Discussion Intel current and future Lakes & Rapids thread

[coercitiv]

The Apple M1 is an absolute killer. The CPU's single thread performance is Tigerlake or better, they are claiming 40% better overall performance so the MT performance might be on par with Renoir, and the GPU is a 2.6TFlop part that performs at least on par with a well-optimized + 28W Iris Xe(which doesn't exist on production laptops) and possibly even faster than that.

Don't get me wrong, I'm not downplaying M1, but I was actually expecting a more aggressive presentation instead. As it stands the most valuable information we have is still based on the performance numbers from GB5 and sites like Anandtech, corroborated with the new resource allocation described by Apple for M1. (number of cores, UMA & memory packaging, active cooling which also mens ST perf scaling etc)

The one thing that hits the hardest when looking at Intel's product stack is the closest product they have to M1 is Lakefield. I'm not going to compare performance and power numbers, but focus on sale strategy instead: they actually attempted to sell Lakefield as a premium product. They had Samsung build a Mac Air equivalent in both form factor and price, and the silicon inside now looks like a toy.

I honestly hope Alder Lake delivers 110% on internal performance goals.

 

[mikk]

Yoga 14s (i5-1135G7)

920 points in Cinebench R15 at 28W, basically zero drop off in following runs. 2345 points in Cinebench R20. The cooling performance is superb from what I can see. Two evenly matched fans, two really long heat pipes of the same size and a relatively big ventilation grill on the bottom. Playing PUBG at 30+ watts less than 70 degrees which is superb. One game he is playing the Soc can use ~40W and CPU goes up to ~70 degrees only.

 

[firewolfsm]

That looks like it's not a yoga? The hinge doesn't look flippable.

 

[mikk]

That looks like it's not a yoga? The hinge doesn't look flippable.

Not all Yoga are flippable. The box says Yoga 14 ITL and apparently the Yoga 14s in asia has a different branding compared to other areas.

The Lenovo Yoga Slim 9i 14-inch laptop (known as Yoga Pro 14s in China and IdeaPad Slim 9i in North America) will start at €1899 (including VAT) and is expected to be available starting November 2020.4

The Latest Lenovo Yoga Laptops Feature New Intel 11th Gen Core Mobile Processors with Intel Iris Xe graphics - Lenovo StoryHub

Today, Lenovo reveals the thin and light Yoga laptops that will offer up to the new 11th Gen Intel® Core™ mobile processors with Intel Iris® Xe graphics.

news.lenovo.com

The Slim isn't flippable. edit: It's the Lenovo Yoga Slim 7i Pro:

 

[TheGiant]

Yoga 14s (i5-1135G7)

920 points in Cinebench R15 at 28W, basically zero drop off in following runs. 2345 points in Cinebench R20. The cooling performance is superb from what I can see. Two evenly matched fans, two really long heat pipes of the same size and a relatively big ventilation grill on the bottom. Playing PUBG at 30+ watts less than 70 degrees which is superb. One game he is playing the Soc can use ~40W and CPU goes up to ~70 degrees only.

Oki
Finally someone Who just copypasted the referencie system
Now we need premium display option with that cooling
New HP spectre maybe?

 

[mikk]

Oki
Finally someone Who just copypasted the referencie system
Now we need premium display option with that cooling
New HP spectre maybe?

Lol the cooling setup from the reference system is much worse than this, the reference was based on a 1 fan cooling with a shorter heatpipe. It's much worse if it's unchanged from the MSI Prestige 14 which I believe it is. It didn't really affect the performance because it can go over 90 degrees without thermal throttling, the MSI prestige 14 is a really fast device but it gets hot (90+ degrees in Cinebench R20). Even the MSI Summit 15 with a 2 fans solution is much worse (the second fan is really small).

 

[coercitiv]

Now we need premium display option with that cooling

That's in the Slim Pro 7i - 2880x1800 / 90Hz / 100% sRGB / 400 nits

 

[IntelUser2000]

That Ryan Shrout guy whose tweet was linked above should be happy that Apple didn't release precise benchmark numbers for the M1. Anandtech just did for the A14, and things don't look so rosy for TGL.
As for the GPU, the A12Z already beats the Intel Xe in 3DMark Wild life (some blame Windows Vulkan drivers, but I don't buy that theory). And consider that the M1 is probably quite a bit faster than the A12Z.

To be fair, its likely a FP16 vs FP32 comparison which will skew the results by 30-40% in favor of the mobile devices. It certainly is for 3DMark Night Raid.

@coercitiv Lakefield is a disappointment. Jasper Lake 6W devices are going to do the same/better but cost <$500. Actually at least for R20 MT the N6000 will perform 25% better.

@mikk Finally a decent Tigerlake implementation!

The branding is crap though. I don't know why they hate themselves. Yoga should have always remained a convertible. Now even I can't keep track. This is like if Macbook Air suddenly was available in a 17" screen and with RTX 2080 mobile graphics, but 2mm thinner than the rest.

give me 16CU Navi2 Rembrandt / 2 GHz and LPDDR5/DDR5 as leaks claim and that M1 GPU wont be that interesting, it's just that amd has been using the same arch for 3 years... and it wouldnt get much bigger, not much further than 30-35 mm2 (@6nm, not even N5)

I hope you are seeing the issue here. Remembrandt is going to be more than a year away. And the performance level of the Apple part is at ~10W.

 

[Antey]

I hope you are seeing the issue here. Remembrandt is going to be more than a year away. And the performance level of the Apple part is at ~10W.

Yes, they are winning that race, if there is one. Apple gpu cores are really big, not very scalable but very power efficient. They are not the only one winning that race, qualcomm adreno gpus are GIGANTIC, each core has 512-768 ALUs, they definitly dont scale well but they are really competitive with apple in the performance per watt area. and they are also coming to notebooks. ARM mali G78 has a more scalable design, computer cores and executions engines, 2 FMA units per EE, 16 FP32 per cycle per FMA Unit. It can scale up to 24 computer cores. AMD and NVIDIA designs are focused on big gpus, 80 compute units or even more. two opposite worlds colliding.

The only problem with those gigantic cores is that you cant rely on lpddr bandwidth forever, they will need hbm if they want to go bigger, that means an HBM phy and hbm dies, and that's doubling or tripling the power draw of the soc.

 

[LightningZ71]

I wonder how good LPDDR5X will be when developed to its fullest potential?

I’ve pointed this out before, but dual channel (quad really) DDR5 implementations can scale out to 100GB/sec with the JEDEC standard, and likely extend it further with higher speeds as it matures. That’s firmly in the territory of the EDRAM on the first XBox One, and also is similar to the RX550/560. That’s more than enough to feed a sizable iGPU implementation of RDNA2 and give excellent 1080p performance in mobile and desktop. On a laptop in the 15” and smaller range, that’s more than enough for all but the most hardcore players.

 

[Exist50]

I wonder how good LPDDR5X will be when developed to its fullest potential?

I’ve pointed this out before, but dual channel (quad really) DDR5 implementations can scale out to 100GB/sec with the JEDEC standard, and likely extend it further with higher speeds as it matures. That’s firmly in the territory of the EDRAM on the first XBox One, and also is similar to the RX550/560. That’s more than enough to feed a sizable iGPU implementation of RDNA2 and give excellent 1080p performance in mobile and desktop. On a laptop in the 15” and smaller range, that’s more than enough for all but the most hardcore players.

Last I heard, LPDDR5 is projected to scale to at least 8000MHz.

 

[IntelUser2000]

Yes, they are winning that race, if there is one. Apple gpu cores are really big, not very scalable but very power efficient.

Why do you think they won't scale? It's a GPU. You put copies of them and it'll scale. Mali didn't grow because it didn't need to. Apple's M1 will because eventually it'll need to displace CPU and GPU in Macs.

And what do you mean by big? Their GPU cores are extremely compact considering their performance.

 

[jpiniero]

Intel Icelake Server Die Size & Floorplan Inefficiencies Revealed

This post was based on the assumption that XCC would be 42C. It is now revealed to be 40C. The various die sizes are slightly off, but much of the analysis within remains accurate with regards to t…

semianalysis.com

First I had seen about Icelake Server's core counts and die sizes - the 16 core LCC is 370 mm2, the 28 core HCC is 505 mm2, and the 42 core XCC is 640 mm2. Wonder how many chips they can really get without having to cut down to like 6 cores.

 

[DisEnchantment]

Why do you think they won't scale? It's a GPU. You put copies of them and it'll scale. Mali didn't grow because it didn't need to. Apple's M1 will because eventually it'll need to displace CPU and GPU in Macs.

And what do you mean by big? Their GPU cores are extremely compact considering their performance.

GPU for compute sure, GPU for gaming is different. (See PS5 vs XSX, N10 vs V64). PC GPUs have way too much fixed function blocks, almost 35% of die size in N10 for example (then add mem controllers etc etc)
MS don't want, can't alter rendering pipelines overnight without breaking a lot of HW and game engines.

 

[DisEnchantment]

Intel Icelake Server Die Size & Floorplan Inefficiencies Revealed

This post was based on the assumption that XCC would be 42C. It is now revealed to be 40C. The various die sizes are slightly off, but much of the analysis within remains accurate with regards to t…

semianalysis.com

First I had seen about Icelake Server's core counts and die sizes - the 16 core LCC is 370 mm2, the 28 core HCC is 505 mm2, and the 42 core XCC is 640 mm2. Wonder how many chips they can really get without having to cut down to like 6 cores.

It is fairly obvious why so much of delays. The glue is not ready.
I imagine Intel's 10SF yields should be comparable to early N7 now(for sure Intel 10SF has higher density and perf), problem is, they have monolithic designs only. Can't go big without taking the defect hits.
That glue ironically will be their saving grace going forward.

 

[IntelUser2000]

First I had seen about Icelake Server's core counts and die sizes - the 16 core LCC is 370 mm2, the 28 core HCC is 505 mm2, and the 42 core XCC is 640 mm2. Wonder how many chips they can really get without having to cut down to like 6 cores.

The issue with Intel is that they are targetting for unrealistic clocks so the cores are too large. It's fine for gaming desktops but nothing else. It hurts them in servers and in laptops.

They really need a third core for that. The core size ratio compared to Atom has grown from 4:1 in the Skylake/GLM+ era to 5:1 in SNC/TNT era. It's ridiculous. With the added caches Willow Cove is at 9mm2. With AVX-512, it looks like Icelake-SP's core portion is nearly 11mm2. No wonder AMD can put double the amount of cores in similar die area!

I imagine Intel's 10SF yields should be comparable to early N7 now, problem is, they have monolithic designs only.

Icelake-SP uses Icelake's 10nm process, not Tigerlake's 10nm SF.

Monolithic isn't that big of an issue as Nvidia creates 800mm2+ dies.

Intel would also prefer the first generation dies on the new process being closer to 450mm2 as they did for many generations, and ramping up to 700mm2 on the second one. Ideally, the first 10nm server chip would have been Cannonlake-SP and be 400-450mm2, with Icelake-SP being the second 10nm and 700mm2.

Apple GPU: I think the team that created the phenomenal M1 chip knows what's necessary to make it scale to bigger devices. Maybe they'll use 3rd party for the biggest parts but I don't think they'll stop at M1.

 

[DisEnchantment]

Monolithic isn't that big of an issue as Nvidia creates 800mm2+ dies.

Whats stopping them from shipping high volume 10nm server parts then?

 

[Antey]

It's not that they can't put many copies of them and call it a day. my understanding is that they don't scale well because the wider your cores are the harder to scale frequencies (eg adreno 640 clocks at 500MHz while Adreno 630 could achieve 700+ MHz). higher clocks means faster rasterization, more cache bandwidth, processing the command buffer goes faster, etc. you would need a redesign, extra resources to feed that gpu, etc. One thing is computational performance, but that's not everything.

With bigger i mean wider, many vector ALUs.

 

[RTX]

Are the new SF features also going to be found in Intel's 7nm nodes? Is it likely the first 7nm won't have the SF features?

 

[NTMBK]

Charlie is talking about another big Intel server delay: https://semiaccurate.com/2020/11/12/intel-delays-a-mainstream-server-platform-once-again/

What do we reckon- is Sapphire Rapids delayed?

 

[Hitman928]

Charlie is talking about another big Intel server delay: https://semiaccurate.com/2020/11/12/intel-delays-a-mainstream-server-platform-once-again/

What do we reckon- is Sapphire Rapids delayed?

The DOE just announced last month that the Aurora supercomputer slated to use Sapphire Rapids and Intel's 7 nm GPU would be delayed yet again and no new date was given for when it is expected to come online. It was originally speculated that it was due to the delay of the Intel 7 nm GPU but it's certainly possible that SR won't be ready either. That would push SR to some time in 2022 where it will be facing off against whatever AMD is cooking up with Zen4 using TSMC 5nm.

 

[jpiniero]

The DOE just announced last month that the Aurora supercomputer slated to use Sapphire Rapids and Intel's 7 nm GPU would be delayed yet again and no new date was given for when it is expected to come online. It was originally speculated that it was due to the delay of the Intel 7 nm GPU but it's certainly possible that SR won't be ready either. That would push SR to some time in 2022 where it will be facing off against whatever AMD is cooking up with Zen4 using TSMC 5nm.

Weren't there rumors that the end of 2021 was the drop dead date for Aurora? If not, perhaps the DOE should consider dropping Intel and going with AMD or nVidia/ARM.

As for the article, if SPR is delayed that's really bad.

 

[Hitman928]

Weren't there rumors that the end of 2021 was the drop dead date for Aurora? If not, perhaps the DOE should consider dropping Intel and going with AMD or nVidia/ARM.

Yes, there have been rumors that the DOE has basically forced Intel to present a backup plan where they will need to source components from other manufacturers (e.g. use other GPUs if Intel's weren't going to be ready) if they want to have a hand in Aurora at all. I don't think it's much of a secret that the DOE is pissed at Intel right now and if Intel can't deliver this time at all, it wouldn't surprise me if they don't scrap Aurora entirely as it currently exists and move to AMD/Nvidia/IBM/ARM.

 

[positivedoppler]

The DOE just announced last month that the Aurora supercomputer slated to use Sapphire Rapids and Intel's 7 nm GPU would be delayed yet again and no new date was given for when it is expected to come online. It was originally speculated that it was due to the delay of the Intel 7 nm GPU but it's certainly possible that SR won't be ready either. That would push SR to some time in 2022 where it will be facing off against whatever AMD is cooking up with Zen4 using TSMC 5nm.

This deserves its own thread. To quote Charlie
" Don’t believe it, things just went off the rails again and the situation is beyond dire, Intel is out of the server game for the foreseeable future. "

Charlie has been doom and gloom when it comes to Intel for the past few years, but the problem is he's been accurate. He's been calling out Intel's 10nm delay for years now. If he's right about this one, and he's been accurate, Intel is about to totally flip server position markets with AMD. Remember when AMD had a 2% share? Wow...just wow...

 

[ondma]

This deserves its own thread. To quote Charlie
" Don’t believe it, things just went off the rails again and the situation is beyond dire, Intel is out of the server game for the foreseeable future. "

Charlie has been doom and gloom when it comes to Intel for the past few years, but the problem is he's been accurate. He's been calling out Intel's 10nm delay for years now. If he's right about this one, and he's been accurate, Intel is about to totally flip server position markets with AMD. Remember when AMD had a 2% share? Wow...just wow...

I would not say Charlie has been "accurate". More like the title of his rag "semi-accurate". He seem to go by the theory that if you spread enough FUD, some if it will come true.