Intel Prepares 500-Watt Xeon 6 SKUs of Granite Rapids and Sierra Forest

[AleksandarK]

Intel is preparing to unveil its cutting-edge Xeon 6 series server CPUs, known as Granite Rapids and Sierra Forest. These forthcoming processors are set to deliver a significant boost in performance, foreshadowing a new era of computing power, albeit with a trade-off in increased power consumption. Two days ago, Yuuki_Ans posted information about the Beechnut City validation platform. Today, he updated the X thread with more information that Intel is significantly boosting core counts across its new Xeon 6 lineup. The flagship Xeon 6 6980P is a behemoth, packing 128 cores with a blistering 500 Watt Thermal Design Power (TDP) rating. In fact, Intel is equipping five of its Xeon 6 CPUs with a sky-high 500 W TDP, including the top four Granite Rapids parts and even the flagship Sierra Forest SKU, which is composed entirely of efficiency cores. This marks a substantial increase from Intel's previous Xeon Scalable processors, which maxed out at 350-385 Watts.

The trade-off for this performance boost is a dramatic rise in power consumption. By nearly doubling the TDP ceiling, Intel can double the core count from 64 to 128 cores on its Granite Rapids CPUs, vastly improving its multi-core capabilities. However, this focus on raw performance over power efficiency means server manufacturers must redesign their cooling solutions to accommodate Intel's flagship 500 W parts adequately. Failure to do so could lead to potential thermal throttling issues. Intel's next-gen Xeon CPU architectures are shaping up to be one of the most considerable generational leaps in recent memory. Still, they come with a trade-off in power consumption that vendors and data centers will need to address. Densely packing thousands of these 500-Watt SKUs will lead to new power and thermal challenges, and we wait to see future data center projects utilizing them.



View at TechPowerUp Main Site | Source

 

[xtreemchaos]

and i thought my Xeon "2699" 145w TDP was high, i feel so much better now . yea fancy water cooler needed.

 

[FoulOnWhite]

crazy. Imagiine the juice that's gonna be wasted powering hundreds of these greedy mo's

 

[AusWolf]

500 W? Is that how we measure performance goals these days instead of in Gigahertz or some other unit that actually measures performance?

2014: "Intel's new 4 GHz CPU comes with better efficiency."

2024: "Intel's new 500 W CPU comes with more cores."

What a way to sell literal shit!

 

[Tropick]

...good lord, wasn't Sierra Forest supposed to be targeted at efficiency or something? What with it being made entirely out of E-cores?

From Anandtech:

Industry: "So by using E-cores we can expect a significant amount of power savings while retaining high density per rack unit?"

Intel:

 

[overclockedamd]

I am sure these will be priced at a reasonable amount.

On a side note with nvidia and intel just generationally ramping up power usage to get better numbers are we going to need 2k watt psu's in two generations?

 

[usiname]

Those xeons will use the same cores as Meteor Lake, so they are DOA

 

[AnarchoPrimitiv]

If I'm not mistaken, aren't the 128 core chips apart of the AP line and requires a socket almost twice as large as the "normal" upcoming xeon socket? In the past, like for Cascade lake if I remember correctly, this AP line has had no documented MSRP, no provided performance figures, no third party reviews or benchmarks, and for all intents and purposes is non-existent unless you're a Fortune 500 company. I mean that literally, the last time Intel announced this AP line they refused to provide any information like price or performance.

 

[wNotyarD]

Intel needs to get Terry Crews yelling POWER for a commercial

 

[Nioktefe]

...good lord, wasn't Sierra Forest supposed to be targeted at efficiency or something? What with it being made entirely out of E-cores?

From Anandtech:
View attachment 344130

Industry: "So by using E-cores we can expect a significant amount of power savings while retaining high density per rack unit?"

Intel:
View attachment 344131

Efficiency is a ratio between power and compute

288 cores at 500W is less than 2W per core, 128 at 500 is 4W per core, it's not like it's worse than what they do on desktop with proper power restriction

If you mix the numbers for e-core and p-core it would be the same configuration as a 13900k at 65W, and by the efficiency test that have been done it's close to peak efficiency for their architecture (and it's also very close to what amd can achieve).

I really don't know from what point of view do you think those are not efficient, I'm sure amd will be better but it's still the best intel can provide.

Having bigger socket with more silicon also means less servers for a given performance target, which cuts the number of servers, switchs and others non-compute related electronics.
All of that means more efficiency, it will also mean big scary number per socket.

You can expect next generation to be even more dense due to next gen packaging enabling more and more silicon per socket, we will reach 700W, then 1kw per socket in the coming years.

 

[Vya Domus]

and by the efficiency test that have been done it's close to peak efficiency for their architecture (and it's also very close to what amd can achieve).

I really don't know from what point of view do you think those are not efficient

The problem is getting better efficiency from multiple slower cores is not equivalent to the same efficiency obtained from fewer, faster cores. It's not a surprise that smaller, lower clocked cores are more efficient, there is a reason neither AMD or Intel has done this until know and the reason is this limits the range of applications these can be used for.

 

[Gmr_Chick]

Intel needs to get Terry Crews yelling POWER for a commercial

I got you, homie!

You're welcome

 

[RandallFlagg]

The problem is getting better efficiency from multiple slower cores is not equivalent to the same efficiency obtained from fewer, faster cores. It's not a surprise that smaller, lower clocked cores are more efficient, there is a reason neither AMD or Intel has done this until know and the reason is this limits the range of applications these can be used for.

They haven't done it before because they didn't have the high density fabs (Intel) or allocation on the 3rd party fabs they have access to (AMD).

AMD has taken to stripping cache off the Zen 4 / 5 CPUs, calling them 4c and 5c. That's a bit of a hack. Intel's solution is likely going to be much more effective in the long run.

As far as the use case - high core count for front end and low tier VMs is one of the biggest use cases in a data center. Outside of edge compute and some specific serialized workflows, data centers are mostly serving up jvms and web pages. These take a lot of cores if traffic is high, but not a lot of individual core performance is needed as in most cases they are merely shuffling data back and forth. You throw your low performance high core count chips at those front end use cases, and high performance lower core counts at the back end (database servers).

 

[Minus Infinity]

Efficiency is a ratio between power and compute

288 cores at 500W is less than 2W per core, 128 at 500 is 4W per core, it's not like it's worse than what they do on desktop with proper power restriction

If you mix the numbers for e-core and p-core it would be the same configuration as a 13900k at 65W, and by the efficiency test that have been done it's close to peak efficiency for their architecture (and it's also very close to what amd can achieve).

I really don't know from what point of view do you think those are not efficient, I'm sure amd will be better but it's still the best intel can provide.

Having bigger socket with more silicon also means less servers for a given performance target, which cuts the number of servers, switchs and others non-compute related electronics.
All of that means more efficiency, it will also mean big scary number per socket.

You can expect next generation to be even more dense due to next gen packaging enabling more and more silicon per socket, we will reach 700W, then 1kw per socket in the coming years.

What all core frequency are we getting for 128 P cores to ONLY be at 500W? I mean it's infinitely better than cRaptor Lake's energy efficiency (out of the box) and I doubt Epyc Turin with 128 Zen 5 cores will much less either.

 

[Dr. Dro]

and i thought my Xeon "2699" 145w TDP was high, i feel so much better now . yea fancy water cooler needed.

I mean, if we're talking performance per watt, you'd be surprised how miserable both the 2699 v3 and v4 would look next to these.

I'd bet a single CPU of these Granite Rapids Xeon 6's can outperform a whole rack with like 4 blades of S8S 22-core Broadwell-EX servers in some workloads.

What all core frequency are we getting for 128 P cores to ONLY be at 500W? I mean it's infinitely better than cRaptor Lake's energy efficiency (out of the box) and I doubt Epyc Turin with 128 Zen 5 cores will much less either.

Considered the 10 nm lithography and extreme clock speeds pushed on i9 SKUs, I'd argue we can't even call Raptor Cove power inefficient. It's just that the TSMC node that Ryzen uses is like three generations ahead by now. These will certainly be energy optimized to reach that wattage target, so it'd be interesting to see metrics against a similar Epyc system.

 

[RandallFlagg]

What all core frequency are we getting for 128 P cores to ONLY be at 500W? I mean it's infinitely better than cRaptor Lake's energy efficiency (out of the box) and I doubt Epyc Turin with 128 Zen 5 cores will much less either.

Current Epyc 9754 128 core is 400W. It uses Zen 4c cores. I would imagine Zen 5c will use more power.

 

[AnotherReader]

Current Epyc 9754 128 core is 400W. It uses Zen 4c cores. I would imagine Zen 5c will use more power.

The default TDP for the EPYC 9754 is 360 W. It can be configured from 320 to 400 W.

I mean, if we're talking performance per watt, you'd be surprised how miserable both the 2699 v3 and v4 would look next to these.

I'd bet a single CPU of these Granite Rapids Xeon 6's can outperform a whole rack with like 4 blades of S8S 22-core Broadwell-EX servers in some workloads.



Considered the 10 nm lithography and extreme clock speeds pushed on i9 SKUs, I'd argue we can't even call Raptor Cove power inefficient. It's just that the TSMC !0 nnode that Ryzen uses is like three generations ahead by now. These will certainly be energy optimized to reach that wattage target, so it'd be interesting to see metrics against a similar Epyc system.

Intel 7 is slightly denser than TSMC N7; electrical characteristics are difficult to judge as there is no product that uses both nodes. So AMD is about a node ahead. As far as Sierra Forest is concerned, it should be using Intel 4 which is supposed to be denser than the node used for Zen 4; this is why this TDP of 500 W is so underwhelming.

 

[RandallFlagg]

The default TDP for the EPYC 9754 is 360 W. It can be configured from 320 to 400 W.

We don't know the default TDP of GR or SF, we just know the max of 500W.

This TPU write-up is pretty sensationalistic, but short on basic math skills, so you should give it the cynicism it deserves.

From the article:
"By nearly doubling the TDP ceiling, Intel can double the core count from 64 to 128 cores.."

Yeah but no, going from 385W to 500W is a 30% increase. It's nowhere near double, except in core count.

It also has 1/3 more memory channels (16 vs 12) rated for DDR5-6400 vs DDR5-4800 on Bergamo. That alone could account for all of the TDP expansion.

 

[AnotherReader]

We don't know the default TDP of GR or SF, we just know the max of 500W.

This TPU write-up is pretty sensationalistic, but short on basic math skills, so you should give it the cynicism it deserves.

From the article:
"By nearly doubling the TDP ceiling, Intel can double the core count from 64 to 128 cores.."

Yeah but no, going from 385W to 500W is a 30% increase. It's nowhere near double, except in core count.

It also has 1/3 more memory channels (16 vs 12) rated for DDR5-6400 vs DDR5-4800 on Bergamo. That alone could account for all of the TDP expansion.

Four more channels would impact system power consumption, but the effect on SOC power consumption shouldn't be 140 W. I hope you're right and 500 W is like 400 W for Bergamo: the upper limit in a configurable range.

I treat all leaks and rumours with skepticism. We can only judge a product when it launches and is reviewed by competent reviewers.

 

[RandallFlagg]

Four more channels would impact system power consumption, but the effect on SOC power consumption shouldn't be 140 W. I hope you're right and 500 W is like 400 W for Bergamo: the upper limit in a configurable range.

I treat all leaks and rumours with skepticism. We can only judge a product when it launches and is reviewed by competent reviewers.

It actually does affect the CPU. Its memory interface has to clock up to support the higher speed memory. So what you've got on Intel vs AMD, is a memory interface that runs 33% faster and has 33% more channels - which should result in 77% more bandwidth + lower latencies - with a TDP of 100W more.

I also suspect this is specifically for the GR part. I'd be surprised if the SF part is going to draw as much power as GR, so that generalized statement of 500W TDP across two wildly different chips for very different purposes is very suspect. It might be that they both have a 500W max, or maybe the socket has that max, but I can't imagine both chips having the same power profile - that would be absurd.

And here is probably the 500W profile.

288 e-cores, 16 channel DDR5-8800 JEDEC.

or

128 P-cores

AMD better get on it.

Intel Granite Rapids "Up To 128 P-Core" & Sierra Forest "Up To 288 E-Core" CPU Specs Leak, Massive LGA 7592 & LGA 4710 Motherboards Pictured Too

Intel's Granite Rapids P-Core & Sierra Forest E-Core CPU specs have leaked out along with huge LGA 7592 & LGA 4710 motherboards.

wccftech.com

 

[AnotherReader]

It actually does affect the CPU. Its memory interface has to clock up to support the higher speed memory. So what you've got on Intel vs AMD, is a memory interface that runs 33% faster and has 33% more channels - which should result in 77% more bandwidth + lower latencies - with a TDP of 100W more.

I also suspect this is specifically for the GR part. I'd be surprised if the SF part is going to draw as much power as GR, so that generalized statement of 500W TDP across two wildly different chips for very different purposes is very suspect. It might be that they both have a 500W max, or maybe the socket has that max, but I can't imagine both chips having the same power profile - that would be absurd.

And here is probably the 500W profile.

288 e-cores, 16 channel DDR5-8800 JEDEC.

or

128 P-cores

AMD better get on it.

View attachment 344339

Intel Granite Rapids "Up To 128 P-Core" & Sierra Forest "Up To 288 E-Core" CPU Specs Leak, Massive LGA 7592 & LGA 4710 Motherboards Pictured Too

Intel's Granite Rapids P-Core & Sierra Forest E-Core CPU specs have leaked out along with huge LGA 7592 & LGA 4710 motherboards.

wccftech.com

288 E cores for 500 W looks a lot better.

 

[mrdumbasspremium]

crazy. Imagiine the juice that's gonna be wasted powering hundreds of these greedy mo's

You and many others complain about peak power usage rather than power usage to complete the task. Most XEON's coming out now are simply more efficient despite having higher peak power draw over the previous generations. While some may think it's a problem it's not at all, It's simply a matter of designing the architecture around specific voltages on top of the processing node being used with die sizes increasing.

 

[SOAREVERSOR]

500 W? Is that how we measure performance goals these days instead of in Gigahertz or some other unit that actually measures performance?

2014: "Intel's new 4 GHz CPU comes with better efficiency."

2024: "Intel's new 500 W CPU comes with more cores."

What a way to sell literal shit!

It's not shit you are looking at this all wrong.

These are for enterprise where it's about how much performance you can smash into a space. What once took rooms and rooms or acres and acres of equipment can now be smashed into a single rack.

When your customers are Amazon, Microsoft, facebook, Oracle, and all the others who do real work not idiotic video games these things make sense. If you go around the DC area there are buildings of insane size for miles and miles with their own power plants. It doesn't matter if the power budget of a single CPU doubles if you can pack the performance of an entire row of servers into one rack. You're still ahead at the end of the day.

 

[unwind-protect]

Professional computing sometimes needs high core speed, too. For example, compiling large software packages might be very parallel as far as the compiler itself is concerned. But linking usually happens single thread.

 

[efikkan]

and i thought my Xeon "2699" 145w TDP was high, i feel so much better now . yea fancy water cooler needed.

There is no need for water cooling. On LGA 4677 the specialized NH-U14S can cool about ~680W sustained, which is a lot more than the "mainstream socket" version of the same cooler, which tops out about ~225-250W depending on the die.

What all core frequency are we getting for 128 P cores to ONLY be at 500W?

Probably very low.

Professional computing sometimes needs high core speed, too. For example, compiling large software packages might be very parallel as far as the compiler itself is concerned. But linking usually happens single thread.

These are the server versions of the Xeons. The workstation counterparts ("HEDT") will probably have a lot fewer cores and higher clock speeds.

And it's true as you say, but even if a company is building a lot of large software for distribution etc., that probably happens on a build server, so a few seconds there doesn't really matter. What probably matters a lot more is smaller (re-)compilations on the workstation, which a developer may do dozens of times a day or more, and in this case, higher core performance is a clear benefit.