Intel Sapphire Rapids 56-Core ES Processor Boosts to 3.3 GHz at 420 Watts

[AleksandarK]

Intel is slowly transitioning its data center customers to a new processor generation called Sapphire Rapids. Today, thanks to the hardware leaker Yuuki_ans we have more profound insights into the top-end 56-core Sapphire Rapids processor and its power settings. According to the leak, we have information on either Xeon Platinum 8476 or Platinum 8480 designs that are equipped with 56 cores and 112 threads. This model was running at the base frequency of 1.9 GHz and a boost frequency of 3.3 GHz. Single-core can boost to 3.7 GHz if the report is giving a correct reading. Remember that this is only an engineering sample, so the final target speeds could differ. It carries 112 MB of L2 and 105 MB of L3 cache, and this sample was running with 1 TB of DDR5 memory with CL40-39-38-76 timings.

Perhaps the most exciting finding is the power configuration of this SKU. Intel has enabled this CPU to consume 350 Watts in PL1 rating, with up to 420 Watts in PL2 performance mode. The enforced BIOS power limit rating is set at an astonishing 764 Watts, which could happen with AVX-512 enabled. Final TDP ratings are yet to be disclosed; however, these Sapphire Rapids processors are shaping to be relatively power-hungry chips.


View at TechPowerUp Main Site | Source

 

[Rares]

This is not innovation anymore is regression. Sapphire Rapids is DOA.

 

[Fourstaff]

Alder Lake - P for mobile processor has 6+8 P/E cores for 28w base and 64w boost - If we are using the same power budget then Sapphire Rapids TDP should be ~225w for PL1 instead of 350w? Frequency is 1.8Ghz/4.8Ghz base/boost for Alder Lake - P, roughly the same as this. Chip overhead power consumption sounds quite high?

 

[Valantar]

Alder Lake - P for mobile processor has 6+8 P/E cores for 28w base and 64w boost - If we are using the same power budget then Sapphire Rapids TDP should be ~225w for PL1 instead of 350w? Frequency is 1.8Ghz/4.8Ghz base/boost for Alder Lake - P, roughly the same as this. Chip overhead power consumption sounds quite high?

Interconnects consume a lot of power at this type of scale, and Intel does seem to be on a roll with poorly implemented high core count server hardware. Of course it's technologically very different, but Epyc Milan can easily exceed 100W just in Infinity Fabric power. This should be more efficient than that (fewer interconnects, shorter travel distances, EMIB instead of signalling through the package), but we don't know the details yet.

Still, I can't really imagine anyone wanting this over even Epyc Milan, let alone Genoa.

 

[Sabotaged_Enigma]

Good job Intel, keep pushing *everything* to the next level.

 

[Chrispy_]

I find it amusing that AMD launched an unusual, 'odd-number of CCD's' 56-core EPYC a year ago, almost exactly when news of Intel's Sapphire Rapids core count first leaked. Spite, or taunt? It doesn't matter, I still find it amusing!

The real question is whether Sapphire Rapids will be out before the current 7773X is superseeded by Zen4 Genoa. With 768MB of cache, and even without Linux kernel optimisations for the extra cache, the 7773X is an absolute beast at the sort of database/simulation work that these chips often get purchased for. Just like the 5800X3D, the 3D V-Cache has no benefit in situations like encoding or tile-based rendering, but if you're buying a server CPU to do what a GPU should be doing at 20x the efficiency, then you're an idiot, or a clueless PHB.

 

[Fourstaff]

I find it amusing that AMD launched an unusual, 'odd-number of CCD's' 56-core EPYC a year ago, almost exactly when news of Intel's Sapphire Rapids core count first leaked. Spite, or taunt? It doesn't matter, I still find it amusing!

View attachment 243970

The real question is whether Sapphire Rapids will be out before the current 7773X is superseeded by Zen4 Genoa. With 768MB of cache, and even without Linux kernel optimisations for the extra cache, the 7773X is an absolute beast at the sort of database/simulation work that these chips often get purchased for. Just like the 5800X3D, the 3D V-Cache has no benefit in situations like encoding or tile-based rendering, but if you're buying a server CPU to do what a GPU should be doing at 20x the efficiency, then you're an idiot, or a clueless PHB.

Answer you are looking for is die harvesting. 64 cores are built with 8x8 chiplets/clusters/etc, if one of them is faulty they can just disable it to give 8x7 = 56. If 2 of them are faulty, that gives you 8x6 = 48, and I bet some of the 32's are 8x4s.

 

[user556]

"... these Sapphire Rapids processors are shaping to be relatively power-hungry chips."
Same as the 12 series.

 

[Vya Domus]

Answer you are looking for is die harvesting. 64 cores are built with 8x8 chiplets/clusters/etc, if one of them is faulty they can just disable it to give 8x7 = 56. If 2 of them are faulty, that gives you 8x6 = 48, and I bet some of the 32's are 8x4s.

I doubt that's the case, it's probably just product segmentation. From their point of view it doesn't make sense to think of it in terms of "a chiplet is broken in this CPU, we're just going to disabled it". The chiplets go through a validation process before they are packaged in the final product, so a 56 core CPU is almost definitely intentional and not the result of defects. It's not like if you have a defective chiplet you absolutely have to put it into a CPU and disable it, it's too wasteful, that's the point of having chiplets, it just means you have to replace it with a working one. The defective ones probably end up in Ryzen products.

 

[TheinsanegamerN]

Interconnects consume a lot of power at this type of scale, and Intel does seem to be on a roll with poorly implemented high core count server hardware. Of course it's technologically very different, but Epyc Milan can easily exceed 100W just in Infinity Fabric power. This should be more efficient than that (fewer interconnects, shorter travel distances, EMIB instead of signalling through the package), but we don't know the details yet.

Still, I can't really imagine anyone wanting this over even Epyc Milan, let alone Genoa.

The other thing to consider is the DDR5 controller. Based on alder lake, it seems that controller is quite a hot headed little thing, and on sapphire rapids with 8 channel controllers that's gotta eat some watts.

 

[ncrs]

I doubt that's the case, it's probably just product segmentation. From their point of view it doesn't make sense to think of it in terms of "a chiplet is broken in this CPU, we're just going to disabled it". The chiplets go through a validation process before they are packaged in the final product, so a 56 core CPU is almost definitely intentional and not the result of defects. It's not like if you have a defective chiplet you absolutely have to put it into a CPU and disable it, it's too wasteful, that's the point of having chiplets, it just means you have to replace it with a working one. The defective ones probably end up in Ryzen products.

They are not disabling an entire chiplet in the 56 core version - they are disabling one core in each of the 8 chiplets. Both L3 cache amount and TDP give clues to this, but this information has been disclosed by AMD anyway and is on Wikipedia. There are 7 (out of 23 available in retail, but cloud vendors get special SKUs) models in the Milan(-X) generation that have 4 active chiplets instead of 8, with the other 4 spots being populated by blank pieces for mechanical stability.
Even the 8 core EPYCs have 8 chiplets, because the L3 cache is too good to waste and memory bandwidth would be imbalanced with just 1 chiplet for the whole CPU. For the latter I mean the InfinityFabric bandwidth, not the IMC since those are on the IO die. One chiplet wouldn't be able to consume the whole bandwidth of the IMCs.

 

[Solaris17]

These are server processors where power consumption, heat and sound are the lowest on the docket. They are bought for compute.

I guess I don’t understand the hate. Must be the audience.

 

[Chrispy_]

and I bet some of the 32's are 8x4s.

They absolutely are.

I just found it strange that AMD went for an odd number of enabled CPU cores in a single CCD, since that only did that for 7, but not 5 or 3. The timing of the single EPYC 7663 was coincidentally close to the first news of Sapphire Rapids topping out at 56 cores, and the 7663 is a single, isolated example of a 7-core CCD across all of the several hundred Zen-architecture CPUs models ever made.

• One singular model only, no low-TDP, or high-clock variants of it.
• The first odd-numbered core count in any Zen-based CCD, ever.
• Released in the same week that Intel confirmed the leak of Sapphire Rapids having 56 cores was accurate.
• The InfinityFabric root down the center of each CCD has a pair of cores off either side. It's apparently difficult to isolate one of these cores from the other as much of that IF interconnect is shared between both. This is kind of backed up but AMD never making any 7, 3, or 5-core Ryzens. If I had to guess, 56-core models are built with fully-functional CCDs where one core is simply told not to process any calculations in microcode to meet power-budget-per-core (or spite/brag) requirements.

I mean, there's no hard evidence, but I can take an educated guess that the 7763 is a 'spite' or 'brag' edition to prove to shareholders/customers/Intel that AMD already has the 56-core product on the market and we won't see other 7-core CCDs either in desktop, mobile, or servers without good reason.

 

[AnarchoPrimitiv]

The other thing to consider is the DDR5 controller. Based on alder lake, it seems that controller is quite a hot headed little thing, and on sapphire rapids with 8 channel controllers that's gotta eat some watts.

Yeah, but Genoa has a 12 Channel DDR5 controller and in leaks, it's said that the 96 Core model will have a 320 watt TDP that's configurable up to 400 watts, so we're talking 3.33 - 4.16 Watts/core (I'm not saying that's how the CPU actually uses power, I'm just doing some quick math to compare Genoa and Sapphire Rapids with a common unit of measurement). For this sapphire rapids 56 core model, if we use the 420 watt PL2 figure, we get 7.5 watts/core. Now, the Milan 7763 64 core model has a boost frequency of 3.5Ghz, so based on the fact that Genoa will be moving to 5nm CCDs and AMD has been hyping up 5ghz+ frequencies for Zen4, I think it's safe to say that Genoa will boost higher than Milan, and I wouldn't be surprised if they reach 4.0Ghz+ on many SKUs, even the 96 core model all while keeping that TDP between around 320 watts....it really shows how power hungry Sapphire Rapids is going to be, and while that apparently doesn't matter to a lot of people in the desktop segment, it does matter in the enterprise segment.

 

[Adam Krazispeed]

Not impressive at all and i bet Zen4 Genoa will be out before Zen4 AM5 Desktop platforms!! so Intel is LATE once again and way too high power!!!!!!!

is intel and nvidia fighting for highest power draw as a higher is better thing or what cuz nvidia 40 series 600watts and intel sapphire rapids 350watts Nvidia wins lol 180w is too much for me on cpu and 350 is too high on a gpu ) 6800xt or less (maybe 6900xt as same pwr as 6800xt ?)

 

[Crackong]

420W and that thing must be watercooled

 

[TheinsanegamerN]

These are server processors where power consumption, heat and sound are the lowest on the docket. They are bought for compute.

And those people are buying milan, which quite frankly ROFLstomps intel in most compute tasks. Only the niche of niche users, those who wants AVX 512, are well served by intel at this point. Genoa is going to nuke this thing from orbit.

I guess I don’t understand the hate. Must be the audience.

It's not "hate", it's dunking on inferior projects launched by a floundering chip giant with a side of shaudenfreud watching the blue gorilla that used underhanded tactics for years get pounded into the sand over and over again.

 

[Solaris17]

And those people are buying milan, which quite frankly ROFLstomps intel in most compute tasks. Only the niche of niche users, those who wants AVX 512, are well served by intel at this point. Genoa is going to nuke this thing from orbit.

It's not "hate", it's dunking on inferior projects launched by a floundering chip giant with a side of shaudenfreud watching the blue gorilla that used underhanded tactics for years get pounded into the sand over and over again.

Odd I’m literally surrounded by tens of thousands of servers and the amount of AMD usage is relatively low. Maybe they are used more in private data centers or on prem.

 

[DeathtoGnomes]

Odd I’m literally surrounded by tens of thousands of servers and the amount of AMD usage is relatively low. Maybe they are used more in private data centers or on prem.

Yeah but how old are those servers?

These are server processors where power consumption, heat and sound are the lowest on the docket. They are bought for compute.

I guess I don’t understand the hate. Must be the audience.

is that a hint of sarcasm? 420w for compute seems like an uncalculated tradeoff.

 

[Solaris17]

Yeah but how old are those servers?

Pretty new we use ice lakes for the most part every multi socket system runs intel. We have a sprinkling of AMD but those are 7xx1 or 7xx2 and only go in 1U short depth chassis think h11/h12 of course we do thise machines as intel as well e3-2xxG we do have some older racks those are e3-xxxxv4 skylakes though.

Yeah but how old are those servers?


is that a hint of sarcasm? 420w for compute seems like an uncalculated tradeoff.

Oh Did it sound like it? No I wasn’t joking. It’s not like you hang out or have desks in a DC we have routers and switches that pull more power than tjat and SANs. BIOS settings are generally set to max power consumption via automation. Excluding collocation who knows what they do with there systems.

 

[Punkenjoy]

All the Epyc server we have in our company are Azure Nodes. the purchase team looked at them to renew our on premise environment and it wasn't easy. They were priced very high, mostly non-negotiable and they would not ship quickly.

Intel on the other side, are negotiable and available right away.

AMD seems to be selling them like hot cake to cloud vendor and don't have that much left for the business market.

 

[Panther_Seraphin]

These are server processors where power consumption, heat and sound are the lowest on the docket. They are bought for compute.

Actually Perf/Watt is king in the data center.

AMD seems to be selling them like hot cake to cloud vendor and don't have that much left for the business market.

I wonder why they are being bought out like crazy? /s

Also the fact that Intel have their own fabs dedicated to production makes it far easier to get stock out the door. When the server chips are also fighting against GPU's, APU's, Console Parts, Steam Deck Parts, Consumer Parts etc on top of the 1-2 year lead times. You bet if AMD could go back 12 months and double their order books they would.

 

[Deleted member 24505]

These are server processors where power consumption, heat and sound are the lowest on the docket. They are bought for compute.

I guess I don’t understand the hate. Must be the audience.

Because it says intel on it, be different if it was AMD

get pounded into the sand over and over again.

Odd I’m literally surrounded by tens of thousands of servers and the amount of AMD usage is relatively low

I don't see no pounding here

 

[eidairaman1]

"... these Sapphire Rapids processors are shaping to be relatively power-hungry chips."
Same as the 12 series.

Pentium 4 preshot anyone?

 

[Dr_b_]

"Alder Lake. All of a sudden...Boom! We are back in the game. AMD in the rearview mirror in clients, and never again will they be in the windshield; we are just leading the market" ....

IN power consumption, yes you are correct Pat, if you ever do release SR that is....