dollar company with a monopoly on it’s own platform baffles me.

Please dispense with the hardware tribalism. Can't you at least address the technical merits (or lack thereof) of the post instead of just reflexively assuming that someone is attacking a sacred cow? M1 Max is an infuriatingly difficult piece of hardware to evaluate on its own thanks to the hw acceleration it uses in much of the software used to showcase its power.

Again, i will repeat:

Yes, maybe Alder Lake-P will show up @ 35W against M1 Max @30W in CBR23 and produce those results. And that will show you a head-to-head comparison of how those products will perform in the market. But you aren't saying much about efficiency when you consider the amount of silicon area in use on Alder Lake-P vs. M1 Max to produce those results. M1 Max is mostly cache, iGPU, and dedicated hw.

Would you compare a 64c Milan downclocked to pull 142W to a 5900X and then make comparisons of efficiency? If so, you get to move to the back of the class.

Keep in mind I am anti-Apple, and I think their "walled garden" is garbage. But I'm not going to blind myself to poor comparisons, like trying to prove that a 12900k with disabled cores and power limits is somehow intrinsically more-efficient than a bone stock M1 Max running a benchmark that for whatever reason runs like crap on M1 Max (at least according to the Apple people anyway; CBR23 may not have proper Neon support, though it's hard to tell, would rather see standalone Blender or something tested to be sure).

The intended behavior from intel is Pl2 for tau seconds and then Pl1, even for the unlocked K parts.

So if Intel puts that in their design docs and nobody follows it, I guess we just blame the OEMs? Also, I thought PL1=PL2 now?

Intel Reportedly Doubles Alder Lake Power Consumption, PL1 is Now the Same as PL2 @ 241W | Hardware Times

Intel appears to have gone a bit overboard with Alder Lake, not only in terms of performance but also the power draw. As per data unearthed by our German friend Igor, the BIOSes of the upcoming Z690 motherboards have set the PL1 values equal to PL2, making the latter rather redundant. This...

www.hardwaretimes.com

 

[Zucker2k]

Non-K parts not having infinite turbo is a myth, the CPU type has no influence on turbo settings: it's all decided by the motherboard, at UEFI level. I've seen both Z and B chipset boards with unlocked power & turbo timer settings for non-K CPUs at stock. Everything comes down to VRM quality and whatever the manufacturer decides.

Here's a very clear example from ASRock, see how VRM quality has a massive influence on max power (and may have further effect on Tau):

View attachment 52598

BFB stands for Base Frequency Boost.

Here's Igor on the subject....

Intel Alder Lake Test: When motherboard settings affect the power consumption and temperature of the CPU | Practice | igor´sLAB

Are you still amazed about the so different power consumption values at the different launch reviews of Intel’s Alder Lake? In the meantime I have the data of a second motherboard model and I could…

www.igorslab.de

 

[TheELF]

The Intel 12th Gen Core i9-12900K Review: Hybrid Performance Brings Hybrid Complexity

www.anandtech.com

That's what I understood...

So if Intel puts that in their design docs and nobody follows it, I guess we just blame the OEMs? Also, I thought PL1=PL2 now?

Intel Reportedly Doubles Alder Lake Power Consumption, PL1 is Now the Same as PL2 @ 241W | Hardware Times

Intel appears to have gone a bit overboard with Alder Lake, not only in terms of performance but also the power draw. As per data unearthed by our German friend Igor, the BIOSes of the upcoming Z690 motherboards have set the PL1 values equal to PL2, making the latter rather redundant. This...

www.hardwaretimes.com

No, we blame reviewers that take super hardcore overclocking mobos, don't change any setting and then try to sell the performance but also the power draw as normal, stock, default, or whatever else.

Also intel didn't say PL1=PL2, intel said PL1=PL2=241w=maximum turbo power/power limits lifted, you can't just truncate half the equation and claim that that's what they said.
TDP is still PL1=125W,PL2=241W=processor base power/TDP
It's the same two power settings you could have with previous generations.

 

[igor_kavinski]

Intel Core i9 12900K "Alder Lake" AVX-512 On Linux - Phoronix

Serious AVX-512 gains in mining algorithms.

 

[Det0x]

Some testing done on Norwegian fourm:

Stock win 11 pro from a z490 install:

12900k(8c/16t)@5.3Ghz / 5Ghz ring
16GB DDR4 @4000 MT/s Cl15 Gear1
3090@2130ish Mhz



Testing 8c/16t vs 16c/24t

SoTR 8c/16t: 277fps vs 16c/24t: 279fps
CS:GO 8c/16t: 898fps vs 16c/24t: 950fps
Division 2 8c/16t: 315fps vs 16c/24t: 296fps
FarCry6 8c/16t: 188fps vs 16c/24t: 189fps
R6 Siege 8c/16t: 823fps vs 16c/24t: 816fps

He is getting DDR5 testing rig up and running next week which will be tested aginst DDR4 system, using the same SP90 CPU

*edit*

Some other highly tuned DDR5 system for comparison:

Dram 6200, 1.35v VDDQ 1.2



*edit 2*
Even 4100 can be done in gear1 mode. (slower uncore speed = worse latency then first ddr4 picture)

4100MT/s CL14 Gear1

 

[JoeRambo]

I wonder if 16c/24T is performing better in those tests due to AVX512 not being available. Cheaper context switches win out over potential AVX512 benefits that these games don't use at all.

Overall very good results for ADL + DDR4. I also saw some 4100 gear1 results on cheapest MSI Z690 results.

 

[DrMrLordX]

No, we blame reviewers that take super hardcore overclocking mobos, don't change any setting and then try to sell the performance but also the power draw as normal, stock, default, or whatever else.

Isn't that what the customer gets when they buy the product?

Overall very good results for ADL + DDR4. I also saw some 4100 gear1 results on cheapest MSI Z690 results.

At least this iteration of the DDR4 controller is better than what was in Rocket Lake-S.

 

[dmens]

Turbo/boost settings for ST workloads are far more aggressive than in MT, though. Plus you aren't necessarily saturating interconnect (where applicable) and caches, if that matters.

That is fine because if your goal is to chart core efficiency and how far some product is willing to go to score a benchmark win, you need to let it turbo/boost to its programmed range to plot the full curve. If it matters to you, prune out workloads that you believe are skewing results due to DRAM/caching behavior.

 

[lobz]

@JasonLD

Mac crowd is already calling out CBR23 for being "poorly optimized" on M1-series chips.

The Intel-x86-Apple-GeekBench-Cinebench free for all deathmatch has been very long in the making 😆
AMD, once a mere observer, has also found itself in the crossfire every now and then, but the aforementioned 2 ultra-elitist crowds can present a really good show if they take it (themselves) seriously enough.

 

[clemsyn]

My 12600K GeekBench 5 Score with 40% overclock with iGPU (Vulcan/OpenCL)

 

[pakotlar]

Alder Lake-S is designed to operate at 241W and beyond, and Intel's performance claims would be inaccurate without that power draw . . .

(also, you should look at how much of M1 Max's power spec is dedicated to the actual CPU cores, vs. other parts of the die).

> Alder Lake-S is designed to operate at 241W and beyond, and Intel's performance claims would be inaccurate without that power draw . . .

It's designed to operate in a variety of envelopes. The M1 Max is designed for the MacBook Pro 14/16 cooling system, power draw constraints, it's a single design (or 2 if you consider the 2 GPU configs). Alder Lake currently has at least 4 configurations, each with a different power point. There will likely be both 15W and 45W, and potentially 35W configurations of Alder Lake as well.


> Alder Lake-S is designed to operate at 241W and beyond, and Intel's performance claims would be inaccurate without that power draw . . .

The link in the post you're replying to does exactly that. It's up to ~45W drawn by the CPU in standard operation, depending on load.

 

[pakotlar]

Some testing done on Norwegian fourm:

Stock win 11 pro from a z490 install:

12900k(8c/16t)@5.3Ghz / 5Ghz ring
16GB DDR4 @4000 MT/s Cl15 Gear1
3090@2130ish Mhz

View attachment 52602

Testing 8c/16t vs 16c/24t

SoTR 8c/16t: 277fps vs 16c/24t: 279fps
CS:GO 8c/16t: 898fps vs 16c/24t: 950fps
Division 2 8c/16t: 315fps vs 16c/24t: 296fps
FarCry6 8c/16t: 188fps vs 16c/24t: 189fps
R6 Siege 8c/16t: 823fps vs 16c/24t: 816fps

He is getting DDR5 testing rig up and running next week which will be tested aginst DDR4 system, using the same SP90 CPU

*edit*

Some other highly tuned DDR5 system for comparison:

Dram 6200, 1.35v VDDQ 1.2
View attachment 52603
View attachment 52604

*edit 2*
Even 4100 can be done in gear1 mode. (slower uncore speed = worse latency then first ddr4 picture)

4100MT/s CL14 Gear1

Wow 51ns at 100GB/s; that's lower latency than my 3600mhz DDR4 2 * 32GB kit on X570.