Epyc deserves its name – Review Roundup

Resurgent AMD

AMD has long been the underdog in markets it competes in, and its fans have been waiting for it to unequivocally retake the performance crown from Intel. It’s held it before with the launch of the Athlon processor way back in my high school days. AMD has often been competitive through a price/performance lens, particularly for graphics cards, but their products often come with a caveat: power usage is higher than the competition, lower performance under specific workloads, etc. Well, it appears the AMD fans finally got what they wanted. Introducing the new EPYC 2, code named “Rome” server platform.

Back in the summer of 2017 AMD announced its original EPYC Server processors. Built around the original Zen cores, AMD brought some pressure on Intel due to competitive performance and lower costs in an extremely profitable market that Intel has long dominated. The new EPYC 2 processors look to continue that pressure.

The Meat

Introducing almost 20 new models, AMD has covered the bases from the EPYC 7251 8 core/16 thread CPU all the way up to an EPYC 7742 64 core/128 thread CPU rocking 256MB of L3 cache, 128 lanes of PCIe 4.0 (256 GB/s in both directions), and an 8 channel DDR4 memory setup. Built on an evolution of AMD’s chiplet process, these have a separate I/O chip, which AMD says has decreased latency. These are also the first x86 chips built on 7nm technology, though not the entire processor is 7nm, as can be seen below. AMD makes some strong claims about this new 7nm process too, claiming double the core density. Apparently these new bad-boys get 4x the Floating-Point of the first generation EPYC chips as well. There is also a new security chip which is used to ” help defend against certain side-channel attacks, encrypt main memory and virtual machine memory, and cryptographically help secure the boot process.” I look forward to those in the security industry testing out the effectiveness of this new approach.

Claiming potentially 90% higher integer performance and up to 79% better floating point performance than a Xeon Platinum 8280 and a #1 position in benchmarks, AMD is planting some clear flags here. AMD says these new chips do all their magic using as little as half the power consumption of the previous generation EPYC chips. Not content to be simply a discount supplier, it’s clearly pushing to significantly shake up the market entirely, and since Intel gets almost half of its revenue from servers, it’s not surprising AMD covets its position.

AMD has not artificially segmented these products off either. Every Rome cpu will support the same 128 PCIe 4.0 lanes, the same infinity fabric speeds, up to 4 TB per socket, and identical security and virtualization features. Anandtech points out that the 7502 is a particularly excellent deal. With 32/64 cores, a frequency max of 3.35GHz, 128MB cache, and a 200w TDP, the $2600 price of the EPYC 7502 seems like a steal, especially when compared to the Xeon Platinum 8253, a 16 core 3.0GHz competitor which rings in at $3115. The pricing bloodbath continues across the board, with Intel charging almost twice as much for roughly half as many cores as competing AMD chips. Never mind that AMD supports 4TB for free and Intel has zero support for PCIe 4.0 currently.

 

Get to the reviews!

Anandtech concludes that “AMD offers you up to 50 to 100% higher performance while offering a 40% lower price” and it states that in their testing AMD has beaten Intel by such a huge margin there is potentially “no contest.” It finishes with a recommendation and call AMD’s new technology “stellar.”

Hexus.net concludes that AMD has set a new performance record and that Intel is “far in the rear view mirror for heavily threaded applications.” It also doubts that Intel will be able to respond in the near future, especially once total cost is factored in. The writer does acknowledge the server market is a complicated one and Intel has been nurturing relationships for a long time. That being said, Hexus has said that the industry support for EPYC 2 vastly outstrips the launch of EPYC 1. How this plays out remains to be seen, but it’s a solid win for AMD. Hexus gives the new platform their stamp of approval.

Phoronix’s Michael Larabel did what he does best and published Linux benchmarks. He calls the EPYC 7742 “nothing short of remarkable.” There are some excellent graphs here which show AMD winning literally every single performance-per-dollar benchmark he ran, and often by significant margins. AMD also takes the top 3 spots for the geometric mean of all tests. He concludes that Rome’s performance-per-watt is either comparable or vastly better than Intel’s. It’s a clear victory for AMD over at Phoronix, and he throws out the adjectives “phenomenal” and “fantastic.” Phoronix will be doing addition testing, but Michael seems quite impressed so far.

There does seem to be a consensus around these new chips and it’s that they’re purportedly excellent. What’s perhaps most striking to this writer is the 225w TDP on the 64 core 7742 running up to 3.4GHz. That’s a heck of a lot of computing performance, and when I remember back to my Pentium D 830 with it’s 130w TDP I am blown away at the progress made. Intel has a huge advantage in market share and relationships, but they are absolutely feeling some heat here. When AMD ran a public demonstration this summer of EPYC 2 Intel was quick to redo the test and complain about AMD’s methods (though Intel’s own results still showed it losing, just by less), and it’ll be interesting to see whether they make any public statements about these new chips or choose to just hunker down and come up with an improved product to answer AMD’s. Regardless of future products, right now it appears that for a great many companies AMD has become a heck of a lot more compelling. Apparently the glue works.

avatar
13 Comment threads
25 Thread replies
0 Followers
 
Most reacted comment
Hottest comment thread
24 Comment authors
XaeusSpunjjiRISC5jarderJ_D_ Recent comment authors
  Subscribe  
newest oldest most voted
Notify of
Xaeus
Guest

Why most still insist is calling it the x86 architecture ?
All x64 current architecture from both AMD and Intel is actually AMD64.
I guess intel’s scam marketing is still strong enough.

Spunjji
Guest
Spunjji

That’s a really nice summary – something for everyone in there, from the old hands to new readers. Great work 🙂

J_D_
Guest
J_D_

About the glue.

10+ years ago AMD was laughing about Intel’s glue. Intel’s glue worked great back then. Today Intel is laughing at AMD’s glue. AMD’s glue works probably even better than Intel’s.

Also from another perspective, in the last 15 years we have AMD copying one mistake from Intel and suffering (Pentium 4 >>> AMD FX) and one correct choice from Intel, the glue and manages to triumph.

JustAnEngineer
Guest
JustAnEngineer

Thanks for the round-up, Renee. I’m not personally in the market for a new 256-thread 2P server at the moment, but it’s exciting to see AMD push progress along by offering better performance at a lower price.

barich
Guest
barich

I’ve been quite negative about the site redesign and some recent articles, but this is a very nice writeup. Keep it up!

Teppo
Guest
Teppo

Last Intel product I bought was Celeron 400. Not going to get back Intel.

chuckula
Guest
chuckula

So here’s a glowing article from Phoronix: https://www.phoronix.com/scan.php?page=news_item&px=AMD-EPYC-7742-2P-Summary

But here’s one little gem: “With the dozens of tests run for this comparison, the EPYC 7742 2P has 22% better performance than the dual Platinum 8280 server based on the geometric mean”

That’s 128 cores of Zen 2 magic vs. 56 cores of Cascade Lake.

We’ll see if it’s physically possible to get 22% more performance [at a minimum] when it’s 112 cores of Cooper Lake or… I dunno maybe 104 Cores of Icelake rather soon.

Sweatshopking
Guest
Sweatshopking

ARE YOU SAYING THAT A FUTURE PRODUCT MIGHT BE FASTER THAN A CURRENT PRODUCT???

AMAZING!!

chuckula
Guest
chuckula

OF COURSE IT’S AMAZING! WHICH IS WHY WE ARE ALL CALLING AN EPYC FROM 2019 A MIRACLE COMPARED TO SERVER PARTS THAT ARE JUST SLIGHTLY TWEAKED FROM THE ONES IN 2017 AND JUMPING UP AND DOWN!

Sweatshopking
Guest
Sweatshopking

What’s ridiculous is that they’re still selling server parts from 2017.

just brew it!
Guest
just brew it!

A pair of 8280s will set you back around 20 grand. The take-home here isn’t just that the 7742 bested the 8280, it’s significantly cheaper too!

chuckula
Guest
chuckula

The price of server-grade Xeons is not dictated by ARK.

Waco
Guest
Waco

No, but it’s also not anywhere near price-competitive to Rome at the moment even with the standard discounts you can get from MSRP.

Klimax
Guest

OEMs often pay far lower prices then is list/ARK price. Intel will simply sacrifice some of margins until 10nm/7nm nodes are fully online.

And memory latency is very high (especially after 16MB mark), so quite few major sue cases like databases will suffer on AMD servers.
(See Anandtech’s review – no databases but memory access is explored)

psuedonymous
Guest
psuedonymous

Price competitive purely per CPU. When you start adding up all the other costs (like per-core pricing of software which often dominates total cost, or replacing your management system which has assumed universal presence of AMT) then margins start to narrow.

Regardless, almost anyone actually *buying* these chips are mostly ignoring any consumer benchmarking sites opinion of price/perf, and instead running their actual workloads on sample systems to get real world data.

just brew it!
Guest
just brew it!

Sure, per-core licensed software is going to be a pain. OTOH, for stuff that is licensed per-socket EPYC is going to be a huge win.

Gondalf
Guest
Gondalf

But yes 🙂 upcoming Cooper Lake will be clearly faster and think about a 2X26 and a 3X26 cores Ice Lake MCMs, moarrr cores is always better :)). And why not to a 4X26 Ice lake??? Intel can easily glue four 350mm2 dies on the same package. And remember that in 2020 Intel will ship stacked SKUs, with two dies stacked via TSV. Anyway there is too much hipe on the core number, only few customers can take advantage from an high core number, expecially because the single thread performance suffer too much. All this hipe on Epyc is ridicule… Read more »

jarder
Guest
jarder

That 22% is a reflection of Amdahls law and Phoronix’s testing methods. Do you really thing PHPBench will scale well to 128 cores? Even highly threaded applications like Blender quickly lose the ability to scale linearly past a dozen or so cores, never mind 128. Let me turn the question around, do you think 104 cores of Icelake CPU will perform 100% better than cascade lake based on the benchmarks chosen? Running a single application at a time is definitely not what 64 core chips were designed for, their target market is virtualization. Replacing eight 16-core servers with a single… Read more »

Mr Bill
Guest
Mr Bill

Read about it over coffee this morning on Anandtech. It sure sounds like a winner.
Waiting on Chuckula’s comment….

chuckula
Guest
chuckula

WE’VE INVENTED A TIME MACHINE TO GO BACK AND CANCEL SANDY BRIDGE!

RISC5
Guest
RISC5

Oh no you didn’t! I still running 2600K it’s ring bus runs circles around that IF trash what 96GB/sec per core vs IF @ 25GB/sec time; So wheres my data, oh ya on the bus mine already went the shortest path. When either of them can off double IPC & Clocks call my wallet – never mind I will be dead by then.

Klimax
Guest

It got some nice wins, but there were some workloads missing. (Like databases) Also any workload that is bound by memory latency might suffer on Rome.

As for Intel? They will likely delete portion of margins until 10/7nm nodes are online. (Not first nor last time)

Although I wonder what margins are on Rome chips.

Sweatshopking
Editor

AMD claims significant latency improvements with these new badboys. I’m looking forward to seeing more testing.

RISC5
Guest
RISC5

2 NUMA in 2S. is acceptable to address 8TB.

Yan
Guest
Yan

Typo right in the title: “Epyc deserves *its* name”.

cygnus1
Guest
cygnus1

By far, one of the better reviews out there for the new Epyc is over at Serve The Home. Way more detail than AT or the others, and more relevant insight since they do more server related coverage normally anyway.

https://www.servethehome.com/amd-epyc-7002-series-rome-delivers-a-knockout

cygnus1
Guest
cygnus1

also, nice write up Renee.

Neutronbeam
Guest

AND she used “purportedly”!

jihadjoe
Guest
jihadjoe

How to make gerbils accept the site redesign:

☑ Use ‘purportedly’ in as many articles as possible

cygnus1
Guest
cygnus1

another also. blows my mind that Intel has lost the lead in platform capability, micro architecture, and most shocking to me, manufacturing node. The manufacturing thing is the craziest to me because of them just historically being 18+ months ahead of every other foundry company on the planet and now they’re behind. I’m sure they’ll come back with a vengeance in a year or two, but gotta love watching the mighty go down every now and then.

GG AMD

Klimax
Guest

In a way Intel got too cocky and thought they can go 150% then expected. As for microarchitecture it was in a way bound to happen, because general code is nasty thing and there is only so much one can do in only so many ways. (Aka convergence of designs)

Take a look at Ice Lake. Major changes? Mostly structures fueling OOO core are larger and L1 got size bump too. (One of the biggest blocks in modern CPU because of SRAM needing many transistors for a single bit)

usacomp2k3 (AJ)
Guest
usacomp2k3 (AJ)

Agreed. I saw theirs first last night and wow it looks impressive. That certainly has expanded their horizons since the WHS days.

MOSFET
Guest
MOSFET

I was somewhat shocked STH wasn’t first on TR’s list.

Waco
Guest
Waco

Did Renee pick up on the glue jokes or did a TR regular write this article?

Guest
ludi

Good summary! Hoping to see more news posts like this in the future.

usacomp2k3 (AJ)
Guest
usacomp2k3 (AJ)

AMD does seem to have knocked this one out of the park. Thanks for the mini-review roundup.

Pin It on Pinterest

Share This