Discussion Intel current and future Lakes & Rapids thread

[Doug S]

Charlie had an article yesterday about SPR's successor, Emerald Rapids. He said Intel still maintains it will come around a year after SPR, but I wonder how much of that is "we need SPR on the market long enough to make all this worth it" versus "Emerald Rapids is trouble plagued too so that number is just a hope and a dream".

If Emerald Rapids really is a year away today they should have a good idea of how well it is working, and if things are going well at some point it makes sense to throw in the towel on SPR and concentrate all your resources on pulling in Emerald Rapids a quarter or two instead.

 

[Markfw]

Fastest Single Thread Performance posted so far.

There is another entry where the MT is 78,000 points but the ST is 36% lower

Quanta Cloud Technology Inc. QuantaGrid D54Q-2U - Geekbench

Benchmark results for a Quanta Cloud Technology Inc. QuantaGrid D54Q-2U with an Intel Xeon Platinum 8480+ processor.

browser.geekbench.com

Also, you can bet that 2P system is running over 1000 watts vs the genoa at probably 300. or less.

 

[Exist50]

Also, you can bet that 2P system is running over 1000 watts vs the genoa at probably 300. or less.

Sapphire Rapids is up to 350W per socket, while Genoa is up to around 360W (socket and some SKUs up to 400W). We can reasonably assume the top CPUs on each platform are in the highest TDP brackets.

 

[nicalandia]

Charlie had an article yesterday about SPR's successor, Emerald Rapids. He said Intel still maintains it will come around a year after SPR,

The latest info we have is that Intel plans to Ramp Up both SPR-SP and EMR-SP at the same time. This means that Emerald Rapids could be release at mid to late 2023

 

[nicalandia]

Also, you can bet that 2P system is running over 1000 watts vs the genoa at probably 300. or less.

closer to 800 Watts for those Intel 2S Systems.

 

[Markfw]

closer to 800 Watts for those Intel 2S Systems.

and 1/2 that for Genoa at the same performance. You think data centers are going to deal with that without considering Genoa ? Power nowadays is king. Not just me, its a real consideration now.

 

[jpiniero]

If yields were actually the problem, why would they only delay the launch of the MCC die, which has ~half the silicon of a full XCC package? And why would those yield issues not apply to ICL-SP, ADL, RPL, etc? And how would a single quarter be enough to materially change them? There's a lot of things that make no sense here.

Because the bigger the die, the more painful having defects is. This is why doing a massive monolithic die is such a bad idea.

 

[moinmoin]

It's really the Whole Enchilada...! Lower Yields and Bugs that are compounding the issue. more than 13 Steppings is something really unheard of before.

The latest info we have is that Intel plans to Ramp Up both SPR-SP and EMR-SP at the same time. This means that Emerald Rapids could be release at mid to late 2023

I'm still really curious how all that can lead to continuous delays to SPR but not affect EMR as much to the point that their launch time converge. EMR a year after SPR seems more realistic.

 

[nicalandia]

Sorry guys I messed Up..

This is the Sapphire Rapids-SP held by Sandra Rivera(Intel Executive)

148 Complete Dies



This one is for the Massive 34C Monolithic Die with only 68 Complete Dies




So 50% to 60% Yields for SPR-SP is between 74 - 90 fully functional dies so between 15 and 20 CPUs per Wafer which is still very low.

 

[Exist50]

Because the bigger the die, the more painful having defects is. This is why doing a massive monolithic die is such a bad idea.

More silicon and especially more non-redundant silicon makes defects more painful. But Intel's been shipping large monolithic dies on a worse 10nm process in the form of Ice Lake server for a while now. And if you believe the rumor, they seem willing to ship the XCC with ~2x the silicon per package. If it were a yield issue, how are these reconcilable? And then how would they fix it in merely a quarter?

 

[jpiniero]

More silicon and especially more non-redundant silicon makes defects more painful. But Intel's been shipping large monolithic dies on a worse 10nm process in the form of Ice Lake server for a while now.

Icelake Server cores and the monolithic dies themselves are much smaller than Golden Cove Server. And they probably aren't selling that much Icelake, given how far the server volume has fallen.

Another quarter might give them extra time to build up some inventory so they have something to sell. At least initially. Or to buy extra time to decide whether to delay it even more.

 

[Exist50]

Icelake Server cores and the dies themselves are much smaller than Golden Cove Server

The big die is roughly 628mm2. That is quite large by any definition, and they launched it a year and a half ago. And right now, you can go out and buy ~250mm2 of fully enabled Intel 7 for a few hundred bucks, no issue.

And they probably aren't selling that much Icelake, given how far the server volume has fallen.

But then why couldn't they do the same low volume shipments for SPR?

Another quarter might give them extra time to build up some inventory so they have something to sell. At least initially.

I could see them trying to build up volume, but that seems a bit different from claiming yield issues. Also, the current market isn't exactly itching to buy.

 

[jpiniero]

And right now, you can go out and buy ~250mm2 of fully enabled Intel 7 for a few hundred bucks, no issue.

That might be the other issue - client parts are more/actually profitable.

 

[LightningZ71]

Keep in mind, Ice Lake is on an earlier version of the 10nm process (the second generation, which Intel wants to call 10nm, trying to forget cannon lake). It has notable power issues on top of whatever yield issues it may have. Tiger Lake was on a later version called 10sf. That improved on the power and yield characteristics markedly over 10+. AlderLake and Saphire Rapids are on Intel 7, previously known as 10ESF. These are different processes that share some design rules, but are built on different lines with different yields and performance profiles.

 

[Exist50]

That might be the other issue - client parts are more/actually profitable.

Well they should have the capacity to do both. And their server margins suck because they haven't actually been shipping SPR.

 

[Doug S]

The latest info we have is that Intel plans to Ramp Up both SPR-SP and EMR-SP at the same time. This means that Emerald Rapids could be release at mid to late 2023

What's the point of that? Isn't EMR supposed to be better than SPR? That's like if Apple ran into issues with M2 Max that delayed it until M3 Max was ready...they would just skip M2 Max entirely if that happened.

 

[nicalandia]

What's the point of that? Isn't EMR supposed to be better than SPR?

Yes, but not by much and Intel seem to think that having two different generations of about similar product lines will not have an impact due to different customer's need.

 

[poke01]

oh my. companies should go for AMD. They have roadmap that works.

 

[IntelUser2000]

From Intel's latest article: https://www.intel.com/content/www/u...tegy-design-new-xeon-processor.html#gs.gwv54z

Moving forward, now that 4th Gen Intel Xeon Scalable processor volume SKUs have met product release qualifications and the company is starting to ramp-up deployment, Intel – along with its partners and customers – will be hosting a launch event on Jan. 10. Additional details to follow.

Sapphire Rapids official launch date?

 

[DrMrLordX]

Also, you can bet that 2P system is running over 1000 watts vs the genoa at probably 300. or less.

Genoa is going to use more power than that per socket, depending on the SKU. We'll get a better idea of the exact numbers later, but assume that Genoa can use up to 350-400W per socket.

 

[Markfw]

Genoa is going to use more power than that per socket, depending on the SKU. We'll get a better idea of the exact numbers later, but assume that Genoa can use up to 350-400W per socket.

Since Milan does at the most 280, and future nodes should do better on power, I would doubt they go over 320 or so got 96 cores. But we will see.

 

[Exist50]

Since Milan does at the most 280, and future nodes should do better on power, I would doubt they go over 320 or so got 96 cores. But we will see.

These numbers have been leaked, if not outright published, several times. And I'm not sure why after Raphael you'd doubt that they'd need ≥350W for +50% cores + DDR5 + PCIe 5.0/CXL. I'd imagine the cores themselves only get ~half of it.

 

[Markfw]

These numbers have been leaked, if not outright published, several times. And I'm not sure why after Raphael you'd doubt that they'd need ≥350W for +50% cores + DDR5 + PCIe 5.0/CXL. I'd imagine the cores themselves only get ~half of it.

Servers are designed and set to to be more efficient. I run my 7950x at 142 watt. It smokes my new Milan 7763's in performance and they are 280 watt,

Let just see benchmarks when they are both available. But even at 350, its still more efficient by 100% than SR. (350 vs dual 350)

 

[Exist50]

Servers are designed and set to to be more efficient

Yes, but there're limits. 350W would be only ~3.6W/core if you ignore IO entirely. If, as a first order estimate, half the power is going to fabric and IO, that's <2W a core, or similar to (if not less than) a smartphone core's sustained power. It would be pointless to push a core below its power at Vmin, so I'd imagine that will set the floor.

But yes, I expect we will see a ~2x efficiency gap.

 

[DrMrLordX]

Since Milan does at the most 280, and future nodes should do better on power, I would doubt they go over 320 or so got 96 cores. But we will see.

Early spec leaks indicated around 360W:

https://www.tomshardware.com/news/specs-of-amd-epyc-genoa-cpu-leak

Intel isn't the only one pushing up per-socket power draw.