Question Raptor Lake - Official Thread

[Hulk]

Since we already have the first Raptor Lake leak I'm thinking it should have it's own thread.
What do we know so far?
From Anandtech's Intel Process Roadmap articles from July:

Built on Intel 7 with upgraded FinFET
10-15% PPW (performance-per-watt)
Last non-tiled consumer CPU as Meteor Lake will be tiled

I'm guessing this will be a minor update to ADL with just a few microarchitecture changes to the cores. The larger change will be the new process refinement allowing 8+16 at the top of the stack.

Will it work with current z690 motherboards? If yes then that could be a major selling point for people to move to ADL rather than wait.

 

[H433x0n]

That certainly tracks if Steve's hypothesis is correct. Which means this is far from the killer app needed to stimulate sales. Unless we suddenly see dozens of newer titles get boosted in a similar fashion. These products have a rather limited shelf life, they have to get cracking.



They have been doing this for decades. I also would not be surprised if windows 12 breathes new life into hybrid CPUs by having better scheduling and optimizations built in. Wintel isn't going to suddenly stop being a thing.

It’s simpler than it seems and doesn’t require any input from game developers. It’s overriding thread director behavior to only enable a single e-core in a cluster so it has access to 4mb of L2$. So effectively a 13900K will have 8P+4E but with 4mb of L2$ for each e-core.

Some games it benefits to enable - some it doesn’t. It requires somebody going in and testing each game individually and white listing whether or not it’s beneficial. I’ve got no idea if they’ll keep up with it though. Also seems like 13th gen is arbritrarily blocked. You can make an argument it doesn’t make sense for 12th gen since only the 12900K has 2 e-core clusters and there are architectural changes between Alder / Raptor Lake (e-core cluster has less L2$ for Alder Lake for example) that may make this tweak impossible for 12th gen processors.

 

[igor_kavinski]

Intel really have no excuse to not at least support 13th gen, which is essentially the same exact part. There is absolutely no excuse for not supporting 13th gen, and probably 12th gen as well.

Their excuse is, they are INTEL. They can do whatever they want: https://videocardz.com/newz/intel-c...n-optimization-apo-on-12th-13th-gen-core-cpus

 

[ondma]

Their excuse is, they are INTEL. They can do whatever they want: https://videocardz.com/newz/intel-c...n-optimization-apo-on-12th-13th-gen-core-cpus

That is pretty much any company isn't it? If the consumers are unhappy, they can let it be known by not buying their products.

 

[DAPUNISHER]

GN tested APO too.

They will need to replicate that ARC team energy, to get a bunch of games supported.

 

[igor_kavinski]

What I would do is, change a bunch of different applications or game EXEs to the same filename as the supported game executables and see what happens when APO is forced on unsuspecting applications! I bet Intel didn't foresee anyone being this nefarious

P.S. I want that power interposer thingy!!!!

 

[igor_kavinski]

Interview - Intel's new APO tech and How It Boosts Gaming Performance on Intel Core 14th Gen

We sit down with Intel's Roger Chandler to find out how the new APO multi threading tech that boost game performance on the new 14th Gen CPUs works.

www.tweaktown.com

"We now understand that some threads are in high demand at one point but not so high demand at another. You don't want to tax the CPU with that because it could impact power and so forth," Roger says. "So, the next iteration is this. What APO does is we test the games and see, okay, this might benefit from a fine-tuning of the policy just because of the unique way this game behaves."

If there's a game that can benefit from Intel Application Optimization (APO), the policy is changed within the Thread Director. The game is simply added to the list, and PC gamers rocking the new Intel Core i9 14900K or 14700K get a sizable performance uplift - sometimes in the region of 20+% in specific workloads or scenarios compared to the top-end Ryzen CPU.

"The user does not have to do anything; there's nothing we have to do with the driver," Roger adds. "It's all just a policy thing we manage based on how the system interacts with the application. Not every application benefits from this; we test it and find what works flawlessly with Thread Director, and there are no issues."

Users can also turn the feature off via a Microsoft Store application, and when it comes to game updates, Intel works closely with developers to re-test as time goes on to ensure there aren't any issues. "We'll try to make it as seamless for the end-user as possible," Roger explains. Currently, the Intel team working on Intel Application Optimization (APO) is focused on games. Still, it could expand to other workloads, such as those in the creative space.

"Thread Director and our dynamic tuning technology policies are pretty elegant," Roger says. "They cover 99% of most use cases, but games differ. They have spiky behavior for things that happen, so that's how APO works. If we find the opportunity, we apply it. If the game doesn't need it, great. We don't put it on the list."

So the bad news is, these gains can't be had with every game or application. APO can only benefit workloads that are too unpredictable for the Thread Director.

 

[biostud]

Interview - Intel's new APO tech and How It Boosts Gaming Performance on Intel Core 14th Gen

We sit down with Intel's Roger Chandler to find out how the new APO multi threading tech that boost game performance on the new 14th Gen CPUs works.

www.tweaktown.com

So the bad news is, these gains can't be had with every game or application. APO can only benefit workloads that are too unpredictable for the Thread Director.

Or rather the good news is that most games work very well with Windows and only a few have problems that APO can fix.

 

[igor_kavinski]

Ah, finally someone realized that this needed to exist: https://bitsum.com/apps/coredirector/

The Efficiency Mode off option is brilliant! No need to disable E-cores!

 

[Timur Born]

Ah, finally someone realized that this needed to exist: https://bitsum.com/apps/coredirector/

The Efficiency Mode off option is brilliant! No need to disable E-cores!

Hm, actually the other way around is much more useful/needed. I use Process Lasso "Sets" to push almost all background processes/services to E cores, which leaves P cores free to boost and be used for more important things. "Sets" even allow to push those services to E cores that cannot be moved via "Affinities". The current description of "Sets" is a bit misleading, though, because Windows doesn't seem to really "decide" much about them yet (aka seems as fixed as Affinities). I hope that they will get more "decision making" in future iterations of Windows.

• CPU Sets
  This method is somewhere in between the other two and leaves a lot of the decision to the OS.

Now why do I say that "Efficiency Mode off" isn't really needed/useful, except maybe for niche cases like running memory stability tests that spawn dozens of background processes?

Windows current logic for pushing background load to E cores is as follows (based on my own tests):

- A process is not in focus (or has at least a good part of its window hidden for a time) while doing Integer calculations, because E cores are efficient for Integer. Processes doing Floating-Point calculations are *not* pushed to E cores, because E cores are inefficient for Floating-Point.

- A process sets its own process priority to "low". This affects both Integer and Floating-Point calculations then! To disable this for a particular program (only) you can manually set the process priority to "Above normal" or higher.

Switching Windows' power MODE (aka classic "Balanced" PROFILE + new MODE) to "Max Performance" keeps Windows from pushing all/multi-core background processes to E cores, no 3rd party software needed! It might still push single/low-core processes to E cores then, though.

Overall having Windows push background/low priority processes to E cores is a good thing! The logic may be a little too simplistic, but power-users can use Process Lasso or power-profile/priority Windows settings to accommodate for their needs until Windows get better with it (which likely includes improving the "Sets" feature to do real decision making).

 

[Timur Born]

My 13900K's all-core 5.8 GHz VIDs with all P- and C-states disabled and AC/DC LL 0.010 mOhm:

 

[Hulk]

My 13900K's (interpolated) all-core 6.0 GHz VIDs with all P- and C-states disabled:

View attachment 89542

That's a very nice bin. Looks to be SP probably around 103 with P SP of about 114. Not a God bin but I'd say a Golden bin!
Good for you!
What are your day-to-day settings? What is Vcore under load at max all-core frequency you're running? What's your cooling?
I had a 14900K when they first came out and it was the worst possible 1.504V at 6GHz. I returned it and just stayed with my 13600K.

 

[Timur Born]

That's a very nice bin. Looks to be SP probably around 103 with P SP of about 114. Not a God bin but I'd say a Golden bin!
Good for you!
What are your day-to-day settings? What is Vcore under load at max all-core frequency you're running? What's your cooling?
I had a 14900K when they first came out and it was the worst possible 1.504V at 6GHz. I returned it and just stayed with my 13600K.

I had to correct this. The 13900K uses the same maximum VID for 6.0 GHz that it uses for 5.8 GHz, so it's 1.434 V at 5.8 GHz only. I got the impression that this is about 105 SP equivalent or so. My Gigabyte Biscuit rating started at 89.9 and decreased to 86.7 with the KS BIOS upgrade. Interpolating from this I assume it would ask for about 1.49-1.5 V for 6 GHz, which would put it in the lower KS/14th tier.

So my CPU is somewhere between middle of the pack to "good enough". Which is not a bad thing, because that means that many people should be able to reproduce my undervolting (+per core OC) settings somewhat easily.

Currently I am using AC LL 0.180 mOhm, LLC about 0.620 mOhm, 8-56x offset -50 mV and 57-60 offset -75 mV. Turbo ratios are 60/59/57/57/56/56/55/55..44/44/43/43 with per core ratios 55/55/60/59/56/56/57/57../43/43/45/44 and a -1 AVX offset (aka 54/54/59/58/55/55/56/56 per core AVX ratios). CB23 get 40-40.5k scores, depending a bit on the CPU temperature, slightly throttling at the 253 W power limit at higher temps. All realworld load usually runs below the 253 W power limit and 1-core CB load at 6.0 GHz runs below 40 W package power.

 

[Hulk]

I had to correct this. The 13900K uses the same maximum VID for 6.0 GHz that it uses for 5.8 GHz, so it's 1.434 V at 5.8 GHz only. I got the impression that this is about 105 SP equivalent or so. My Gigabyte Biscuit rating started at 89.9 and decreased to 86.7 with the KS BIOS upgrade. Interpolating from this I assume it would ask for about 1.49-1.5 V for 6 GHz, which would put it in the lower KS/14th tier.

So my CPU is somewhere between middle of the pack to "good enough". Which is not a bad thing, because that means that many people should be able to reproduce my undervolting (+per core OC) settings somewhat easily.

Currently I am using AC LL 0.180 mOhm, LLC about 0.620 mOhm, 8-56x offset -50 mV and 57-60 offset -75 mV. Turbo ratios are 60/59/57/57/56/56/55/55..44/44/43/43 with per core ratios 55/55/60/59/56/56/57/57../43/43/45/44 and a -1 AVX offset (aka 54/54/59/58/55/55/56/56 per core AVX ratios). CB23 get 40-40.5k scores, depending a bit on the CPU temperature, slightly throttling at the 253 W power limit at higher temps. All realworld load usually runs below the 253 W power limit and 1-core CB load at 6.0 GHz runs below 40 W package power.

Thanks for the info, what is your Vcore from HWinfo when running CB at those settings?

 

[Timur Born]

Thanks for the info, what is your Vcore from HWinfo when running CB at those settings?

1.242 V (average over 10 minutes) for all-core scores of 40k+ at the 253 W power limit on the more reliable sensor 2. That sensor is closer to the CPU, shows more droop and seems to be used by Gigabyte for their stock DC LL + LLC setting to match Vcore to VIDs. It an extra sensor on the Aero G that the Aorus Elite doesn't offer, but it doesn't seem to be a VOUT sensor like the Aorus Master offers.

 

[Hulk]

1.242 V (average over 10 minutes) for all-core scores of 40k+ at the 253 W power limit on the more reliable sensor 2. That sensor is closer to the CPU, shows more droop and seems to be used by Gigabyte for their stock DC LL + LLC setting to match Vcore to VIDs. It an extra sensor on the Aero G that the Aorus Elite doesn't offer, but it doesn't seem to be a VOUT sensor like the Aorus Master offers.

That's a safe voltage to avoid any possibility of degradation. I think setting the 253W power limit is a good idea too. Any additional power will bring negligible performance at the expense of heat, power, and possible long term cpu damage. You have that rig well set up.

 

[Timur Born]

That's a safe voltage to avoid any possibility of degradation. I think setting the 253W power limit is a good idea too. Any additional power will bring negligible performance at the expense of heat, power, and possible long term cpu damage. You have that rig well set up.

Thank you!

253 W is just at the edge of slightly throttling CB23 when CPU temps increase, stock was closer to 290 W. I still get full 40k+ scores there and in practice all realworld applications use less power than that. So the power-limit throttles power-virus load, which in turn allows to use lower voltages without losing *any* performance in realworld applications.

I optimize my undervolted (!) CPU for stability up to the power-limit and optimize for *realworld* performance. I do not optimize for power-virus "performance", but only use stability tests to test stability. Increasing Vcore to keep power-virus load stable above the realworld power-limit is drowning the CPU in voltage/heat/cooling-noise for no performance benefits whatsoever.

And if someone fears to run into power/temp/current limits regularly (or at all) then I advice to enable Ring downbin. This will keep core clocks running at considerably higher frequency in return for slightly lower Ring frequency (up to 300 MHz, but usually lower). This only kicks in when the CPU runs into a limit, though, it's *not* a permanent downbin.

Single-core CB23 at 6.0 GHz uses about 1.34 V (aka much less droop happening due to light load) at 38 W on my CPU. That only works on my best core 2, though.

Unfortunately all those overclocking minded forums/Discords still keep propagating to not use an AVX offset on modern Intel CPUs. Bad advice, as the offset doesn't trigger 99% of the time anyway and being adamant to using a zero (0) offset either increases voltage demands for no practical benefit or leaves extra performance on the table.

 

[Hulk]

Thank you!

253 W is just at the edge of slightly throttling CB23 when CPU temps increase, stock was closer to 290 W. I still get full 40k+ scores there and in practice all realworld applications use less power than that. So the power-limit throttles power-virus load, which in turn allows to use lower voltages without losing *any* performance in realworld applications.

I optimize my undervolted (!) CPU for stability up to the power-limit and optimize for *realworld* performance. I do not optimize for power-virus "performance", but only use stability tests to test stability. Increasing Vcore to keep power-virus load stable above the realworld power-limit is drowning the CPU in voltage/heat/cooling-noise for no performance benefits whatsoever.

And if someone fears to run into power/temp/current limits regularly (or at all) then I advice to enable Ring downbin. This will keep core clocks running at considerably higher frequency in return for slightly lower Ring frequency (up to 300 MHz, but usually lower). This only kicks in when the CPU runs into a limit, though, it's *not* a permanent downbin.

Single-core CB23 at 6.0 GHz uses about 1.34 V (aka much less droop happening due to light load) at 38 W on my CPU. That only works on my best core 2, though.

Unfortunately all those overclocking minded forums/Discords still keep propagating to not use an AVX offset on modern Intel CPUs. Bad advice, as the offset doesn't trigger 99% of the time anyway and being adamant to using a zero (0) offset either increases voltage demands for no practical benefit or leaves extra performance on the table.

You are soooo right about optimizing for actual applications rather can CPU degrading benches. Outside of things like Handbrake you'll probably never see power/temps near 250W but you will see high clocks so you're getting the performance you paid for.

 

[Timur Born]

Those degrading benches still have to be stable up to the power-limit, though. At least for my own "as stable as stock", aka "it's only stable when it's stable" goals. ;-)

Handbrake creates one of those super demand load scenarios where the average power for some 4K AV1 videos is well below 200 W and H.265 is below 240 W. Even at a high polling rate the maximum power peaks can hardly be measured to hit the 253 W power-limit by HWinfo. But it still regularly creates transients that trigger the PL register of the CPU and would show up on an oscilloscope. So some encoding tasks are kind of throttled for peak, but for the most part they are not and even for the peaks I wonder if the PL throttling catches up fast enough.

Curiously AV1 encoding is pushed to E cores when done in the background, H.265 encoding is not. That would suggest that the former is Integer load and the latter Floating-Point load.

My way of stability testing for these kind of loads - below the power-limit, but with crazy transients - is to use Folding@home (and Handbrake). FAH is the bane of stability testing for me! Not only because at the same voltage it catches instabilities that all others don't, but also because it takes a loooong time to find edge cases, because each work unit creates different load and the units are chosen randomly.

One of the most demanding realworld applications is: Windows Defender! Part of its Full Scan is so superbly multi-threaded that it hits the same power usage (over 230 W average) as Handbrake 4K H.265 when allowed to run on P cores, just without the crazy transients. It achieves the highest throughput of all anti-virus scans I tested on my PCIe 3.0 M.2 SSD.

Defender Full Scan:

 

[jpiniero]

Don't think it got mentioned much but Intel finally bothered to make Xeon E branded Raptor Lake. P cores only.

Oh and if you were wondering, AVX-512 is still disabled.

 

[DrMrLordX]

Oh and if you were wondering, AVX-512 is still disabled.

Sad. How many cores on the Xeon-E? 8?

 

[jpiniero]

Sad. How many cores on the Xeon-E? 8?

4, 6 or 8.

 

[DrMrLordX]

4, 6 or 8.

So they just disabled the e-cores and called it a day.

 

[FlameTail]

This is something I have been wondering for a long time:.

How much does Intel charge laptop companies for their processors?

 

[igor_kavinski]

How much does Intel charge laptop companies for their processors?

Intel® Core™ i7-1365U Processor (12M Cache, up to 5.20 GHz) - Product Specifications | Intel

Intel® Core™ i7-1365U Processor (12M Cache, up to 5.20 GHz) quick reference with specifications, features, and technologies.

www.intel.com

I read years ago that CPU manufacturing cost is about $50 per CPU for fabs. Lets assume due to inflation, it has risen to $100. So anything above that would be profit for Intel/AMD, especially with a volume discount.

 

[FlameTail]

Intel® Core™ i7-1365U Processor (12M Cache, up to 5.20 GHz) - Product Specifications | Intel

Intel® Core™ i7-1365U Processor (12M Cache, up to 5.20 GHz) quick reference with specifications, features, and technologies.

www.intel.com

View attachment 90578
I read years ago that CPU manufacturing cost is about $50 per CPU for fabs. Lets assume due to inflation, it has risen to $100. So anything above that would be profit for Intel/AMD, especially with a volume discount.

Oh wow. They actually list it in their website?

And that thing costs $430. That's like 1/3 the price of a laptop with it!