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.

 

[coercitiv]

Also, behold the new branding for Raptor Lake U:

Intel’s CPU branding was already confusing, and today’s new CPUs made it worse

Some are 14th-gen Core and some are Core (Series 1), but they're the same thing.

arstechnica.com

 

[Doug S]

Also, behold the new branding for Raptor Lake U:
View attachment 91614

Intel’s CPU branding was already confusing, and today’s new CPUs made it worse

Some are 14th-gen Core and some are Core (Series 1), but they're the same thing.

arstechnica.com

They need the word Core in there twice to catch up with "series" and "processors"

 

[lightmanek]

They need the word Core in there twice to catch up with "series" and "processors"

Would something like this work?
"Intel Core Processors (Series-1) U-series Processors Core 1xx"

 

[Hulk]

I realize I previously stated I wasn't going to play the 14900K lottery after I returned one to Amazon but I broke down and bought one from Microcenter with the intention that no matter what I got I would keep it.

Since my 13900K degraded after only a few months running auto settings I decided I'd play it safe with the 14900K and not even test the limits of this CPU. All I wanted was 5500MHz for the P's and 4400MHz for the E's with HT turned off. I don't have any applications that require more than 24 physical cores so other than reporting big CB numbers I can do without HT given the number of actual cores on the 14900K.

After some experimenting I settled on the following.
Manual voltage setting of 1.3V with LLC4.

This setting keeps the CPU running less than 1.3V at all times and under heavy load at about 1.15V. Power and thermals are quite good at this setting as well. CB R24 55/44 pulls about 206W package power. The CB score is only about 34,000 because HT is off but it's still a decent score. I'm losing about 5,400 CB points with HT off but AFAIK I don't run any apps that utilize that many threads anyway. Doing some quick napkin math it looks like Intel would only have to add about 1 additional E core cluster to make up for HT. This leads me to believe HT is going the way of the dinosaur for 3 reasons and I see quite clearly why there are rumors of Intel killing HT in client.

1. E cores are more area efficient than P's for MT. That die area is better spent on E's if the point is to increase MT performance.
2. P's can be designed solely with ST performance in mind and additional transistors previously used for HT can be designated to increase ST performance.
3. As nodes get smaller thermal density is becoming more of an issue. It's not that the total thermal load is unmanageable, but the density of the hotspots is hard to control with ambient cooling solutions (ie need for higher heat transfer coefficients to increase flux). By in effect moving the MT performance to additional E cores, away from the P cores, that additional heat moves away as well. This is of course one of the reasons I turned of HT on my current CPU.

Thinking through this more deeply, let's assume my applications 99% of the time would not saturate all cores with HT on. But for that 1% of the time they did my current voltage setting would not work and I get a lock up. Would the performance benefit for that 1% total time be worth having to increase voltage and heat across the system when 99% of the time things would be fine at the lower voltage? I don't think so.

Now I fully understand my thought experiment is based on the arbitrary values of 99% and 1%. If in reality those numbers are 60% and 40% then the analysis changes completely. But I am quite sure that outside of artificial benchmarks my apps don't saturate more than 24 threads. This is especially true now that some of my more compute heavy applications like Topas Photo/Video AI, PureRaw, and Vegas Pro are utilizing the GPU quite a bit.

Anyway, I thought this might be interesting for anyone with a LGA1700 system considering a Raptor Refresh part.

 

[coercitiv]

Now I fully understand my thought experiment is based on the arbitrary values of 99% and 1%. If in reality those numbers are 60% and 40% then the analysis changes completely. But I am quite sure that outside of artificial benchmarks my apps don't saturate more than 24 threads. This is especially true now that some of my more compute heavy applications like Topas Photo/Video AI, PureRaw, and Vegas Pro are utilizing the GPU quite a bit.

I think your thought experiment has bigger issues than setting up arbitrary values like 99% and 1%.

You argue HT increases voltage requirements at ISO frequency, have you considered that HT gains can be partially traded for lower frequency and thus lower power requirements? In other words, have you considered ISO power instead of ISO clocks? SMT roughly increases perf and power linearly, while dropping frequency has a non-linear effect on power. So what happens if you leave HT enabled for that CB run, but enforce a 200W limit instead?

You say 24 threads is the upper bound for your workloads. In your case this also happens to be 24 cores. If that is truly the case wouldn't you be better served by less but uniformly faster cores than the ones in your CPU? There's a consumer product on the market with 16 cores of comparable performance with your P cores. For the answer to be "no", your workloads would have to be uniquely shaped as to saturate 16P cores, strongly leverage 8 more threads, then fall of a cliff from 24 threads onward. I would find that very peculiar.

Ever since I bought my 12700K, I have used it in various configs for many months at a time: 8+0 /w HT, 8+4 /w HT, and 8+4 w/o HT. I usually go for 3-6 months before making a change, and I have never been able to subjectively tell the difference between these configs. I'd be curious to know if your subjective experience with the 14900K degrades when running in 8+8 configuration, for a total of 16 cores with or without SMT enabled.

 

[Hulk]

I think your thought experiment has bigger issues than setting up arbitrary values like 99% and 1%.

You argue HT increases voltage requirements at ISO frequency, have you considered that HT gains can be partially traded for lower frequency and thus lower power requirements? In other words, have you considered ISO power instead of ISO clocks? SMT roughly increases perf and power linearly, while dropping frequency has a non-linear effect on power. So what happens if you leave HT enabled for that CB run, but enforce a 200W limit instead?

You say 24 threads is the upper bound for your workloads. In your case this also happens to be 24 cores. If that is truly the case wouldn't you be better served by less but uniformly faster cores than the ones in your CPU? There's a consumer product on the market with 16 cores of comparable performance with your P cores. For the answer to be "no", your workloads would have to be uniquely shaped as to saturate 16P cores, strongly leverage 8 more threads, then fall of a cliff from 24 threads onward. I would find that very peculiar.

Ever since I bought my 12700K, I have used it in various configs for many months at a time: 8+0 /w HT, 8+4 /w HT, and 8+4 w/o HT. I usually go for 3-6 months before making a change, and I have never been able to subjectively tell the difference between these configs. I'd be curious to know if your subjective experience with the 14900K degrades when running in 8+8 configuration, for a total of 16 cores with or without SMT enabled.

I like your questions/analysis.

I have not run the test with CB and ISO power as you have suggested but I agree that there is a good chance that for a perfectly multithreaded application like Cinebench that can use every thread available that there would be various power level settings there HT would be more efficient than non HT. In general, as we know, when talking about well threaded apps like CB, doubling the cores and cutting frequency in half is always the more efficient option. The fact that the v/f is nonlinear makes that obviously true. But unfortunately reality has that unfortunate way of asserting itself and few applications are as well threaded as CB, and more importantly none that I use.

As or your second thought, wouldn't more uniform cores be better than a hybrid approach? The answer is that it depends on the application. Most of my applications seems to hit 5 or 6 cores really hard and then after that core usage backs off quite a bit.

But I get your point, maybe we both would have been better served with Zen 4. Maybe next round. Right now I'm very happy with running 55/44 at reasonable power levels.

As for "noticing" differences in performance with HT on and lower clocks or HT off and higher clocks I have done some testing. Studio One will do better with the P's running at higher frequency (no HT). Same with Vegas Pro frameserving. Alas most of my bottleneck is still ST performance.

Is it enough difference to "feel?" Probably not, but I do see a project in Studio One updating at 10x speed rather than 9x with HT, so I do know I'm saving a little time.

 

[Khato]

I was curious about the CB23 performance of HT on/off and P vs E core at various per-core power levels a few weeks ago with a 14900k. So I ran three different core configurations of 1P+16E (processlasso used to constrain CB23 to the E cores only), 8P+0E no HT, and 8P+0E with HT at 18W, 33W, 65W, and 120/150W PL1. The IA Cores Power readout always showed 1-2W below the selected PL1, so used that for approximating the per core power.

Results of HT on/off matched coercitiv's explanation exactly as expected. At low power levels HT is showing minimal benefit as we're in the favorable power vs frequency range, eg 1.5x the power for 1.4x the frequency. Whereas at higher power levels you're paying a lot more power for small gains in frequency, eg1.5x the power for 1.05x the frequency. By comparison HT is always 1.5x the power for 1.5x the performance. (Note that these examples are by no means exact/correct, just illustrative of the concept.)

Anyway, figured I'd share these as they seem pertinent to the discussion. A fun tangent to go off on with them though is how such applies to the server realm where they're not likely to even reach the 8W/C level. The E cores running in the 1-2W/C range are looking quite good.

Config	Score	Score/Core	Score/Thread	HT % Increase	P/E Same W/C
16E 1W/C, 18W PL1	7202	450.13	450.13
8P/8T 2W/C, 18W PL1	4386	548.25	548.25		0.77
8P/16T 2W/C, 18W PL1	4715	589.38	294.69	1.08	0.83
16E 2W/C, 33W PL1	11334	708.38	708.38
8P/8T 4W/C, 33W PL1	7577	947.13	947.13		1.01
8P/16T 4W/C, 33W PL1	8901	1112.63	556.31	1.17	1.18
16E 4W/C, 65W PL1	15031	939.44	939.44
8P/8T 8W/C, 65W PL1	11868	1483.50	1483.50		1.30
8P/16T 8W/C, 65W PL1	14521	1815.13	907.56	1.22	1.59
16E 7.5W/C, 120W PL1	18268	1141.75	1141.75
8P/8T 18W/C, 150W PL1	15825	1978.13	1978.13
8P/16T 18W/C, 150W PL1	20621	2577.63	1288.81	1.30

 

[Timur Born]

Concerning degradation on "Auto" settings: One thing to keep in mind is that many mainboards set both CPU package power (PL1/2) and peak current (ICCmax) limits to "unlimited". Even when you run a load that measures as 210-230 W package power there can be short transients hitting over 253 W and/or over 400 A (don't even think about 307 A).

 

[igor_kavinski]

Even when you run a load that measures as 210-230 W package power there can be short transients hitting over 253 W and/or over 400 A (don't even think about 307 A).

Can you please show screenshots of something like HWinfo and point out the values that we can check to ensure that our systems are not running beyond spec? Does Intel XTU show anything in RED or warn the user if something is running outside safe limits?

 

[Hulk]

Concerning degradation on "Auto" settings: One thing to keep in mind is that many mainboards set both CPU package power (PL1/2) and peak current (ICCmax) limits to "unlimited". Even when you run a load that measures as 210-230 W package power there can be short transients hitting over 253 W and/or over 400 A (don't even think about 307 A).

I had my PL1/PL2 set to 225W with the 13900K but Auto setting will still throw crazy high voltage at the CPU even though it's not needed.

There is debate whether it is voltage, current, or temp that kill/degrades a cpu. I think any of them can do it so I limit voltage to 1.3V, power (current as well by Ohms Law) to 225W, and temps to 80C. I feel these are safe numbers for a daily work PC that might render or do heavy compute for hours on end. Time will tell if my current rig holds up at these settings...

 

[igor_kavinski]

There is debate whether it is voltage, current, or temp that kill/degrades a cpu. I think any of them can do it so I limit voltage to 1.3V, power (current as well by Ohms Law) to 225W, and temps to 80C.

Great. Seem like pretty sane settings.

 

[Timur Born]

Can you please show screenshots of something like HWinfo and point out the values that we can check to ensure that our systems are not running beyond spec? Does Intel XTU show anything in RED or warn the user if something is running outside safe limits?

For testing run OCCT Linpack 2021 at 204 mb data size. Both HWinfo and XTU show the corresponding CPU triggers which do *not* depend on polling rate. The CPU sets the trigger and they are cleared after software reads them out, which means that running HWinfo and XTU at the same time can lead to some conflict (one software may "steal" the readout of the other).

This is at 253 W PL limits + 415 A ICCmax (needed so high to allow both limits to trigger at the same time, else ICCmax is faster to trigger than PL) on my undervolted (!) 13900K:



Folding@home creates this type of load for many of its work units and Handbrake 4K H.264 does as well (and likely is the most "realworld" example).

 

[Timur Born]

I experimented a lot with ICCmax lately, using it to stabilize these kind of high transients load where the PL isn't enough to tame them. It did indeed work to a certain degree. Unfortunately I still had to increase my all-core voltage offset by 5 mV, because I noticed that Cinebench v15 load was unstable - spamming the "Run" button) while not triggering *any* limit at all (temp/TVB, power, current, AVX offset).

My TVB downbin -1 is at 86°C btw and I use an AVX offset of -1, which is not triggered by most realworld load (including CB) anyway and even then only affects two of my 8 P cores for all-core load. 13/14th gen users should *really* make use of the AVX offset to either allow for higher OC or lower voltages.

 

[Hulk]

When I replaced the 13600K for the 14900K I noticed that one of the tabs on the mount for the AIO was broken. I had noticed temps higher than they should have been. I read some great reviews on the Peerless Assassin 120 so for $42 gave it a try. What a beast of a cooler! Better than my Noctua U12A at nearly 1/3 of the price.

I sold my 13600K on e-bay and the seller lodged a "DOA" complaint. Strange since I pulled it from my computer working and it somehow got broken in the mail? You know e-bay, they force the return. I told the guy to make sure he sends me back the chip I sent him with the S/N and everything as described in the auction. If it is broken he smoked it overclocking I bet. e-bay did tell me if the chip comes back broken they will reimburse me for the e- bay sale price. I'll let you guys know how this goes. About 20 years ago I was scammed by a guy who bought my P4 3.06, claimed it was broken and then sent me back a different broken part. After that I learned to include photos and description of the specific part you are selling. e-bay reimbursed me half of the sale price on that one after I calmly told them I was scammed.

 

[Rigg]

When I replaced the 13600K for the 14900K I noticed that one of the tabs on the mount for the AIO was broken. I had noticed temps higher than they should have been. I read some great reviews on the Peerless Assassin 120 so for $42 gave it a try. What a beast of a cooler! Better than my Noctua U12A at nearly 1/3 of the price.

I sold my 13600K on e-bay and the seller lodged a "DOA" complaint. Strange since I pulled it from my computer working and it somehow got broken in the mail? You know e-bay, they force the return. I told the guy to make sure he sends me back the chip I sent him with the S/N and everything as described in the auction. If it is broken he smoked it overclocking I bet. e-bay did tell me if the chip comes back broken they will reimburse me for the e- bay sale price. I'll let you guys know how this goes. About 20 years ago I was scammed by a guy who bought my P4 3.06, claimed it was broken and then sent me back a different broken part. After that I learned to include photos and description of the specific part you are selling. e-bay reimbursed me half of the sale price on that one after I calmly told them I was scammed.

These are even better. It's a newer design. The PA is still very good though. I've tried a few different iterations of each and they've all performed well. The fans are where they save the money. Even if you put some nicer fans on them they're still a great value though. The Phantom Spirit 120 EVO at least has the more upscale gentle typhoon clones that the lower price Thermalright dual 120 tower coolers lack. For an extra $5-10 it seems like the way to go. As a bonus it is also completely blacked out if that suits your aesthetic taste.

Hopefully the ebay dude doesn't pull an IHS swap on you. You never know with ebay buyers and sellers. Ebay is like a scumbag tractor beam. I wish you the best.

 

[Hulk]

I know what you mean about e-bay. I've had over 500 transactions and all but 3 have been smooth. If this CPU comes back and it is the one I sent and it's DOA then I really have to just accept it even though it's very unlikely it was damaged in shipping in it's retail packaging and then shipping inside another box with packing material. Whatever, it's not a big enough deal to waste time and thought on.

That cooler you linked is nuts! Like $45 and even better than my PA. Both air and AIO's have gotten ridiculously good and cheap over the last few years. The AIO manufacturers are going to have to seriously up their game to compete with this new breed of super cheap and good air coolers. Seems like the only way to do significantly better than a top air cooler these days is using an expensive custom loop.

Also, if Arrow Lake doesn't burn the watts like Raptor then air is all anyone outside of extreme overclockers are going to require I would think.

 

[Rigg]

I know what you mean about e-bay. I've had over 500 transactions and all but 3 have been smooth. If this CPU comes back and it is the one I sent and it's DOA then I really have to just accept it even though it's very unlikely it was damaged in shipping in it's retail packaging and then shipping inside another box with packing material. Whatever, it's not a big enough deal to waste time and thought on.

That cooler you linked is nuts! Like $45 and even better than my PA. Both air and AIO's have gotten ridiculously good and cheap over the last few years. The AIO manufacturers are going to have to seriously up their game to compete with this new breed of super cheap and good air coolers. Seems like the only way to do significantly better than a top air cooler these days is using an expensive custom loop.

Also, if Arrow Lake doesn't burn the watts like Raptor then air is all anyone outside of extreme overclockers are going to require I would think.

I really don't like selling online and shipping anymore because of the obnoxious selling fees and scammers. Especially since the IRS started making payment services issue 1099 forms. I'll pass on that hassle. Everything about it sucks. I strongly prefer to put complete gaming rigs together and flip them for cash locally when i have parts to move.

I hope it all works out for you. Some people are just pieces of ....


The CPU cooler market has gotten a bit more interesting lately. I've been using Deep Cool, Thermalright, and Scythe for years. All 3 brands have made good quality yet cost effective coolers for quite some time now. They all seem to have stepped up their performance game a bit more the last few years (especially Thermalright) and more people have noticed.

At this point I'm kind of over custom water cooling. The juice ain't worth the squeeze IMO. I do still have a custom loop rig though. I'm keeping it around to use as my DAW PC since It's pretty hard to get rid of used water cooling stuff unless you pretty much give it away. I've long since switched to air cooling for my main gaming PC.

 

[SteinFG]

New rumor, from redgamingtech

Quick summary: Bartlett lake-S, New desktop line of processors for LGA 1700, most likely Raptor lake Refresh Refresh, aimed at a lower end market, as it seems like ARL-S and LGA 1851 platform will cost a lot.

 

[jpiniero]

Quick summary: Bartlett lake-S, New desktop line of processors for LGA 1700, most likely Raptor lake Refresh Refresh, aimed at a lower end market, as it seems like ARL-S and LGA 1851 platform will cost a lot.

RGT?

Bartlett Lake would imply that they are doing something like Rocket Lake. Which for only low end desktops is not worth the effort.

 

[SteinFG]

RGT?

Bartlett Lake would imply that they are doing something like Rocket Lake. Which for only low end desktops is not worth the effort.

Bartlett Lake doesn't mean anything. Skylake got released under many names (Kaby, Comet, Coffe), and it didn't stop being skylake. Bartlett seems to be another Golden Cove Processor.

 

[DrMrLordX]

If they lower the price enough, Bartlett Lake could be good for the budget sector, especially if people learn to detune power usage. Otherwise it'll be like Pentium D all over again.

 

[uzzi38]

Just take a look at mobile if you want to know what Intel are doing.

 

[SteinFG]

Just take a look at mobile if you want to know what Intel are doing.

We're getting Ultra (new and shiny) and non-ultra (old and renamed) naming on desktop too? heh

 

[coercitiv]

We're getting Ultra (new and shiny) and non-ultra (old and renamed) naming on desktop too? heh

Intel Core Processors Series 1

Intel® Processor Names, Numbers and Generation List

Understanding Intel® processor names and numbers helps identify the best laptop, desktop, or mobile device CPU for your computing needs.

www.intel.com

 

[igor_kavinski]

They hired ex-AMD marketing. Ah crap!