Discussion Intel current and future Lakes & Rapids thread

[dmens]

That's every rumor mill and product hype train in the history of the universe. Occasionally they manage to get one right, but they have no shame about being wrong, even if by a massive amount.

For some I suspect there's little joy even if they do manage to get it right, because it's the potential and excitement about what could be that drives people to participate and wrong or right as soon as a product launches, there's always a new hype train for the next product that's just about to depart the station.

You are likely right, it is all emotional. Is that why when I said RKL will completely suck months before launch, I got such a negative reaction in this very thread? I did get some good laughs from the insults.

 

[mikk]

Actually the early RKL-S IPC rumors/speculations were too pessimistic and the first subpar Geekbench entries were misleading. Some expected 10% improved IPC or something like that, in the end we got 15-20% depending on the test. Intel claimed up to 19% improved IPC, so it's clearly not a real up to figure (a real up to would involve AVX512), similar to their UHD750 up to 50% figure which turned out it's an average increase.

 

[KompuKare]

There is no issue with a chip consuming even 200W of power as long as the 15W TDP is met (I feel like yelling at AMD, Intel, NVIDIA, shoot, anyone making "processors" for allowing this to ben open definition).

Maybe from a cooling and chassis point of view, but such a spike will put a huge strain on the PSU and mobo power delivery system. Such spiky CPUs could be designed for, but that is hardly a zero cost thing.

 

[ashFTW]

Intel Sapphire Rapids to feature up to 56 cores, 350W TDP and 64GB of HBM2 memory - VideoCardz.com

Intel 4th Gen Xeon Scalable (Sapphire Rapids) More cores, more PCIe lanes, and support for next-gen interfaces. We received an anonymous tip with a slide featuring Intel’s future Xeon platform specifications. The slide is dated, as it still lists the Ice Lake server as a 2020 product, however it...

videocardz.com

There you go, 56 cores. Any other questions?

There you go. 80 cores.

It will be up to 80 cores of SPR vs up to 96 cores of Zen 4 EPYC.

Intel's Next-Gen 10nm ESF Based Sapphire Rapids Xeon CPU Die Shot Unveils MCM Design & Up To 80 Cores In 4 Chiplets

New die shots of Intel's next-gen Sapphire Rapids Xeon CPUs have surfaced which show an MCM design that could house up to 80 cores.

wccftech.com

 

[jpiniero]

There you go. 80 cores.

80 might be what's physically there but between yields and power consumption who knows what Intel will do. That slide says 56 cores and 350 W.

Maybe you will see a 500 W off roadmap CPU?

 

[Mopetar]

Maybe you will see a 500 W off roadmap CPU?

At 500W Intel needs to switch their codenames to volcanoes instead of bodies of water.

 

[LightningZ71]

There you go. 80 cores.

It will be up to 80 cores of SPR vs up to 96 cores of Zen 4 EPYC.

Intel's Next-Gen 10nm ESF Based Sapphire Rapids Xeon CPU Die Shot Unveils MCM Design & Up To 80 Cores In 4 Chiplets

New die shots of Intel's next-gen Sapphire Rapids Xeon CPUs have surfaced which show an MCM design that could house up to 80 cores.

wccftech.com

From the WCCFTech article (for whatever that may be worth),
" With all four chiplets exposed, we can see that underneath them is a 5x4 core configuration which means each die consists of up to 80 cores. However, the entire 80 core silicon will never be released to the public due to the mesh layout."

"Theoretically, Intel's Sapphire Rapids-SP Xeon CPUs could feature a maximum of 72 cores and 144 threads but we know from previous leaks that the maximum configuration is going to end up at 56 cores and 112 threads."

There you both go, Max 72 cores, retail at 56.

 

[ashFTW]

From the WCCFTech article (for whatever that may be worth),
" With all four chiplets exposed, we can see that underneath them is a 5x4 core configuration which means each die consists of up to 80 cores. However, the entire 80 core silicon will never be released to the public due to the mesh layout."

"Theoretically, Intel's Sapphire Rapids-SP Xeon CPUs could feature a maximum of 72 cores and 144 threads but we know from previous leaks that the maximum configuration is going to end up at 56 cores and 112 threads."

There you both go, Max 72 cores, retail at 56.

physical evidence is 80 cores, rest is WCCF conjecture. I know which one I would rather trust

 

[ashFTW]

80 might be what's physically there but between yields and power consumption who knows what Intel will do. That slide says 56 cores and 350 W.

Maybe you will see a 500 W off roadmap CPU?

maybe the 80 core version is just for Aurora, which has cooling for 500W for both CPUs and GPUs

 

[jpiniero]

physical evidence is 80 cores, rest is WCCF conjecture. I know which one I would rather trust

Some of the tiles might not be cores.

 

[ashFTW]

Some of the tiles might not be cores.

I think it’s a symmetrical design, which seems like a logical extension of Icelake, with the mesh being connected via EMIB. Each tile should be identical with 1/4th each of cores, I/O, and memory. There could be two tile versions as well, but with a different a different tile that has more I/O and mem channels.

 

[yuri69]

There are 5 * 4 surface structures visible on the photos. No doubt. However, what are those structures? To me it seems to be a (micro) bump structure not the actual silicon patterns.

 

[ashFTW]

There are 5 * 4 surface structures visible on the photos. No doubt. However, what are those structures? To me it seems to be a (micro) bump structure not the actual silicon patterns.

One of those 20 structures could be dual memory channel, and another could be UPI to support dual socket. PCIe could be the structures seen on the top and the side, part of which could be to support EMIB. So it’s probably 72 cores, not 80.

 

[tomatosummit]

There are 5 * 4 surface structures visible on the photos. No doubt. However, what are those structures? To me it seems to be a (micro) bump structure not the actual silicon patterns.

Must be the micro bumps, they correlate with the leftover mess on the substrate.
I wish the extraction didn't damage the edges so much because you can't see the edge interfaces on the dies clearly. What you can see is 3 on the top/bottom edge and 2 on the L/R edges. Unless there's another 2 on the opposite side of the die it'd require a mirrored die which is very unlikely.
What it would also do is leave 8 interfaces on the outer edges of the die, could they be multi purpose for the hbm or cross socket interface perhaps.

 

[Exist50]

It's not 80 cores, nor even 72. The structures aren't all cores.

 

[uzzi38]

It's not 80 cores, nor even 72. The structures aren't all cores.

Don't bother, he won't listen.

There'll be a die shot taken with one of the tiles sanded a bit soon which will likely make for some interesting conversations, and later when the SKU list gets leaked some even more interesting ones. Right now you're not going to convince him of anything I'm afraid.

 

[ashFTW]

It's not 80 cores, nor even 72. The structures aren't all cores.

maybe some of them are unicorns lol

 

[DrMrLordX]

At 500W Intel needs to switch their codenames to volcanoes instead of bodies of water.

The lakes are suggestive of recommended cooling.

 

[eek2121]

The lakes are suggestive of recommended cooling.

I’ve never seen a lake boil before. 🤣

 

[mikk]

Did someone notice that on the partner event 2021 product roadmap there is no Alder Lake-P, it's only Alder Lake-S: https://images.anandtech.com/doci/16629/image_2021_04_20T15_23_40_681Z.png

Of course this isn't a surprise, this was expected. But it's Intels first confirmation. ADL-S BGA is a possibility which is likely more a showcase mobile.

I wonder if TGL-R will support LPDDR5.

 

[dmens]

Actually the early RKL-S IPC rumors/speculations were too pessimistic and the first subpar Geekbench entries were misleading. Some expected 10% improved IPC or something like that, in the end we got 15-20% depending on the test. Intel claimed up to 19% improved IPC, so it's clearly not a real up to figure (a real up to would involve AVX512), similar to their UHD750 up to 50% figure which turned out it's an average increase.

Hope you realize all that "IPC" talk is utterly meaningless without the additional context of power and frequency limits.

Also it does seem ironic to me for "enthusiasts" to even bother discussing the performance of an integrated graphics IP which is literally worst in class by a large margin in terms of framerate per die area and is really nothing more than a bare minimal monitor driver for the lowest possible cost.

 

[jpiniero]

really nothing more than a bare minimal monitor driver for the lowest possible cost.

That's really all you need. Frame rates, yes, AMD is always going to be faster just because of driver optimizations.

 

[mikk]

Hope you realize all that "IPC" talk is utterly meaningless without the additional context of power and frequency limits.

Singlethread performance is a result of frequency and IPC, this should be more than obvious.

Also it does seem ironic to me for "enthusiasts" to even bother discussing the performance of an integrated graphics IP which is literally worst in class by a large margin in terms of framerate per die area and is really nothing more than a bare minimal monitor driver for the lowest possible cost.

Simply ask yourself why they even bother working on faster iGPUs when it's just a monitor driver. Saying worst in class is strange, you should understand that not all mobile devices are suited for a dedicated GPU, be it for cost reasons, cooling reasons, size constraints etc. There is a big market for non dGPU devices and not everyone is fine with an UHD 620 like performance forever. A decent iGPU also helps removing the crap dGPUs from Nvidia, MX130 and MX230 are no longer a thing since Tigerlake.

That's really all you need. Frame rates, yes, AMD is always going to be faster just because of driver optimizations.

Right now AMD is losing mostly in frame rates, check out ultrabookreview or notebookcheck. Maybe they should work on better driver optimizations. But why even bother? According to some people it's just a monitor driver and there is no need for something faster.

 

[dmens]

Singlethread performance is a result of frequency and IPC, this should be more than obvious.

And yet somehow so many people ignored RKL's obvious frequency limitations within a reasonable power envelope while drinking the IPC gainz kool-aid.

Simply ask yourself why they even bother working on faster iGPUs when it's just a monitor driver. Saying worst in class is strange, you should understand that not all mobile devices are suited for a dedicated GPU, be it for cost reasons, cooling reasons, size constraints etc. There is a big market for non dGPU devices and not everyone is fine with an UHD 620 like performance forever. A decent iGPU also helps removing the crap dGPUs from Nvidia, MX130 and MX230 are no longer a thing since Tigerlake.

Then why did you bring up the supposed perf gains of UHD750?

Also the reason you work on faster iGPU is so you can shrink it down even more for parity performance, thus saving cost and power.

 

[dmens]

Right now AMD is losing mostly in frame rates, check out ultrabookreview or notebookcheck. Maybe they should work on better driver optimizations. But why even bother? According to some people it's just a monitor driver and there is no need for something faster.

Hah. You might want to do a area/power normalized comparison. Also, consider what AMD was able to achieve on CPU performance by saving area on graphics. As in, the point I just made in the last post.