Discussion Intel current and future Lakes & Rapids thread

[IntelUser2000]

Aside from that the top Tremont cores get about ~750 on Geekbench for ST: https://browser.geekbench.com/v5/cpu/search?utf8=✓&q=n6005 . An expected +30% on top of that is ~1000 which is pretty close to the original Skylake (core 6000 series), but current rocket lake is closer to 1800.

Here are the comparisons of some CPUs. Tremont, Goldmont Plus refresh, and Goldmont Plus.

GB5 Integer
N6000: 702
N5030: 520
N5000: 488
6700K: 1171 @ 4.6GHz
11900K: 1673

I expect the 10W version of Tremont to be few % faster and get 750. Such 10W chips are more representative of the Gracemont that will be in Alderlake as it won't be power bound as the 6W mobile chips are.

Still, overall you are right. I believe it can beat 5900X, but not 5950X.

Golden Cove = 1.5x Skylake/Gracemont
Hyperthreading = 1.2x
Clock speed = 1.3x

You are looking at each Gracemont core being equal to about 40% of the Golden Cove core in MT performance.

I wonder if this prediction is as accurate as your "Tigerlake graphics performing about 30-40% better than Icelake" last year, it was a bit off.

Actually in that HW Unboxed review it's not far off. Yes I know they do better, but when you parade that it'll be so good for so long, you expect much more. Not to mention they still suffer from driver issues. It's not just under Batman! It's not a make or break issue but a noticeable downgrade from AMD/Nvidia.

Also I don't believe Tigerlake is bandwidth bound as you think. I think the CPU cores are too power hungry which is why Renoir gets the advantage under 15W.

LPDDR5-5400 is a mere 25% bandwidth improvement. If it was 100% bound by bandwidth, then it's a possible 25% gain. If that's a 60% gain(which is a LOT) then we'll get 15%.

Remembrandt I think will be a big deal on the iGPU side and I expect 60-80% improvement as well.

 

[Exist50]

In terms of Alder Lake multithreaded performance, and Gracemont vs Golden Cove, need to consider sustained vs burst performance. Peak clock speeds are only relevant for short loads. For sustained, you need to also factor in how much power the core consumes. That's probably going to significantly hurt Golden Cove and relative help Gracemont.

 

[IntelUser2000]

In terms of Alder Lake multithreaded performance, and Gracemont vs Golden Cove, need to consider sustained vs burst performance.

A great example here is Zen3, which got like +19% IPC and >+25% ST performance but only <+10% MT performance from the 3950x to the 5950x.

In case of Zen 3, it's probably limited by power/thermals. Golden Cove being bigger will be offset by the more efficient 10nm process.

I hope the uarch gains are closer to 25% and the clock speed goes down to 5GHz for the claimed "up to 20% gain". Therefore getting power under control.

When is Tiger Lake-H coming? Feels like a perpetual wait.
Also, is it going to be replaced by Alder Lake-S BGA or -P?

Both it seems. 45W is replaced by -P, while 55W is -S BGA.

The latter being based on -S tells me Alderlake desktop is really a souped up mobile chip, because they are desperate to move to a new process. It'll probably put Intel in a better position than they are now. Especially since we know Zen 3's successor is Zen 3+.

 

[Exist50]

In case of Zen 3, it's probably limited by power/thermals. Golden Cove being bigger will be offset by the more efficient 10nm process.

Don't forget uncore power. Not sure how AMD vs Intel compare, but I think both are pretty bad compared to e.g. Apple or Qualcomm.

Both it seems. 45W is replaced by -P, while 55W is -S BGA.

The latter being based on -S tells me Alderlake desktop is really a souped up mobile chip, because they are desperate to move to a new process.

I bet the vast majority of high-TDP laptops will stick with ADL-P. Few enough laptops would use 55W to begin with, and all it gets you is +2 cores and more IO.

 

[IntelUser2000]

New graphics driver: https://downloadmirror.intel.com/30381/eng/ReleaseNotes_100.9466.pdf

• Intermittent crash or hang seen in Cyberpunk 2077*i (DX12), Ridge Racer Unbounded*, StarCraft 2*, Horizon Zero Dawn* (DX12).

Took them look enough on Horizon Zero Dawn. It's not whether the performance is crap or not, it has to be feature complete and stable first.

They still got lot of issues just by looking at the Known Issues section. Intermittent crash or hang seen in a super popular game like WoW is troubling. Also the fixes regarding SC2. Really?

I bet the vast majority of high-TDP laptops will stick with ADL-P. Few enough laptops would use 55W to begin with, and all it gets you is +2 cores and more IO.

The extra GCM cores and I/O probably matters in the few DTR gaming laptops it gets into.

Can't compare exactly to the Apple/Qualcomm ones. AMD would benefit power-wise just by moving to a monolithic die.

I'm going to give them the benefit of the doubt second time and see how Alderlake-M does. https://adoredtv.com/exclusive-alder-lake-m-and-p-power-consumption-figures/

Lakefield seems to regress significantly in the idle power department(despite the high level details suggesting otherwise), and Alderlake-M is supposed to get that in control.

The 200-300mW difference seems small but platform power loosely follows SoC power management, and may end up adding 2-3 hours. You add 3 hours to a device like this: https://www.notebookcheck.net/Samsu...Lakefield-with-initial-problems.480990.0.html

And it would significantly reduce the gap it has against ARM devices.

 

[TESKATLIPOKA]

Aside from that the top Tremont cores get about ~750 on Geekbench for ST: https://browser.geekbench.com/v5/cpu/search?utf8=✓&q=n6005 . An expected +30% on top of that is ~1000 which is pretty close to the original Skylake (core 6000 series), but current rocket lake is closer to 1800. So Skylake level performance is likely more like 50% of big core than just 10% slower. That said see my advice above about not looking at ST perf for MT predictions.

I think he meant big core is 20% faster than Rocket Lake and small core is 10% slower than Rocket Lake.

 

[Exist50]

The extra GCM cores and I/O probably matters in the few DTR gaming laptops it gets into.

Maybe, but I'm not terribly convinced. 6+8 should be enough CPU, gaming-wise, for any single GPU you could fit in a laptop. Likewise for x8 PCIe 4.0. I suspect you're right about "DTR" laptops using S-BGA, but I'm doubtful of the real value it'd add. Hell, in some cases it might even be a regression with LPDDR5 vs DDR5.

Can't compare exactly to the Apple/Qualcomm ones. AMD would benefit power-wise just by moving to a monolithic die.

It's not just disaggregation. IIRC, Apple can put >80% of the chips power to the cores. Think <70% for Intel and AMD. The ratio rises at higher power levels, but <15W, it's a huge issue. I have no numbers, but I wouldn't be surprised if Lakefield were <60% running at TDP.

I'm going to give them the benefit of the doubt second time and see how Alderlake-M does. https://adoredtv.com/exclusive-alder-lake-m-and-p-power-consumption-figures/

I'm generally pessimistic about ADL-M. Way back I mentioned that I have a friend that used to work at Intel. Well he couldn't/didn't talk much about products beyond the occasional tease, but politics is fair game. Apparently IDC wasn't very power-aware, and they're in charge of ADL. Though this was a couple of years ago, so maybe someone's given them a kick in the pants.

Though it's not like it would be hard to do way better than Lakefield.

 

[TESKATLIPOKA]

I personally like the concept, but Lakefield was simply a FLOP. If you look at the notebookcheck review then the CPU and IGP was very underwhelming. I think the biggest culprit for super low performance was TDP. Normally this Lakefield has 7W(9.5W Bust) TDP, which is already very low, but this one had only 5W(9.5W) TDP. I think with 15W and one more big core It could have been way better than this, but then It would have been compared to Ice Lake U, which would most likely be still better.

Some new napkin math!
28W Intel Core i7-1185G7: 4C8T -> 3GHz base and 4.8GHz turbo

28W Alder Lake could look like this:
6x Golden Cove(big core): 2GHz base and 4.8GHz turbo
8x Gracemont(small core): 1.8GHz base and 3.6GHz turbo
Golden cove would have 50% higher IPC than Gracemont and also having 33% higher clockspeed would mean 2x higher ST performance, on the other hand Gracemont could be 2x more power efficient, in other words having only 50% of performance but needing only 25% of power.
If you ask why the base clock difference is only 11% even though turbo is 33% then the answer is to compensate for the missing HT in Gracemont, which would improve Golden Cove performance by 20%, this way even at base clocks the difference in performance would be kept at 100%.

Now what about theoretical ST and MT performance comparison? Let's set IPC of Golden Cove is 20% higher than Tiger Lake.
ST(turbo clocks):
28W Intel Core i7-1185G7 -> 100% at 4.8GHz
28W Alder Lake -> 120% at 4.8GHz

MT(base clocks):
28W Intel Core i7-1185G7 -> 100% at 3GHz
28W Alder Lake -> 120(big cores: +20% IPC, +50% cores, -33% frequency) + 80(small cores: 1/2 performance of big cores, +33% more cores) -> 200%
So Alder Lake could be 100% faster in MT than Tiger Lake.

I think this looks quite reasonable for a speculation. What do you all think?

 

[TESKATLIPOKA]

Now let's see about Tiger Lake vs Alder Lake at 15W TDP.
15W Intel Core i7-1180G7: 4C8T -> 2.2GHz base and 4.6GHz turbo

15W Alder Lake could look like this:
2x Golden Cove(big core): 2GHz base and 4.6GHz turbo
8x Gracemont(small core): 1.8GHz base and 3.45GHz turbo
I will keep the base clocks the same thanks to the missing 4 big cores.

ST(turbo clocks):
15W Intel Core i7-1185G7 -> 100% at 4.6GHz
15W Alder Lake -> 120% at 4.6GHz

MT(base clocks):
15W Intel Core i7-1185G7 -> 100%
15W Alder Lake -> 55(big cores: +20% IPC, -50% cores, -9% frequency) + 110(small cores: 1/2 performance of big cores, +200% more cores) -> 165%

This is a surprise, Alder could be 65% faster by using every single core! To tell you the truth, I was disappointed at first, when I first read Alder Lake P U15 will be only 2+8, but now It doesn't look so bad.
Of course using 3-4 threaded apps would make Tiger Lake faster thanks to using only big cores while Alder would use 2 big and 2 small ones, but 1-2 and 5-12 threaded apps should be faster on Alder Lake.

 

[uzzi38]

I think he meant big core is 20% faster than Rocket Lake and small core is 10% slower than Rocket Lake.

Even that only makes sense when looking at IPC. It's unlikely (to say the least) that Gracemont will clock at 5+GHz. Frankly, I'd be rather surprised if they surpass the 4GHz mark.

 

[uzzi38]

I'd be careful with looking at per core performance improvements to determine MT performance. What often gets quoted are ST performance or IPC numbers and those may not always be an indication of MT performance. A great example here is Zen3, which got like +19% IPC and >+25% ST performance but only

From what I've seen, Zen 3 seems to yield less from SMT on average than Zen 2 did. That's playing a part in terms of MT scoring.

Also, I feel like pointing out that in MT workloads like CB the 5950X (and I also imagine the 3950X) are actually incapable of reaching their maximum 142W PPT and are instead restricted by the stock current limits (TDC and EDC) of AM4 and instead sit at pretty much exactly 125W. I'll include an image below (excuse the formatting if there are issues, typing this from my phone).

This isn't addressed at you or anyone else as far as I can tell, just wanted to point out that this is a long workload where the 5950X doesn't benefit from a higher power limit than 125W TDP Intel parts.

1618746089130-png.png (3840×1235) (discordapp.com)

EDIT: Image replaced

 

[TESKATLIPOKA]

Even that only makes sense when looking at IPC. It's unlikely (to say the least) that Gracemont will clock at 5+GHz. Frankly, I'd be rather surprised if they surpass the 4GHz mark.

Even for IPC It's unlikely in my opinion. What I expect is already above your post and in MT those small weak cores help a lot.

 

[Hitman928]

I'm going to give them the benefit of the doubt second time and see how Alderlake-M does. https://adoredtv.com/exclusive-alder-lake-m-and-p-power-consumption-figures/

I'm going to call absolute BS on this 'leak'. He's claiming ALD-P and TGL use ~1W or less in these tests. There's absolutely no way these are real.

 

[eek2121]

I think he meant big core is 20% faster than Rocket Lake and small core is 10% slower than Rocket Lake.

I was referring to SKL vs Zen 3. According to Anandtech, SKL is anywhere from 3%-13% slower in SPEC 1T workloads. Gracemont will likely be competitive with skylake. Based on the numbers we have been able to find, I expect integer performance to be close to skylake (possibly a bit better), while floating point performance will trail skylake a bit due to slower AVX performance. We will see.

Even that only makes sense when looking at IPC. It's unlikely (to say the least) that Gracemont will clock at 5+GHz. Frankly, I'd be rather surprised if they surpass the 4GHz mark.

Current engineering samples appear to have the small cores running at 4.7 ghz. Intel has no real reason to cap frequencies except on low power devices. They also have no competitive reason to do so.

I expect the big reason for hybrid on desktop is die size savings, not power savings.

Also, anyone that thinks Intel’s top of the line 12th gen Core chip won’t beat the 5950X is going to be completely wrong. The golden cove cores alone will get Intel most of the way there. I don’t understand how anyone can can rationalize ADL not being able to perform. If the claimed 20% performance over sunny/willow cove is true, the 8 golden cove cores alone will likely perform as well as a 12 core 5900x.

If Intel thought they were going to be trailing Zen 3, they would have added more big cores.

I have said my piece. We will find out how close I am in a few months. 😊

 

[dmens]

Current engineering samples appear to have the small cores running at 4.7 ghz

LOL. Good one.

 

[TESKATLIPOKA]

Current engineering samples appear to have the small cores running at 4.7 ghz. Intel has no real reason to cap frequencies except on low power devices. They also have no competitive reason to do so.

Please show me those engineering samples, where small cores are clocked at 4.7GHz, I am very interested. BTW why do you think Gracemont is capable of running at those high clockspeeds? That's a mystery to me.

I expect the big reason for hybrid on desktop is die size savings, not power savings.

I agree 16 big core Alder Lake would be pretty big even with only 32EU IGP, yet Intel is putting 96EU IGP in It's 6+8 configuration, so die size shouldn't be such a big issue.
I will try to calculate how much bigger would be a 4C8T Tiger Lake, If I put another 12 cores in It using this dieshot. I got ~43mm2 for 4 Tiger Cores + L3 and the Ring Agent. So a 16C32T 96EU Tiger Lake would be 146+(3*43)= ~275mm2.

Also, anyone that thinks Intel’s top of the line 12th gen Core chip won’t beat the 5950X is going to be completely wrong.

Very strong words. At best I expect comparable performance in Cinebench for example and wouldn't be surprised If It performs worse.

The golden cove cores alone will get Intel most of the way there. I don’t understand how anyone can can rationalize ADL not being able to perform. If the claimed 20% performance over sunny/willow cove is true, the 8 golden cove cores alone will likely perform as well as a 12 core 5900x.

How "up to 20% higher ST performance" compared to Rocket Lake or Tiger Lake could compensate for the missing 4 cores in a battle between 8 Golden Cove cores vs 12 Zen3 cores is beyond my understanding.
BTW Ryzen 9 5900X is 60% faster than Core i9-11900K in Cinebench R23 MT test. If we enable Max power limit or Adaptive boost, then 5900X is only 34% faster, but the power consumption increases from 260W to 307W or 419W for the whole system, while the still faster 5900X system consumes only 183W! Link

If Intel thought they were going to be trailing Zen 3, they would have added more big cores.

I have said my piece. We will find out how close I am in a few months. 😊

Intel is seriously behind AMD in MT performance, yet Rocket Lake has even fewer cores than It's predecessor. Do you know what's the biggest issue with Rocket Lake? Very high power consumption as shown in the above link. So It's not surprising they want to lower It in Alder Lake by using both big and small cores.

 

[AMDK11]

If GoldenCove gives a 20% higher IPC, the advantage over Zen3 will be around 8%. Not too much. I hope the height will be higher.

 

[mikk]

mikk, I highly suggest you lower your expectations of Xe for Alder Lake.

What expectations? It can't be the expectation that Tigerlake graphics is only 30-40% faster than Icelake. Availability of Cezanne is so bad, it wouldn't surprise me if Rembrandt won't be available before H2 2022.

 

[RTX]

Is Tigerlake's IGP is supposed to be running at 1.35Ghz or 1.65Ghz or can it even boost to 1.80Ghz if the vendor allows it with that 65W power limit?

Intel’s 11th Gen Core Tiger Lake SoC Detailed: SuperFin, Willow Cove and Xe-LP

www.anandtech.com

 

[IntelUser2000]

Is Tigerlake's IGP is supposed to be running at 1.35Ghz or 1.65Ghz or can it even boost to 1.80Ghz if the vendor allows it with that 65W power limit?

Don't take that thing seriously.

The iGPU in 1185G7 has a 1.35GHz max clock: https://ark.intel.com/content/www/u...cessor-12m-cache-up-to-4-70-ghz-with-ipu.html

What expectations? It can't be the expectation that Tigerlake graphics is only 30-40% faster than Icelake. Availability of Cezanne is so bad, it wouldn't surprise me if Rembrandt won't be available before H2 2022.

Are you still sticking to that? How long ago was that from? Give it up already. So you don't have a problem with them saying Tigerlake is the best CPU ever made and that Xe would be a game changer? Or that they were hyping it for 6 months? Xe is decent but nothing to be super excited about. 10-30% advantage in graphics over competition using 2 generations old is their best effort? And they have glitches in games that are over a decade old?

 

[IntelUser2000]

I personally like the concept, but Lakefield was simply a FLOP. If you look at the notebookcheck review then the CPU and IGP was very underwhelming. I think the biggest culprit for super low performance was TDP. Normally this Lakefield has 7W(9.5W Bust) TDP, which is already very low, but this one had only 5W(9.5W) TDP.

Not sure if it's entirely due to TDP. The N6000 with 6W TDP gets similar ST score, and beats it in MT by 25%. So they didn't need all that fancy Foveros tech, nor even Sunny Cove.

Base clocks are not that relevant. If you see Notebookcheck tests, Tigerlake routinely can maintain 2.5GHz clocks even running stress tests such as Prime95. Again, look back to Icelake with it's 1.3GHz clocks. It'll run at 2GHz clocks under Prime95 and 15W as well.

Here's my simple analysis:

1165G7 Cinebench R20 ~20W: 2000(not taking burst numbers)
Pentium Silver N6000 "Tremont" R20: 775

1165G7 x 1.2 / 1.85 = 1300
N6000 x 1.85(typical Cinebench scaling) x 1.3 = 1850

Total: 3150

Golden Cove needs to be more efficient(better perf/W) otherwise we can't simply multiply Tigerlake numbers by 1.2x.

28W will not be enough to fully unleash 6 Golden Cove cores, nevermind having to share with Gracemont.

I'm going to call absolute BS on this 'leak'. He's claiming ALD-P and TGL use ~1W or less in these tests. There's absolutely no way these are real.

What seems nonsense can be a matter of misunderstanding.

Those are not peak power numbers. Those are power numbers in bursty workloads such as video playback and web browsing. If you take the low duty cycle into account, it's perfectly reasonable. If you have a decently modern laptop, you can see that for yourself using an application such as HWInfo.

Otherwise how do you think you can get 10 hour battery life on a laptop with a 25W rated chip with just a 50WHr battery?

 

[Exist50]

What expectations? It can't be the expectation that Tigerlake graphics is only 30-40% faster than Icelake. Availability of Cezanne is so bad, it wouldn't surprise me if Rembrandt won't be available before H2 2022.

The expectation that Alder Lake will have any significant GPU improvements over Tiger Lake.

 

[Hitman928]

What seems nonsense can be a matter of misunderstanding.

Those are not peak power numbers. Those are power numbers in bursty workloads such as video playback and web browsing. If you take the low duty cycle into account, it's perfectly reasonable. If you have a decently modern laptop, you can see that for yourself using an application such as HWInfo.

Otherwise how do you think you can get 10 hour battery life on a laptop with a 25W rated chip with just a 50WHr battery?

I think I'll need to see some evidence of Intel's big core SOCs idling at 0.1W or doing any kind of real activity whatsoever at 1W or less (bursty or not) before I give Adored's charts any credence. I have yet to see anything like that from any of Intel's big core SKUs.

 

[TESKATLIPOKA]

Base clocks are not that relevant. If you see Notebookcheck tests, Tigerlake routinely can maintain 2.5GHz clocks even running stress tests such as Prime95. Again, look back to Icelake with it's 1.3GHz clocks. It'll run at 2GHz clocks under Prime95 and 15W as well.

You are right about CPUs working at higher frequencies than the base even in a very demanding soft like Prime95, but this is not new info to me. The reason why I used base to calculate MT performance was because It was the easiest and the fastest way. I could very well say both of them are clocked 20% higher let's say in Cinebench R23 and my MT calculation would stay the same. This is still just a speculative estimate, where I chose IPC and frequencies, I could be very wrong with It.

Golden Cove needs to be more efficient(better perf/W) otherwise we can't simply multiply Tigerlake numbers by 1.2x.

28W will not be enough to fully unleash 6 Golden Cove cores, nevermind having to share with Gracemont.

I expect both the process and Golden Cove to be more efficient than Tiger Lake. I calculated It like Golden Cove has 20% better IPC than Tiger Lake.
Because of more big cores + small cores I lowered the base clocks of Golden Cove to just 2GHz, while Tiger lake's base is clocked 50% higher for the 28W TDP version.
Basically I tried to set 3W per Golden Cove and 0.75W per Gracemont core.
For 28W TDP It would be 3*6+0.75*8=24W and 4W for the rest for a total of 28W.
For 15W TDP It would be 3*2+0.75*8=12W and 3W for the rest for a total of 15W.

Was I too generous with clockspeeds?
After looking at Pentium N6005 which is a 4C Tremont 2GHz base with a 10W TDP, then I can say I was at least with Gracemont's base frequencies. Instead of 1.8GHz I should have lowered It to 1.3-1.5GHz. 2x better power efficiency for Gracemont than Golden Cove is also a bit too optimistic, should have stayed with only 50% better efficiency, in other words 50% of performance needing only 33% of big core power(Golden Cove 3W and Gracemont 1W). This is just my speculation.

 

[coercitiv]

I think I'll need to see some evidence of Intel's big core SOCs idling at 0.1W or doing any kind of real activity whatsoever at 1W or less (bursty or not) before I give Adored's charts any credence. I have yet to see anything like that from any of Intel's big core SKUs.

My understanding is those are per core numbers, not per SOC... otherwise I'd love to find out more about what kind of real-world workload uses just 1W on TGL-U.