Friday, December 15th 2023

Intel "Emerald Rapids" Xeon Platinum 8592+ Tested, Shows 20%+ Improvement over Sapphire Rapids

by
AleksandarK
 Discuss (23 Comments)
Yesterday, Intel unveiled its latest Xeon data center processors, codenamed Emerald Rapids, delivering the new Xeon Platinum 8592+ flagship SKU with 64 cores and 128 threads. Packed into its fresh silicon, Intel promises boosted performance and reduced power hunger. The comprehensive tech benchmarking website Phoronix essentially confirms Intel's pitch. Testing production servers running the new 8592+ showed solid gains over prior Intel models, let alone older generations still commonplace in data centers. On average, upgrading to the 8592+ increased single-socket server performance by around 23.5% compared to the previous generation configs of Sapphire Rapid, Xeon Platinum 8490H. The dual-socket configuration records a 17% boost in performance.

However, Intel is not in the data center market by itself. AMD's 64-core offering that Xeon Platinum 8592+ is competing with is AMD EPYC 9554. The Emerald Rapids chip is faster by about 2.3%. However, AMD's lineup doesn't stop at only 64 cores. AMD's Genoa and Genoa-X with 3D V-cache top out at 96 cores, while Bergamo goes up to 128 cores. On the power consumption front, the Xeon Platinum 8592+ was pulling about 289 Watts compared to the Xeon Platinum 8490H average of 306 Watts. At peak, the Xeon Platinum 8592+ CPU managed to hit 434 Watts compared to the Xeon Platinum 8490H peak of 469 Watts. This aligns with Intel's claims of enhanced efficiency. However, compared to the 64-core counterpart from AMD, the EPYC 9554 had an average power consumption of 227 Watts and a recorded peak of 369 Watts.
Source: Phoronix

Related News

Add your own comment

23 Comments on Intel "Emerald Rapids" Xeon Platinum 8592+ Tested, Shows 20%+ Improvement over Sapphire Rapids

#1
Onasi
I find it interesting how in a single socket configuration the 8592+ is faster than the 9554, even if marginally, but as soon as 2P config is in play they switch places and now EPYC is faster, and even seemingly by a larger amount. Can anyone more knowledgeable about server hardware shed some light on potential reason for this? Does Xeon lose more performance because it becomes more power limited? Worse connectivity between chips? Something else?
Posted on Reply
#2
Imouto
OnasiI find it interesting how in a single socket configuration the 8592+ is faster than the 9554, even if marginally, but as soon as 2P config is in play they switch places and now EPYC is faster, and even seemingly by a larger amount. Can anyone more knowledgeable about server hardware shed some light on potential reason for this? Does Xeon lose more performance because it becomes more power limited? Worse connectivity between chips? Something else?
I'd say AMD scored a home run with Infinity Fabric.
Posted on Reply
#3
HTC
OnasiI find it interesting how in a single socket configuration the 8592+ is faster than the 9554, even if marginally, but as soon as 2P config is in play they switch places and now EPYC is faster, and even seemingly by a larger amount. Can anyone more knowledgeable about server hardware shed some light on potential reason for this? Does Xeon lose more performance because it becomes more power limited? Worse connectivity between chips? Something else?
Better scaling?
Posted on Reply
#4
AnarchoPrimitiv
OnasiI find it interesting how in a single socket configuration the 8592+ is faster than the 9554, even if marginally, but as soon as 2P config is in play they switch places and now EPYC is faster, and even seemingly by a larger amount. Can anyone more knowledgeable about server hardware shed some light on potential reason for this? Does Xeon lose more performance because it becomes more power limited? Worse connectivity between chips? Something else?
Check out the original review, Phoronix usually goes in depth
Posted on Reply
#5
Daven
Intel’s 56 core skylake dual die from 2019:


Eight additional cores in four years using similar packaging. Biggest difference: Intel charged about $40,000 approx back then. Today $11,600. Isn’t competition awesome!
Posted on Reply
#7
Toothless
Tech, Games, and TPU!
KarymidoNThat power draw tho... jeez
Lots of cores, lots of power.
Posted on Reply
#8
Vya Domus
This isn't enough, they really need to pick up the pace. Making better products in the consumer market is one thing and arguably not that important but the fact they let AMD steamroll them in the server space really is embarrassing.
Posted on Reply
#9
bug
Vya DomusThis isn't enough, they really need to pick up the pace. Making better products in the consumer market is one thing and arguably not that important but the fact they let AMD steamroll them in the server space really is embarrassing.
Intel already picked up the pace. Granite Rapids is like 6 months away, that should bring proper upgrades to the table. Emerald Rapids is just a drop-in replacement for the current Sapphire Rapids setups.
Posted on Reply
#10
Squared
Sapphire Rapids was launched less than a year ago, so Emerald Rapids is also part of the quick pace. Then again, their microarchitectures, Golden Cove and Raptor Cove respectively, both came to the consumer market before even Sapphire Rapids was released, so a little over a year between the consumer and server parts. If Granite Rapids does come to market in 6 months, it'll be just six months behind the same tech (Redwood Cove) coming to consumers, which is more similar to the pacing of Epyc after Ryzen.
Posted on Reply
#11
KarymidoN
ToothlessLots of cores, lots of power.
EPYC 9754 has double the cores with similar (yet lower) power draw.
Posted on Reply
#12
bug
KarymidoNEPYC 9754 has double the cores with similar (yet lower) power draw.
That's how it works: less powerful cores, less power drawn. Put together, Epyc does more work per watt than these. But next year Intel will release their all E-core parts, that might change the landscape quite a bit.
Posted on Reply
#13
ncrs
bugThat's how it works: less powerful cores, less power drawn. Put together, Epyc does more work per watt than these. But next year Intel will release their all E-core parts, that might change the landscape quite a bit.
And those E-core parts will not have AMX that is capable of yielding huge performance increases in Sapphire/Emerald Rapids vs. EPYC. Nor will they have AVX-512, which is not the case for AMD's compact Zen 4c/5c cores.
They are also not capable of SMT, but will be available in (currently known) up to 288-core configurations. Depending on workloads it might not be enough to challenge AMD's current 256-thread SKUs.
They will be an interesting offering, especially for light cloud workloads, but it's still an Atom lineage core design.
Posted on Reply
#14
thesmokingman
OnasiI find it interesting how in a single socket configuration the 8592+ is faster than the 9554, even if marginally, but as soon as 2P config is in play they switch places and now EPYC is faster, and even seemingly by a larger amount. Can anyone more knowledgeable about server hardware shed some light on potential reason for this? Does Xeon lose more performance because it becomes more power limited? Worse connectivity between chips? Something else?
Never trust an Intel slide, Semis 101.
Posted on Reply
#15
Assimilator
The problem is that it needs to be about 200% faster to actually compete.
bugGranite Rapids is like 6 months away
... says Intel, the same company that has been unable to hit any of their self-announced deadlines for the last half decade.
Posted on Reply
#16
Redwoodz
thesmokingmanNever trust an Intel slide, Semis 101.
:laugh: sig worthy:toast:


They can release 5 new architectures every year, it's not going to change their fab process anytime soon. When all their cpu's are manufactured elsewhere, then we will see them return.
Posted on Reply
#17
R0H1T
Assimilatorthe same company that has been unable to hit any of their self-announced deadlines for the last half decade.
Close to a decade actually ~ 22nm was the first node that was late by about a quarter or two IIRC.
Posted on Reply
#18
Minus Infinity
bugIntel already picked up the pace. Granite Rapids is like 6 months away, that should bring proper upgrades to the table. Emerald Rapids is just a drop-in replacement for the current Sapphire Rapids setups.
And Granite Rapids will be up against Epyc Turin around the same time assuming Intel delivers, which is a big IF. AMD usually delivers on time as they don't give timetables up to 3 years in advance.
Posted on Reply
#19
R0H1T
I doubt GR is coming in 6 months, maybe a paper launch at best?
Posted on Reply
#21
DavidC1
@Daven It's not the same packaging technology. Cascade Lake-AP just put two dies in the same package. Emerald Rapids connects them using EMIB, which is much lower power per bit transferred.
bugIntel already picked up the pace. Granite Rapids is like 6 months away, that should bring proper upgrades to the table. Emerald Rapids is just a drop-in replacement for the current Sapphire Rapids setups.
Granite Rapids is little more than 6 months away. Sierra Forest is probably close to 6 months and Intel said GNR is a "fast followup" which could make it 2-3 months after potentially making it August if SRF is June.
ncrsAnd those E-core parts will not have AMX that is capable of yielding huge performance increases in Sapphire/Emerald Rapids vs. EPYC. Nor will they have AVX-512, which is not the case for AMD's compact Zen 4c/5c cores.
The Gracemont core that's the basis for Sierra Forest is very competitive in Integer performance which is the basis of uarch and in a cloud workload is really all that matters. Golden Cove is only 25% faster per clock here. SMT also matters lot less in such workloads which is why ARM servers are starting to pick up.
Posted on Reply
#22
ncrs
DavidC1The Gracemont core that's the basis for Sierra Forest is very competitive in Integer performance which is the basis of uarch and in a cloud workload is really all that matters. Golden Cove is only 25% faster per clock here. SMT also matters lot less in such workloads which is why ARM servers are starting to pick up.
Integer workloads are one of the best when it comes to SMT scaling. My 8-core 7800X3D is getting 1583% scaling in 7-zip decompression benchmark due to SMT utilization.
The lack of SMT for Sierra Forest is probably related to the vulnerabilities, which still are a bit iffy when it comes to SMT. The current guidelines for maximum security under Linux is still to disable SMT altogether, regardless of vendor.
Posted on Reply
#23
Assimilator
DavidC1SMT also matters lot less in such workloads which is why ARM servers are starting to pick up.
Arm servers are picking up because they're cheap. Literally no other reason.
Posted on Reply
Add your own comment
Nov 21st, 2024 12:06 EST change timezone

Latest GPU Drivers

New Forum Posts

Popular Reviews

Controversial News Posts