Discussion Intel current and future Lakes & Rapids thread

[SlimFan]

-AP and -SP are very different things for BHS.

No doubt, but given the existence of GNR-AP for BHS, why are you suggesting that SRF-AP can't be successful because of "platform readiness"?

Are you implying that SRF-AP and GNR-AP are completely different platforms, and that Intel didn't realize this and forgot to disclose this to their customers prior to this announcement?

 

[H433x0n]

I don’t think SRF-AP was ever cancelled but was slated to come out alongside GNR-SP & GNR-AP?

 

[adroc_thurston]

why are you suggesting that SRF-AP can't be successful because of "platform readiness"?

No one is/was building scale-out boxes for -AP since it was not supposed to exist.

and forgot to disclose this to their customers prior to this announcement?

It's a damage control part given they've probably reached the acceptance stage wrt Turin-Dense.
Same as CL-AP or CPX-4.

I don’t think SRF-AP was ever cancelled

Oh it was.

 

[H433x0n]

Oh it was.

How do you interpret this then? They figured they could get it out alongside GNR-AP and just YOLO’d it by announcing it at Innovation event?

 

[adroc_thurston]

They figured they could get it out alongside GNR-AP and just YOLO’d it by announcing it at Innovation event?

Yes, SRF-SP is quite literally dead on arrival so they had to do something and that something is SRF-AP.
Wonder how ODMs feel about that (must be livid).

 

[Hitman928]

It's a damage control part given they've probably reached the acceptance stage wrt Turin-Dense.
Same as CL-AP or CPX-4.

As an outsider with no sources at Intel currently, this is basically how I read the situation. Seems Intel caught whiff of Zen5c (and maybe ARM competition?) SKU/performance coming next year and, Intel being Intel, had to make it at least appear that they were still in the lead, even if the product is essentially vaporware. Who knows though, maybe under Pat, they'll buck that kind of behavior and prove me wrong.

 

[Khato]

Contrary to popular belief, Intel not staffed by a pack of incompetent monkeys. Why exactly wouldn't Intel develop a SRF SKU with more than a single compute tile?

It's important to keep in mind that the echo chamber of the internet doesn't necessarily bear any resemblance to reality.

 

[adroc_thurston]

Seems Intel caught whiff of Zen5c

YES.
And they're doing the ol' faithful damage control part.

even if the product is essentially vaporware.

Not vaporware, just that no one built scale-out anything for 7529.

Intel not staffed by a pack of incompetent monkeys

I wish it was so but unfortunately.

Why exactly wouldn't Intel develop a SRF SKU with more than a single compute tile?

Cost, power, server form-factors; literally numerous reasons why SRF-AP was canned like 2 years ago to date.

 

[Khato]

Cost, power, server form-factors; literally numerous reasons why SRF-AP was canned like 2 years ago to date.

Again, only canned 2 years ago according to the internet rumor echo chamber which doesn't bear any resemblance to reality in this case.

 

[gdansk]

I said pretty recently 144c was max, I was wrong. Intel kept that secret pretty well. It was removed of roadmaps months ago. Now it’s back. Is this what it feels like to be MLID? 😂

Nah, being incorrect was based on available information. It is the right stance to take. It wasn't on official slides even a few months ago.

 

[adroc_thurston]

only canned 2 years ago according to the internet rumor echo chamber which doesn't bear any resemblance to reality in this case.

You're coping alright, but there's no reason to make it public.
It wasn't a thing, just like CL-AP wasn't a thing.

 

[coercitiv]

Is this what it feels like to be MLID? 😂

No, if you want to experience the true MLID sensation then go back and delete your original prediction, or change it to vague language open to interpretation.

Then come back here and tell us you were right all along.

 

[Abwx]

Why exactly wouldn't Intel develop a SRF SKU with more than a single compute tile?

8 current e cores using Intel 7 at 3GHz should take something like 20W, with a 20% perf/Watt process improvement for Intel 4 that put one core at 2W@3GHz, so 288 cores will be at 576W, and we re not counting MCs and generaly the I/O with PCI5 that is at least at 100W, so that s rougly a 700W SKU...

At 144 cores that s a little more than half the power, so about 400W.

 

[adroc_thurston]

8 current e cores using Intel 7 at 3GHz should take something like 20W, with a 20% perf/Watt process improvement for Intel 4 that put one core at 2W@3GHz, so 288 cores will be at 576W, and we re not counting MCs and generaly the I/O with PCI5 that is at least at 100W, so that s rougly a 700W SKU...

It's a better node and not 3GHz but again, no one was building scale-out boxes for 7529.

 

[dullard]

8 current e cores using Intel 7 at 3GHz should take something like 20W, with a 20% perf/Watt process improvement for Intel 4 that put one core at 2W@3GHz, so 288 cores will be at 576W, and we re not counting MCs and generaly the I/O with PCI5 that is at least at 100W, so that s rougly a 700W SKU...

At 144 cores that s a little more than half the power, so about 400W.

Intel 3 should be roughly 18% better perf/watt than what you calculated. Are you sure it will be 3 GHz? And finally, the E cores are running at a bad position in Intel 7 -- too high of a voltage without a dual power supply and thus at an inefficient part of the frequency/power curve.

Taking your 8 cores at 20 W -> 4 cores at 10 W, you'll see that we are already reaching the point of diminishing returns. Turn the power down to about half of what you calculated and now you are in a sweet spot for E core power/performance (right around 1.4 W per core).

Data from: https://chipsandcheese.com/2022/01/28/alder-lakes-power-efficiency-a-complicated-picture/

 

[Khato]

8 current e cores using Intel 7 at 3GHz should take something like 20W, with a 20% perf/Watt process improvement for Intel 4 that put one core at 2W@3GHz, so 288 cores will be at 576W, and we re not counting MCs and generaly the I/O with PCI5 that is at least at 100W, so that s rougly a 700W SKU...

At 144 cores that s a little more than half the power, so about 400W.

Won't disagree with those calculations, but 3GHz may well be beyond their target frequency. Drop the frequency to a 1.25W/core target and you get around the 450W mark. Even with ADL a 1.25W/core budget gets up to 2GHz, so still pretty reasonable.

 

[Abwx]

Intel 3 should be roughly 18% better perf/watt than what you calculated. And also are you sure it will be 3 GHz?

I m talking of the technical possibilities given what we know of Intel 7 and current e cores and adding the claimed 20% perf/watt improvement they claim for Intel 4 if that s the process used.

That being said i do no speculation on expected products, i ll just add that 288 cores@400W@intel 4 is possible only if base frequency is something like 2GHz.

 

[dullard]

I m talking of the technical possibilities given what we know of Intel 7 and current e cores and adding the claimed 20% perf/watt improvement they claim for Intel 4 if that s the process used.

That being said i do no speculation on expected products, i ll just add that 288 cores@400W@iNTEL 4 is possible only if base frequency is something like 2GHz.

Intel 4 isn't the process used. They are using Intel 3.

 

[DrMrLordX]

If it is really on 20A, numbers we were "given" don't really make sense at all.

Unless 20a isn't much better than Intel 3 or even Intel 4.

 

[Abwx]

Intel 4 isn't the process used. They are using Intel 3.

So grossly 600W for 288 cores@3GHz instead of 700W.

 

[dullard]

So grossly 600W for 288 cores@3GHz instead of 700W.

But again you are making the assumption that this chip will run way past the point of diminishing returns. I edited my post above for more detail: https://forums.anandtech.com/thread...ure-lakes-rapids-thread.2509080/post-41076438

An E core's sweet-spot of performance/watt is right at about 1.4 W per core. You were calculating them at 2.5 W per core. Sure, it could run that high, but why would you?

 

[H433x0n]

So grossly 600W for 288 cores@3GHz instead of 700W.

Sierra Glen != Gracemont

You're comparing it to a core that's overvolted with its frequency pushed way out of its efficiency range. It's a different core on a significantly better node. It could very well hit 3ghz or get very close to it and remain in that 400W power envelope.

 

[Joe NYC]

Arrow lake is 20A confirmed 😄✅ .. lunar lake is looking good

20A or TSMC N3.

There are multiple versions of the CPU chiplet.

 

[Abwx]

But again you are making the assumption that this chip will run way past the point of diminishing returns. I edited my post above for more detail: https://forums.anandtech.com/thread...ure-lakes-rapids-thread.2509080/post-41076438

An E core's sweet-spot of performance/watt is right at about 1.4 W per core. You were calculating them at 2.5 W per core. Sure, it could run that high, but why would you?

At 2GHz 288 cores would be within 300-350W, but then you have to look at the perf, that s a low throughput overall.

Sierra Glen != Gracemont

You're comparing it to a core that's overvolted with its frequency pushed way out of its efficiency range. It's a different core on a significantly better node. It could very well hit 3ghz or get very close to it and remain in that 400W power envelope.

Have you numbers that suggest that Intel 3 is significantly better than say TSMC s N5.?.

If anything Intel say that it will be about 42% better than Intel 7, that would barely close the gap if we are to look at current number between N5 and Intel 7.

 

[Geddagod]

If anything Intel say that it will be about 42% better than Intel 7, that would barely close the gap if we are to look at current number between N5 and Intel 7.

Why do you think that's true