AMD Looks To 'Crush' Intel's Xeon With New Epyc Server Chips (extremetech.com)

AMD has unveiled the first generation of Epyc, its new range of server processors built around its Zen architecture. Processors will range from the Epyc 7251 -- an eight-core, 16-thread chip running at 2.1 to 2.9GHz in a 120W power envelope -- up to the Epyc 7601: a 32-core, 64-thread monster running at 2.2 to 3.2GHz, with a 180W design power. From a report: These chips are built on the same fundamental architecture as the company's Ryzen CPU cores, and they're aimed at the incredibly powerful data center market. AMD's 32-core / 64-thread Epyc CPUs combine four eight-core dies, each connected to the other via the company's Infinity Fabric. According to AMD, this approach is significantly cheaper than trying to pack 32 cores into a single monolithic die -- that approach would leave the company potentially throwing away huge amounts of silicon during its production ramp. The Infinity Fabric is deliberately over-provisioned to minimize any problems with non-NUMA aware software, according to Anandtech. Each 32-core Epyc CPU will support eight memory channels and two DIMMs per channel, for a total maximum memory capacity of 2TB per socket, or 4TB of RAM in a two-socket system. Each CPU will also offer 128 lanes of PCI Express 3.0 support -- enough to connect up to six GPUs at x16 each with room left over for I/O support. That's in a one-socket system, mind you. In a two-socket system, the total number of available PCI Express 3.0 lanes is unchanged, at 128 (64 PCIe 3.0 lanes are used to handle CPU -- CPU communication). Anandtech has a longer writeup with more details on the CPUs power efficiency and TDP scaling. Further reading: ZDNet, press release.

Comments

Hopefully not too late

[thegreatbob]

Who knows what Intel might have in their back pocket after chilling around for half a decade? I know benchmarks aren't everything, but at least there's some hope for them wiggling their way back into the server market.

Re:Hopefully not too late

[Anonymous Coward]

Who knows what Intel might have in their back pocket after chilling around for half a decade?

Haven't we already seen some indication that Intel's been caught flat footed? Their response to Ryzen has been strange and erratic with confusing product launches (or at least announcements...) the tech media is still trying to untangle, $1500+ 175W i9s that hardly anyone will buy and other bad ideas.

I expect their marketing people will sperg out and invent some bizarre new name for embiggened chips based on existing process tech and blow a few million on the "launch event."

The fact is AMD has caught up. So now there is competition and Intel has to compete. If Intel does have anything "in their back pocket" they'll have to use it to hold the market share they've earned. If not then they'll have to compete on price etc. It's all good. We win either way.

Re:Hopefully not too late

[gweihir]

Intel very likely has nothing. They made huge profits being fat and lazy and they expected that to continue. Remember that comparable small AMD has declassed Intel before. Intel may well need several years to catch up.

Re:Hopefully not too late

[tlhIngan]

I've gotten the impression that Intel has lots of stuff prepped, and ready to deploy, in response to any threat by AMD.

Yes, but only to maintain, but not eliminate AMD.

You see, Intel's goal is to keep AMD where they are. They WANT AMD to have a foothold in the server market where the money is. But not too big a foothold.

For Intel, AMD"s position is perfect - they are the distant second and will nip at the heels like an eager puppy. Which is where Intel wants them. Too powerful and they have competition, too weak and AMD could die off and unleash all sorts of government investigations, regulations, anti-monopoly rulings and other things. Worse yet, AMD's patents that are cross-licensed with Intel might go to ARM or others forcing Intel to pay or cross-license.

So Intel needs AMD to be around, but not too powerful nor too weak. They could crush AMD with what they have, but they won't, so they just need to deploy "just enough" to counter the threat without actually harming AMD.

So it's not a 32 core chip

[dreamchaser]

It's a cluster of 4 8 core CPUs connected via a high speed interconnect. I'm not saying that is bad; I just wish a tech site would have more accurate reporting.

Re:So it's not a 32 core chip

[VernonNemitz]

It looks to me like they finally figured out a small part of something I wrote about more than a decade ago. Too bad I couldn't afford to patent it....

Re:So it's not a 32 core chip

[scumdamn]

It's not 32 core die but the CPU itself is 32 cores. The CPU is what goes into the socket. This whole conversation has been going back to the mid 2000s when Intel released their quad core CPU that was really two different dies glued together. Nobody cared that AMD had a "true quad core" where Intel just had a higher-performing part that had 4 cores spread among 2 dies.

Re:So it's not a 32 core chip

[gweihir]

You need to take into account that AMD CPU interconnect and thread migration always was far, far superior to Intel. Usually you lose low single digits in performance on AMD, while on Intel people hat to use an additional system to stream gaming, for example, because Intel inter-CPU communication is so bad.

Re:So it's not a 32 core chip

[Kryptonut]

Too bad you couldn't afford to patent it....IF you were going to do something practical with it. That I have no problem with.

If not, I'm glad you didn't. That's the problem with patenting "ideas"....a lot fail to implement them, then stifle the progress of the industry because it's easier to litigate than innovate. There's a whole thriving industry around that, which contributes nothing to society.

I swear....if there was a short time limit on how long a patented "idea" had until implementation before being made invalid, we'd be so much further ahead than we are now.

Epyc?

[ArchieBunker]

How do you even pronounce Epyc? Like "epic"?

Re:Epyc?

[BronsCon]

I prefer it's pronounced "e-pike", as in an electronic pike on which they wish to mount Intel's head; in that case, they would have chosen that spelling knowing that most people would think it was to be pronounced "epic", thereby completely hiding their intentions.

Or I just say that for a laugh because yeah they're going for "epic".

Re:Epyc?

[__aaclcg7560]

The "pyc" is silent.

So these new AMD processors run Python compiled bytecode? Sweet!

https://stackoverflow.com/questions/8822335/what-do-the-python-file-extensions-pyc-pyd-pyo-stand-for

Re:My problem with AMD

[tonywong]

AMD made some huge mistakes but Intel was blatantly anti-competitive/abusive against AMD and got fined in the US, Europe and had to settle with AMD for billions of dollars.

https://www.extremetech.com/computing/184323-intel-stuck-with-1-45-billion-fine-in-europe-for-unfair-and-damaging-practices-against-amd

AMD wasn't exactly partying like they ruled the world, the deck was stacked against them by Intel when AMD had a superior product.

Re:My problem with AMD

[Anonymous Coward]

The reason AMD couldn't sustain is that Intel bribed their customers to stay with them as shown by the lawsuit Intel lost and the 1.2 billion dollars that Intel owes AMD. Intel has made many times that by essentially marginalizing AMD so the fine is most likely worth it to them. It is worthy to note that AMD still hasn't received a penny of that fine from Intel since they agreed to pay in 2009.

There was a separate, similar case in the UK with a similar outcome. Intel owes another 1 billion euros to AMD from the second case which it has been appealing unsuccessfully for several years.

Overall Intel owes AMD close to 3 billion dollars in fines but has successfully delayed paying it for the better part of a decade.

Re:My problem with AMD

[Gravis Zero]

Can you buy it? AMD arguably made "ok" CPUs (talking within past 2-3 years), and even some mildly interesting ones. But could you buy them anywhere?

Yes you can.

Re:My problem with AMD

[alvinrod]

This is just dripping with FUD. Several big players in the server market have already announced they'll be shipping products with AMD's CPUs. If you couldn't find any server CPUs from AMD in the recent past its because they didn't bother making any after a point because their Bulldozer architecture was so much of a failure that they left that part of the market. Just look at the Wikipedia article that lists their server chips and notice that the pretty much stopped after 2012 outside of a few ARM or Jaguar-based parts that were for micro-servers.

Also, the last time AMD put a dent into Intel, Intel started fighting back in a large number of ways that were later found to be illegal. Celebrate and relax, indeed.

Re:whay that is

[Kohath]

The other point is, Intel's positive re-enforcement loop of having the best processors because they had the most revenues from selling the best processors to develop new processor fabs has been broken by ARM. Competition from ARM architecture drives fab R&D outside of Intel, so much so that Samsung foundries were shipping Qualcomm ARM processors at 10nm while Intel was still stuck at 14nm.

 

No, Intel just dropped the ball on that (or decided to let go of it). They had an 18-24 month lead and they let TSMC and Apple catch up.

Maybe they decided that dragging out (prolonging the inevitable end of) scaling by 4 years or so would be more profitable between now and 2025 than galloping to the finish line.

Re:My problem with AMD

[PIBM]

I'm pretty sure he was talking of ready to use AMD servers, which have been nowhere to be found for the last 5 years or so.

Support?

[bill_mcgonigle]

Is AMD assigning engineers to help port stuff that has gone Intel-only in the years that AMD failed to compete in the server space?

I'm thinking of some of the newer Xen features, codec assembly, etc. If we have to wait a year to try Ryzen they may have missed their opportunity.

24 hot-swap U.2 NVMe drive bays supermicro

[Joe_Dragon]

24 hot-swap U.2 NVMe drive bays supermicro

now just think if each one had its' own X4 link and you still have PCI-E leftover for say 4 10G links.