Next-gen Xeon CPU leaks with 56 cores and 350W TDP

[Tudor Cibean]

In context: While Intel still commands a massive 90% share of the x86 server CPU market, AMD has been gaining ground ever since releasing its Zen-based Epyc chips. Intel hopes to put a halt to that with their upcoming Xeon Sapphire Rapids lineup, expected to launch later this year.

Intel's next generation of Xeon data center processors, codenamed Sapphire Rapids, just got leaked. Like Intel's consumer Alder Lake lineup, it will feature the chipmaker's Golden Cove architecture and Intel 7 process node, previously known as 10 nm Enhanced SuperFin.

After several delays, the processors are expected to debut later this year, around the same time as Intel's mainstream Raptor Lake series. They will offer up to 56 cores and 112 threads and come with a maximum TDP of 350W. It's worth noting that the enforced BIOS power limit is set at a staggering 764W, a target it could reach under workloads using AVX-512.

As we can see from the leaked screenshots, the 56-core engineering sample features 112MB of L2 cache and 105MB of L3 cache. Intel will also launch models with up to 64GB of on-package HBM2e acting as an L4 cache. The leaked CPU also has a base clock of 1.9 GHz and a 3.3 GHz boost clock, although it can boost up to 3.7 GHz on a single core. Shipping products might run at different frequencies, as this is only an engineering sample.

This design could be either the Xeon Platinum 8476 or the Platinum 8480, as both are predicted to feature 56 cores and 112 threads. The CPU was tested on Intel's C741 (Emmitsburg) platform with 1TB of DDR5 memory running at CL40-39-38-76 timings and support for PCI-Express 5.0.

These CPUs will go up against AMD's Epyc 7004 (Genoa) series, which should arrive around the same time. Those will feature the new Zen 4 architecture, up to 96 cores and 192 threads, and are going to be built using TSMC's 5nm node.

Image credit: @yuuki_ans

Permalink to story.

https://www.techspot.com/news/94256-next-gen-xeon-cpu-leaks-56-cores-350w.html

 

[takaozo]

I cant imagine this CPU in a chassis smaller than 3U or 4U in dual CPU config, air cooled. Or maybe the integrators will usewater cooling. But still 2 PSU arround 2KW required.

 

[Dimitriid]

I cant imagine this CPU in a chassis smaller than 3U or 4U in dual CPU config, air cooled. Or maybe the integrators will usewater cooling. But still 2 PSU arround 2KW required.

I think some modern data centers are moving towards just water cooling instead because of the density they're achieving now.

Still now the fact that 56/112 is what they're planning on when Epyc is *already* 64/128 at lower tdp and will go up to an insane 96/192 fairly soon clearly shows how much of a better proposition going AMD is turning out to be right now: Datacenters will have to consider a move to water cooling to have a reasonable set up with Xeons or yes, updating platforms but keeping their installation and deployment the same and just switching over to Epycs instead.

 

[Squid Surprise]

I think some modern data centers are moving towards just water cooling instead because of the density they're achieving now.

Still now the fact that 56/112 is what they're planning on when Epyc is *already* 64/128 at lower tdp and will go up to an insane 96/192 fairly soon clearly shows how much of a better proposition going AMD is turning out to be right now: Datacenters will have to consider a move to water cooling to have a reasonable set up with Xeons or yes, updating platforms but keeping their installation and deployment the same and just switching over to Epycs instead.

The # of cores/threads aren't as important as the performance of each.... Will be interesting to see the benchmarks when they come out..

I wonder if there will be an HEDT offering based on this though - and how it will compare with Threadripper... I'm looking to upgrade in the next year or 2 and while I planned on Threadripper, it would be nice if Intel could compete and maybe drive the prices down a bit.

 

[HardReset]

The # of cores/threads aren't as important as the performance of each.... Will be interesting to see the benchmarks when they come out..

Server class software usually scales very well. Assuming perfect scaling, to achieve same performance with 56 cores against 96 cores, that 56 core part must have 70% higher performance per core. Not going to happen.

Efficiency wise (performance per watt), it looks like Intel might have problem matching even Epyc 7742 (Zen2, 7nm, 2018) part. And that's two generations behind Genoa (Zen4, 5nm, 2022?) that will also get 128 core part.

 

[envirovore]

Hot damn that's a beast!

 

[DukeJukem]

It's kind of ironic how energy related political subjects have been floating around for a while now yet products like these are being made. The energy crisis really seems like a big concern lol.

 

[ypsylon]

I think Intel lives in parallel Universe. In their reality there is no energy crisis in the World. No need to worry about cost of running their CPUs. For data centers provision for 700+W/socket vs 280W from AMD that's monumental amount of $$$ in energy upkeep over the whole year.

Look even at very conservative power draw you can run 2 EPYCs inside the same power envelope of one Xeon-SR@ AVX-512. Even worse for Chipzilla is that AMD is bringing (supposedly) AVX-512 to next gen EPYCs which will nullify basically the only niche advantage Intel still holds in HPC segment. If AMD can keep next gen EPYCs with AVX-512 inside 300W, then sorry, but Intel has nothing.

 

[Irata]

Server class software usually scales very well. Assuming perfect scaling, to achieve same performance with 56 cores against 96 cores, that 56 core part must have 70% higher performance per core. Not going to happen.

Efficiency wise (performance per watt), it looks like Intel might have problem matching even Epyc 7742 (Zen2, 7nm, 2018) part. And that's two generations behind Genoa (Zen4, 5nm, 2022?) that will also get 128 core part.

Let‘s not forget the additional chipset power consumption - Xeon still needs chipsets, Epyc doesn‘t and we‘re most likely not talking about 5W chipsets for the hpc platform.

 

[wiyosaya]

The # of cores/threads aren't as important as the performance of each.... Will be interesting to see the benchmarks when they come out..

I wonder if there will be an HEDT offering based on this though - and how it will compare with Threadripper... I'm looking to upgrade in the next year or 2 and while I planned on Threadripper, it would be nice if Intel could compete and maybe drive the prices down a bit.

For me, the question will be how much extra will sIntel charge for a version that is not PCI-e crippled compared to ThreadRipper or Epyc? In prior gen Xeons, that was SOP for sIntel.

IMO, sIntel must be desperate - this looks like a part that they are releasing with hopes of eliciting "Wow 56-cores! Lets run right out and convert our entire data center to these!!!!!" from their customers - all the while hoping that those same customers do not pay attention to what is under the hood of these Xeons in comparison to Epyc.

 

[Squid Surprise]

For me, the question will be how much extra will sIntel charge for a version that is not PCI-e crippled compared to ThreadRipper or Epyc? In prior gen Xeons, that was SOP for sIntel.

IMO, sIntel must be desperate - this looks like a part that they are releasing with hopes of eliciting "Wow 56-cores! Lets run right out and convert our entire data center to these!!!!!" from their customers - all the while hoping that those same customers do not pay attention to what is under the hood of these Xeons in comparison to Epyc.

Yeah... I'm not holding out much hope that Intel will be able to match Threadripper... just that they might be close enough to keep AMD from charging north of $4k+ for a CPU...

I'd love to be able to grab a 5990 (or 6990) without having to mortgage my house...

 

[zamroni111]

Intel must use tsmc euv for cpu instead of gpu.
They have to let go that in-house foundry pride

 

[Usukosej]

I think some modern data centers are moving towards just water cooling instead because of the density they're achieving now.

Still now the fact that 56/112 is what they're planning on when Epyc is *already* 64/128 at lower tdp and will go up to an insane 96/192 fairly soon clearly shows how much of a better proposition going AMD is turning out to be right now: Datacenters will have to consider a move to water cooling to have a reasonable set up with Xeons or yes, updating platforms but keeping their installation and deployment the same and just switching over to Epycs instead.

Total performance matters more than core count and stuff like AVX is important for many workloads

Intel sits hard at enterprise markets, both server and laptop segment FYI - I sell hardware b2b for a living.

 

[Usukosej]

Intel must use tsmc euv for cpu instead of gpu.
They have to let go that in-house foundry pride

And pay overprices like AMD? Nah, Alder Lake destroys Zen 3 in price/perf for a reason; Intel can make cheaper chips than AMD because they don't rely on a middleman. Without TSMC, AMD is nothing. They rely 100% on TSMC. GloFo was pure trash compared to TSMC, which is why Ryzen 1000 and 2000 series were mediocre at best.

Intel shipped millions and millions of chips during COVID lockdowns. Especially for laptops (enterprise demand was skyhigh because of working from home). Meanwhile AMD was waiting for the next TSMC shipment and their 7nm line was heavily overbooked + all AMD products were made on this proces.

Intel has best performance per dollar today, not AMD. Go read TechSpots Top 5 CPU article, and you will see.

This will be an even bigger problem for AMD; They will have to pay big to use TSMC 5nm and 3-4nm going forward, because Nvidia and Intel are customers too this time, and they both have way more money than AMD.

AMD _NEEDS_ TSMC to be able to compete, Intel and Nvidia don't really (Intel makes their own chips and Ampere beat RDNA2 using Samsung 8nm which is more like a 10nm node in reality).

Just by reserving TSMC capacity; AMDs potential output goes down and their prices go up. Supply vs demand; Simple.

Being fabless saved AMD when Intel was stuck, AMD being fabless will also be the reason why Intel will come back, because AMD can't match Intel in terms of chip price _IF_ Intel has the need to do aggressive pricing.

Intel will be more advanced than TSMC by 2025, they are building several fabs across the world as we speak. Top-end facilities. TSMC don't have the money to do the same. Which is why they are raising prices, they know Intel will be back in the lead soon. TSMC raised prices back in 2020/2021, and this plays a big role in AMDs Ryzen 5000 pricing (and no NON-X models); AMD simply had to pay alot more for Ryzen 5000 than Ryzen 3000 and did not need to go aggressive on pricing to win marketshare.

Result? Intel won the last 2 quarters in terms of CPU sales. It's a fact.

Intel 7 ~ 7nm TSMC
Intel 4 ~ 4nm TSMC (which is optimized 5nm process)
Intel A20 = 2nm
Intel A18 = 1.8nm

TSMC have always been very optimistic about their true density.

TSMC 7nm is more like Intel 10nm in terms of density. This is the reason why Intel changed the name to Intel 7.

No need to call it 10nm when TSMC calls their 10nm for 7nm.

Nanometer can't be used anyway to compare across different fabs.

Intel 14nm was more dense than GloFo 12nm for example.

Intel has a bright future ahead of them. Tons of money and tons of fabs incoming.

Google Intel factory and you will see that they are building several mega factories around the world as we speak. Some are even finished soon.

 

[HardReset]

Intel will be more advanced than TSMC by 2025, they are building several fabs across the world as we speak.

Whole post was just nonsense. And this is top of that BS.

Intel just released 10nm that was late whopping 3.5 years. But hey, "Intel will lead again in three years"

TSMC is around three times bigger manufacturer than Intel. Yeah, they have money. Yeah, they are building several fabs as we speak.

You really sound like Intel's new CEO. Lots of trash talk, nothing to support it.

 

[zamroni111]

And pay overprices like AMD? Nah, Alder Lake destroys Zen 3 in price/perf for a reason; Intel can make cheaper chips than AMD because they don't rely on a middleman. Without TSMC, AMD is nothing. They rely 100% on TSMC. GloFo was pure trash compared to TSMC, which is why Ryzen 1000 and 2000 series were mediocre at best.

Intel shipped millions and millions of chips during COVID lockdowns. Especially for laptops (enterprise demand was skyhigh because of working from home). Meanwhile AMD was waiting for the next TSMC shipment and their 7nm line was heavily overbooked + all AMD products were made on this proces.

Intel has best performance per dollar today, not AMD. Go read TechSpots Top 5 CPU article, and you will see.

This will be an even bigger problem for AMD; They will have to pay big to use TSMC 5nm and 3-4nm going forward, because Nvidia and Intel are customers too this time, and they both have way more money than AMD.

AMD _NEEDS_ TSMC to be able to compete, Intel and Nvidia don't really (Intel makes their own chips and Ampere beat RDNA2 using Samsung 8nm which is more like a 10nm node in reality).

Just by reserving TSMC capacity; AMDs potential output goes down and their prices go up. Supply vs demand; Simple.

Being fabless saved AMD when Intel was stuck, AMD being fabless will also be the reason why Intel will come back, because AMD can't match Intel in terms of chip price _IF_ Intel has the need to do aggressive pricing.

Intel will be more advanced than TSMC by 2025, they are building several fabs across the world as we speak. Top-end facilities. TSMC don't have the money to do the same. Which is why they are raising prices, they know Intel will be back in the lead soon. TSMC raised prices back in 2020/2021, and this plays a big role in AMDs Ryzen 5000 pricing (and no NON-X models); AMD simply had to pay alot more for Ryzen 5000 than Ryzen 3000 and did not need to go aggressive on pricing to win marketshare.

Result? Intel won the last 2 quarters in terms of CPU sales. It's a fact.

Intel 7 ~ 7nm TSMC
Intel 4 ~ 4nm TSMC (which is optimized 5nm process)
Intel A20 = 2nm
Intel A18 = 1.8nm

TSMC have always been very optimistic about their true density.

TSMC 7nm is more like Intel 10nm in terms of density. This is the reason why Intel changed the name to Intel 7.

No need to call it 10nm when TSMC calls their 10nm for 7nm.

Nanometer can't be used anyway to compare across different fabs.

Intel 14nm was more dense than GloFo 12nm for example.

Intel has a bright future ahead of them. Tons of money and tons of fabs incoming.

Google Intel factory and you will see that they are building several mega factories around the world as we speak. Some are even finished soon.

Tsmc euv is more expensive than Intel in-house foundry, but still gives amd hugh profit margin.
Additionally except in avx 512 test, latest amd epyc thread rips latest xeon, doesn't it?

And except in ray tracing, rdna2 beats ampere in power consumption and performance per transistor.
cdna2 has better fp64 performance than ampere ga100 which is what real compute use case need

 

[Usukosej]

Tsmc euv is more expensive than Intel in-house foundry, but still gives amd hugh profit margin.
Additionally except in avx 512 test, latest amd epyc thread rips latest xeon, doesn't it?

And except in ray tracing, rdna2 beats ampere in power consumption and performance per transistor.
cdna2 has better fp64 performance than ampere ga100 which is what real compute use case need

Hahaha performance per transistor? New term? AMD has terrible ray tracing which lacks tons of features compared to RTX

The high power usage is because of GDDR6X but GDDR6X is also the reason why 4K+ resolution gaming runs much better on GA102 based chips with 320+ bit bus. AMD uses sucky GDDR6 memory even on their flagship. RDNA2 have crappy 4K+ performance as a result; cheap memory and low bandwidth

AMD also gimps lower end RDNA2 with 64-128 bit bus and gimped PCIe speed hahah

AMD lost GPU marketshare for a reason.

Nvidia sits on 85% dGPU marketshare on Steam HW Survey.

Sad but true

 

[Usukosej]

Whole post was just nonsense. And this is top of that BS.

Intel just released 10nm that was late whopping 3.5 years. But hey, "Intel will lead again in three years"

TSMC is around three times bigger manufacturer than Intel. Yeah, they have money. Yeah, they are building several fabs as we speak.

You really sound like Intel's new CEO. Lots of trash talk, nothing to support it.

Haha TSMC does not have money compared to Intel. You know nothing.
Intel has twice the revenue if not triple.

Intel has 10 times the revenue of AMD too. AMD is a small player.

Intel Can Regain Process Tech Lead by 2025, Says IC Knowledge President

Intel's accelerated roadmap is already showing progress with the 18A process.

www.tomshardware.com

Yeah it must be hard being an AMD fanboy these days ... Sadly Intel is not stuck on 14nm anymore AMD needed 7nm TSMC to beat Intel 14nm

Intel is back and AMD already lowered prices across the board. This always happends. Intel always comes back and AMD will go back to being cheap.

 

[zamroni111]

Hahaha performance per transistor? New term? AMD has terrible ray tracing which lacks tons of features compared to RTX

The high power usage is because of GDDR6X but GDDR6X is also the reason why 4K+ resolution gaming runs much better on GA102 based chips with 320+ bit bus. AMD uses sucky GDDR6 memory even on their flagship. RDNA2 have crappy 4K+ performance as a result; cheap memory and low bandwidth

AMD also gimps lower end RDNA2 with 64-128 bit bus and gimped PCIe speed hahah

AMD lost GPU marketshare for a reason.

Nvidia sits on 85% dGPU marketshare on Steam HW Survey.

Sad but true

Performance per transistor matters for intel and amd because more transistors means more production cost, less yield and more power consumption competitiveness.

Surely amd prioritizes to use its limited tsmc allocation to produce the much higher margin cpu chiplet than gpu chip, especially if those cpu chiplets are sold as server cpu.
Profit per transistor: server cpu >>> consumer cpu >>> consumer gpu.

 

[Usukosej]

Performance per transistor matters for intel and amd because more transistors means more production cost, less yield and more power consumption competitiveness.

Surely amd prioritizes to use its limited tsmc allocation to produce the much higher margin cpu chiplet than gpu chip, especially if those cpu chiplets are sold as server cpu.
Profit per transistor: server cpu >>> consumer cpu >>> consumer gpu.

Wrong, you act like 100% of the chip in a Nvidia GPUs is used for rasterization performance like AMD ... AMDs RT performance sucks for a reason and DLSS beats FSR for a reason too; RT and Tensor Cores.

If Nvidia ignored the future, and went for full-on rasterization performance, they would wipe the floor with AMD in terms of power per transistor

RAY TRACING is the future for gaming and AMD better sort out their crappy RT perf. Even 6900XT is useless for RT at 1440p, it will barely does 1080p/60fps RT.

RTX = Class leading RT perf, best features, best performance. And this is why Nvidia dominated AMD this generation.

Nvidia beat AMD with ease using a worse node (Samsung 8nm, which is more like TSMC 10nm). AMD is years behind. Lets see if their MCM approach is gonna save them, and what issues that technology will bring. 1st gen of anything is rarely good...

 

[zamroni111]

Wrong, you act like 100% of the chip in a Nvidia GPUs is used for rasterization performance like AMD ... AMDs RT performance sucks for a reason and DLSS beats FSR for a reason too; RT and Tensor Cores.

If Nvidia ignored the future, and went for full-on rasterization performance, they would wipe the floor with AMD in terms of power per transistor

RAY TRACING is the future for gaming and AMD better sort out their crappy RT perf. Even 6900XT is useless for RT at 1440p, it will barely does 1080p/60fps RT.

RTX = Class leading RT perf, best features, best performance. And this is why Nvidia dominated AMD this generation.

Nvidia beat AMD with ease using a worse node (Samsung 8nm, which is more like TSMC 10nm). AMD is years behind. Lets see if their MCM approach is gonna save them, and what issues that technology will bring. 1st gen of anything is rarely good...

But those rt gaming is consumer use case in which nvidia sacrifices lots of profit per transistor opportunity.

At srp, $450 ryzen 5800x only contains less than 7 billion transistors.
Compared that to $320 GeForce 3060 which the gpu chip itself contains 13 billions transistors and still needs to include 8 GB VRAM, VRM, fans etc.

 

[Usukosej]

But those rt gaming is consumer use case in which nvidia sacrifices lots of profit per transistor opportunity.

At srp, $450 ryzen 5800x only contains less than 7 billion transistors.
Compared that to $320 GeForce 3060 which the gpu chip itself contains 13 billions transistors and still needs to include 8 GB VRAM, VRM, fans etc.

The 3D cache is what drives prices up on 5800X3D. It's just a 5800X without it.

 

[bviktor]

Clearly, Intel has gone the "throw watts at it" route.