A violent delidding reveals that Intel's Sapphire Rapids CPUs could have up to 80 cores

mongeese

Posts: 525   +111
Staff member

Eighty cores is twice the amount you can currently purchase from Intel for up to $8,099, though it’s only a third more than what AMD could sell you for $7,890. But Intel’s flashiest Xeons and AMD’s shiniest Epycs are more similar than you'd imagine; they share the same clock speeds, the same memory options, and the same TDPs. AMD has an upper hand in PCIe connectivity, but Intel’s got better on-chip connectivity, etc.

Sapphire Rapids, Intel’s fourth-generation Xeons, share more emerging technologies with GPUs than with other CPUs. Nvidia’s five-figure A100, for example, is the only GPU to offer up to 80 GB of HBM2e memory. Sapphire Rapids has 64 GB of HBM2e onboard (and eight lanes of DDR5 memory, too). The A100 was the first GPU to support BF16 computation. Sapphire Rapids will support BF16, too.

Sapphire Rapids also has PCIe 5.0, a new core architecture called Willow Cove, the 10nm Enhanced SuperFin node, and the new LGA4677-X socket on its resume.

All this is public information that Intel has divulged, mostly intentionally. Intel had begun sampling Sapphire Rapids to partners by the end of last year, Intel’s now-former CEO Bob Swan said to investors in January. Multiple sources have said that the sample models had 28 cores enabled.

The story takes a violent turn

At the end of January, a regular leaker on the Chinese site Bilibili gained ownership over one of the disseminated samples. He took some photos, then delidded the processor in the most awful way possible. Twenty seconds into this clip, he puts the processor in a regular bench vice and just starts cranking. Before the minute mark, he’s heating it over a gas stovetop flame, and then – brace yourself – he accidentally drops the processor into the flame. (If you’re curious, an oven is usually considered best practice, partly to avoid this issue.)

Afterward, the heat spreader comes off quite easily. An autopsy reveals that the heat spreader was soldered to the die using indium, like Rocket Lake.

Under the heat spreader was a quad-chiplet formation. It’s theorized, but not with much certainty, that Intel is using EMIB (embedded multi-die interconnect bridge) or Foveros to connect the chiplets and the HBM2e memory to gain an inter-chip bandwidth advantage over AMD’s Infinity Fabric.

Leaker commits an atrocity

Last week, the leaker pulled all four of the chiplets off the substrate, then dissected them. It appears as if he successfully removed the HBM2e chips from three of the chiplets. The fourth… it’s best not to look.

Removing the HBM2e chips revealed the cores underneath. From the looks of it, each chiplet contains twenty cores arranged in a five by four grid. There also seem to be several controllers and possibly interconnect arrays around the perimeter of each chiplet.

The twenty cores took the leaker by surprise. It had been leaked, repeatedly, that Sapphire Rapids would top out at 56 cores. If that were the case, then Intel would be disabling six of the twenty cores on each chiplet, a proportion unprecedented in a flagship model. It’s possible that the 56-core leaks were simply wrong, and it’s also possible that not all of the arrangements that look like cores are in fact cores. The most depressing possibility is that Intel’s new node has an absurdly high failure rate.

Image credit: Michael Dziedzic, Yuuki_AnS

Permalink to story.

 

QuantumPhysics

Posts: 5,310   +6,036
When I build my next intel Extreme powered PC, we may be in the 13th or 14th generation. I want DDR5 RAM. I'm very excited.
 

yRaz

Posts: 3,857   +4,046
They may be disabling cores on the chiplets due to yield problems. That, and they may be accumulating fully working chiplets for 80core models. This will be interesting to watch, I look forward to seeing where this goes and AMDs response to it.

I know Intel is "behind" right now, but this has been the most exciting time in the CPU market since Intel best the FX-60 with the Core 2 duo
 

yRaz

Posts: 3,857   +4,046
Yeah, to find that one out of 100 that doesn't explode under load. Some yield problem.
Yield might not just be non-functional, it could be that they can't run at the advertised clock speed. Also, it said it was 10nm. Intel hasn't had the greatest success with 10nm. 4x14 sounds like a weird number, but it is inline with Intel's 28 core CPUs. Having 14 functionally high-speed cores on a 20 core chiplets could be fairly high on Intel's 10nm process. Just look at what AMD did with theadripper. It's a lot easier to have 4x16 core chiplets than it is one monolithic chip. They also use this method on their ryzen chips. Have an 8 core chiplet with only 6 functional cores.

I'm not criticising anyone, disabling cores on chiplets is an incredible way to economically increase yeilds
 
Last edited:

Theinsanegamer

Posts: 2,732   +4,278
80 cores that run hotter than the sun lets goooooo
Depends on clock speed. Keep comet and rocket lake out of boost and it's rather economical. These huge multi core chips rarely touch 4 GHz, let alone 5 like consumer chips. Pushing higher boosts requires a huge amount of power.
 

Dimitrios

Posts: 918   +719
I hope MICROSOFT doesn't get any idea's and make a new version of WINDOWS that we have to buy to take advantage of these new architectures.
 

dragosmp

Posts: 35   +44
This was a fun bit of investigation, but what about the elephant in the room - the power consumption. Intel today on the same node can barely power 40 cores within 270W. Intel has done "miracles" in the past, like the C2D vs FX60 mention above, but it feels far fetched they can cram twice the cores, 64GB! of HBM2, interconnect and all in a package that can be cooled. Call me skeptic, I'd love to be wrong.
 

Bullwinkle M

Posts: 568   +459
"I hope MICROSOFT doesn't get any idea's and make a new version of WINDOWS that we have to buy to take advantage of these new architectures."
-----------------------------------------
That depends on the hardware

If the hardware requires updates to Windows, Microsoft has historically refused to update older versions just to sell you a new version

For example, when Windows XP was still the most widely used O.S. on the planet, Microsoft refused to update XP for SSD partition offset's and 4K hard drive sectors

They went out of their way to kill XP, then 7, and are currently in the process of killing Windows 8.1

There was no need for Microsoft to kill Vista as it was dead on arrival
 

Toju Mikie

Posts: 220   +201
I hope MICROSOFT doesn't get any idea's and make a new version of WINDOWS that we have to buy to take advantage of these new architectures.
This is why I hate Windows Server. They already charge more money for more cores, and this puts AMD CPUs at a disadvantage since they have more cores. Users are forced to get a more expensive license with a higher core-count CPU.
 

Watzupken

Posts: 342   +320
This was a fun bit of investigation, but what about the elephant in the room - the power consumption. Intel today on the same node can barely power 40 cores within 270W. Intel has done "miracles" in the past, like the C2D vs FX60 mention above, but it feels far fetched they can cram twice the cores, 64GB! of HBM2, interconnect and all in a package that can be cooled. Call me skeptic, I'd love to be wrong.
I feel Intel is trying to pull an AMD by ramping up the core count to compete. I agree to your point that power will be sticking point for this processor. Even if Intel can push 80 cores out their door, I don't think their 10nm is capable of reining in the power and heat. So I expect the base clockspeed to be very low. Still and impressive feat if its true, but not sure if this can help them retain/ regain market share losses.
 

Lounds

Posts: 922   +818
I dread to think how much these will cost. AMD has the economics of chiplets right and has done it for 4 years now. Also AMD leaves all the IO stuff to the IO die in the centre of their eypc CPUs, so the chiplets themselves are even smaller having even better yields and more margin per core.
 

Bulllee

Posts: 229   +157
I dread to think how much these will cost. AMD has the economics of chiplets right and has done it for 4 years now. Also AMD leaves all the IO stuff to the IO die in the centre of their eypc CPUs, so the chiplets themselves are even smaller having even better yields and more margin per core.
As sure at the sh$t on your shirt tail I for one can't afford them.
 

Lionvibez

Posts: 2,440   +2,068
I feel Intel is trying to pull an AMD by ramping up the core count to compete. I agree to your point that power will be sticking point for this processor. Even if Intel can push 80 cores out their door, I don't think their 10nm is capable of reining in the power and heat. So I expect the base clockspeed to be very low. Still and impressive feat if its true, but not sure if this can help them retain/ regain market share losses.

I believe the enterprise version of Zen 4 will be going up to 96 cores so intel will still be behind but smaller gap this time, should still run hotter.
 

yRaz

Posts: 3,857   +4,046
This is why I hate Windows Server. They already charge more money for more cores, and this puts AMD CPUs at a disadvantage since they have more cores. Users are forced to get a more expensive license with a higher core-count CPU.
Most enterprise software is sold on a per-core basis these days. Having on rack with multiple cores is still more economical than have more racks with the same amount of licences
 

yeeeeman

Posts: 432   +384
I recall there were rumors of 72 cores somewhere...
People need to understand intel has stayed still about 4 years in the server space. Xeon 8280 with 28 cores is more or less an ancient product. Ice Lake brings that up to 40 and I don't see why Intel wouldn't push it to 2x if they use chiplets.
 

Nanochip

Posts: 45   +47
The 28 core sample could just be a binned chip with several cores fused off, similar to how intel does it on its consumer parts… one core design for multiple skus. On comet lake for instance, top sku is 10 cores, then intel fuses both cores and igpus off to bin the various skus like the 10900F (no igpu) or 10600K (4 cores fused off).

Why sample the 80-core sapphire rapids unit when it could leak like this one has? We should read too much into these leaks until we have final silicon to inspect and comment on.