News [Toms] Intel Announces Socketed 56-Core Cooper Lake Processors

[Hitman928]

https://www.tomshardware.com/news/i...socketed-processor-cooper-lake-ice,40098.html

 

[Markfw]

So 6 months from now, they will come out with a processor that has 8 cores less than AMD. Well thats exciting....

And that 6 months is a maybe....

 

[DrMrLordX]

Well

56-cores is kind of a step up for Intel. Cascade Lake-AP was not available in sockets (you had to buy a "sled"). So now you can get your Cascade Lake-AP fix in a socket instead. It remains to be seen if the market will embrace these chips. They may also have hardware mitigations for Zombieload.

 

[BigDaveX]

Well

56-cores is kind of a step up for Intel. Cascade Lake-AP was not available in sockets (you had to buy a "sled"). So now you can get your Cascade Lake-AP fix in a socket instead. It remains to be seen if the market will embrace these chips.

Gonna depend on how they're priced. If they're 2x-3x the price of Rome for equal or slightly worse performance... well, they'll probably still get some takers thanks to AVX512 and the still-substantial number of buyers who won't stray from Intel, but I wouldn't exactly expect a stampede.

 

[Yotsugi]

well, they'll probably still get some takers thanks to AVX512

Barely anything uses AVX512, and what usually does would like more b/w than CPX could provide.

 

[amrnuke]

I wouldn't exactly expect a stampede.

I don't think there will be a stampede either way. Corporate inertia is so difficult to break. It's a slow slog.

 

[naukkis]

Gonna depend on how they're priced. If they're 2x-3x the price of Rome for equal or slightly worse performance... well, they'll probably still get some takers thanks to AVX512 and the still-substantial number of buyers who won't stray from Intel, but I wouldn't exactly expect a stampede.

To be only 2-3x price of Rome Intel needs to check their prices - and by big factor....

 

[ehume]

Cooper Lake will use the same socket -- LGA4189? -- as Icelake. Is that because one of them will have a limited distribution?

 

[Yotsugi]

Corporate inertia is so difficult to break. It's a slow slog.

AMD is targeting CSPs first so that stage is passed.

 

[Atari2600]

https://www.tomshardware.com/news/i...socketed-processor-cooper-lake-ice,40098.html

Intel plans to heat the place so much they'll melt any icebergs that might be in their road.

Full steam ahead, the unsinkable Intel Itanic lives on!!!

 

[BigDaveX]

Cooper Lake will use the same socket -- LGA4189? -- as Icelake. Is that because one of them will have a limited distribution?

Probably Intel have realised that two-die MCM Xeons are their only hope of avoiding being completely obliterated by Rome and its successors, but LGA3647 doesn't have the physical space and/or electrical infrastructure to support such chips, so they threw yet another new socket together. Plus, the potential for a drop-in upgrade to Icelake if it turns out any good may help sweeten the deal.

In other words, Intel are falling back on their mid-2000s playbook.

 

[Kenmitch]

Intel plans to heat the place so much they'll melt any icebergs that might be in their road.

Full steam ahead, the unsinkable Intel Itanic lives on!!!

400w TDP....Holly crap! I can't wait to see those perf/watt benchmarks against Rome.

 

[positivedoppler]

is this a single 56 core die or two 28 cores glued together. What is the yield on a massive 56 core die at 14nm.

 

[Yotsugi]

is this a single 56 core die or two 28 cores glued together. What is the yield on a massive 56 core die at 14nm.

The latter and you can't make 56c single die on 14, it's gonna be bigger than the reticle.

 

[jur]

It is really beyond me, why doesn't Intel take the same approach as Amd. They are clearly capable of making <100mm^2 dies on 10nm, so an 8 core Icelake chiplet shouldn't be a problem to produce.

The era of large dies seems to be getting to an end, because process complexity and defect rate per mm^2 are going up.

 

[joesiv]

It is really beyond me, why doesn't Intel take the same approach as Amd. They are clearly capable of making <100mm^2 dies on 10nm, so an 8 core Icelake chiplet shouldn't be a problem to produce.

The era of large dies seems to be getting to an end, because process complexity and defect rate per mm^2 are going up.

Takes times to develop these things like this. AMD has been working on chiplets for a long time, and have been able to bring it to market first.

Though Intel has also been working on multichip designs (FOVEROS), they're just a bit behind AMD at brining them to market. The next few years should be really interesting.

It's definitely a reactionary move by Intel. In the past intel probably had designs up their sleeve, where they could release things when they wanted, and most designs likely never saw the light of day even. Now, it's almost like we're watching intels reaction to AMD real time, as they iterate and develop their products. Before Cooper Lake, they had the platinum 9200, to react to the ROME announcement, now they have iterated down to more end user platform/socket.

 

[DrMrLordX]

To be only 2-3x price of Rome Intel needs to check their prices - and by big factor....

Intel rarely charges "list price" to anyone except really small buyers. If you're a big boy making big orders, they will negotiate down on price. Especially through ODM channels where details of transactions go largely unreported.

Cooper Lake will use the same socket -- LGA4189? -- as Icelake. Is that because one of them will have a limited distribution?

Icelake-SP will likely be next-to-non-existent in server rooms. There will be enough out there to keep Intel looking relevant and get some benchmarks up on all the key websites. The real product for the new socket is Cooper Lake. Sharing the socket with Icelake-SP keeps costs down since Intel doesn't have to produce a separate socket for Icelake-SP.

is this a single 56 core die or two 28 cores glued together. What is the yield on a massive 56 core die at 14nm.

2x28. The difference here is the "glue" on Cooper Lake is EMIB I think? Cascade Lake-AP doesn't use that. The electrical arrangement with Cascade Lake-AP is different, and there are some consequences for Cascade Lake-AP buyers in that regard.

BTW I would be angry had I been one of the few adopters of Cascade Lake-AP since that product offers no real advantage over Cooper Lake. Unless you really just love those sleds for density purposes.

It is really beyond me, why doesn't Intel take the same approach as Amd. They are clearly capable of making <100mm^2 dies on 10nm, so an 8 core Icelake chiplet shouldn't be a problem to produce.

The era of large dies seems to be getting to an end, because process complexity and defect rate per mm^2 are going up.

Cooper Lake appears to be Intel's first foray into EMIB in the server world. I think? It's nothing Earth-shattering, it just lets them cram 56 cores under one IHS by using two dice without doing what they did with Cascade Lake-AP.

As for using a bunch of little chiplets together using EMIB/Foveros, I think you have to wait for Sapphire Rapids to see that in action.

 

[Kenmitch]

Though Intel has also been working on multichip designs (FOVEROS), they're just a bit behind AMD at brining them to market. The next few years should be really interesting.

It will be interesting to see what comes of it in the future. It would be interesting to see if there are some underlying patents that make it more difficult.

 

[jpiniero]

Cooper Lake will use the same socket -- LGA4189? -- as Icelake. Is that because one of them will have a limited distribution?

There's two different sockets actually (for the Cooper/Icelake Server generation). The LGA 4189 might be the smaller one.

 

[Yotsugi]

There's two different sockets actually (for the Cooper/Icelake Server generation). The LGA 4189 might be the smaller one.

They're both 4189, just keyed differently.
Different platforms too, ICL/CPX-4 is Whitley, CPX-6 is Cedar Island.

 

[NTMBK]

One particular reseller, when asked if they expect to sell even one unit, said ‘no’.

How to Tarnish Platinum: Sell It as Xeon 9200

www.anandtech.com

Yikes. A pretty scathing write up from Ian. Definitely sounds like this product was built for marketing, not for customers.

EDIT: Wait, sorry, this thread was for the socketed version, not the BGA.

 

[DrMrLordX]

@NTMBK

It's also Cascade Lake, not Cooper Lake. Intel was already given a good flogging by SemiAccurate months ago over the "sled" configurations, lack-of-availability, and other problems observed by Ian in the linked article. At least there's one vendor that's going to try to sell Cascade Lake-AP to the public, which is certainly a change from ~6 months ago.

That being said, if Intel tries the same crap with Cooper Lake (creating Cooper Lake-AP) then I don't even know what to say.

 

[liahos1]

So 6 months from now, they will come out with a processor that has 8 cores less than AMD. Well thats exciting....

And that 6 months is a maybe....

they will probably sell more of this product than amd will sell of the 64c rome chip they have.

 

[joesiv]

they will probably sell more of this product than amd will sell of the 64c rome chip they have.

The article doesn't paint a good picture for vendors/buyers being excited about the platform. Specifically the water cooling requirement and lack of flexibility (soldered CPUs make upgrades/replacements impossible). Who knows the cost.

 

[liahos1]

thast cas

The article doesn't paint a good picture for vendors/buyers being excited about the platform. Specifically the water cooling requirement and lack of flexibility (soldered CPUs make upgrades/replacements impossible). Who knows the cost.

that is cascade lake not cooper lake.