AMD Launches 64-core / 128 Threaded 2nd Gen AMD EPYC Processors

I'm not at all a software expert. But if you can't meet your expectation with 40 cores, that may be that the app ( the task the app has to solve or the implementation ) is not a fit for multi threading. The 256 megs of cache are to be divided by the 8 chips, so is like 32 every 8 core, not a super jump compared from the 28 of xeon.

 

lets see if dont have compatilibity issues like Ryzen =)

 

Some tasks just need huge amounts of processing power.
We make frame rate conversion software where I work, and for the quality and to be in real-time we use three top end NV cards all crunching away in openCL.

We looked at doing a pure software version, but didn't get close with a dual Broadwell Xeon machine.
Perhaps with 8 channel memory and dual Rome machine, it may be possible.

 

You overlooked one thing - as you mentioned, the cache on EPYC is divided for each chip module, but the cache for the Xeon scalable family is also divided for each of its cores. The Xeon's 28MB is effectively divided by 20 so that's about 1.4MB per core while the EPYC has 4MB per core. That is a BIG improvement. As is the clock rate. This should be a really good processor for servers.

 

Will the new Xeon present itself as two Numa nodes, I wonder?

Without an I/O die, it will have issues. Until Intel can go full chiplet and port whatever core they have to a configuration matching what AMD is currently doing, it will only be their market share and software people that keep them afloat.

Hardware wise I cannot see them recovering before 2022. Next year AMD will most likely present EPYC CPUs with hardware accelerated chiplets. These will probably be dedicated AVX-512 units, or Javascript accelerators or anything in between. Until 5nm they won't have the die space for making larger CCXs, so dedicated hardware accelerators are the next step.

Apple is already doing it on mobile, their CPUs are the most interesting hardware I've read about for a long time.

 

For desktop use you should never fall below 3GHz from my own experience (2X xeon 10c/20t ) and even in hard personal use i keep thinking that i don't need so much core (virtual and real)

to play Fornite? lol

If you don't know the awnser is that you don't need them, because the reason why you need them is in your question

 

On other hand it was when AMD started fail over fail on server segment that the company hit the ground everywhere...
Xeon where not exeptional at this time but Intel was the best that left on the ring.

Sadly for them they should have done that years ago, because it's not as easy as "normal" market segment.
I guess it would take time.

 

8-Core Ryzen
----------------------
24 PCI-Express lanes
32MB of L3 cache
Dual-Channel ram
Not ECC compatible
Single-Socket Only

8-Core EPYC
----------------------
128 PCI-Express Lanes
32MB of L3 cache for the lowest end, 128MB of L3 cache for the highest end 8-Core EPYC
8-Channel RAM
ECC Compatible
Single, or Dual socket configurations
Dual-Socket 8-Core per CPU system allows up to 256MB of L3 cache for a 16-core system, whereas the highest end 16-core EPYC single CPU only has L3 128MB of cache.

Not every workload takes advantage of a ton of cores, but that doesn't mean the workloads that don't, also don't take advantage of everything listed above, and more.

 

To my recollection, all AMD CPUs support ECC, but, not all Ryzen motherboards support it
Generally speaking though, you should assume it won't work.

The idea of having a dual-socket 8-core Epyc is kinda funny to me, but, maybe there's a use for that much L3 cache. I'm not sure what it'd be but I sure would be interested in seeing benchmarks comparing to a single 16-core CPU, and, if/when the 16-core Ryzen gets released (with it's "measly" 32Mb).

 

I want this 8 core cpu even more just for the cache! Us ARMA freaks understand the need for this.

I wonder can we get 4.5 GHZ off the 8 core?

 

Hi there, @rl66, please note those dual xeons of my are UNLOCKED by hack, meaning they live in the 3.5ghz range for all 72 threads, when avx2 is used it is a lot more from the stock 2.3ghz all core/ 2.00 ghz base clock, i.e. around 3ghz, also my TDP is as well unlocked by efi hack, as in unlimited. ( water cooled). Dont look at the cpu z screenshot that says 2ghz, as that is the spec of the cpu, not the actual state.

https://imgur.com/isQmEOH

BTW, I consider this system future proof against upgrades for the next 10-20 years as I see it. Buying a spare identical motherboard to keep for any case. I dont see a use case where I will need more performance than this.

 

The problem with that CPU is that the inter-CCX latency will now increase every two cores instead of every four.
Also the L3 is per CCX, it's still massive but a CCX has to go to the I/O die to access the same data as in the cache of another CCX.

I can't really imagine an 8-chiplet 8-core. If only a single core is activated per chiplet, then it will have 128MB of L3 accessible.

 

See this:

https://www.servethehome.com/new-supermicro-amd-epyc-7002-platform-innovation/
Looks like EPYC's 7742 competitor from Intel is actually 8280L!@ $17K with 4.5TB mem support vs 4TB on EPYC!,so AMD is actually were very conservative to compare their pricing vs 8280M @ $13K with only 2TB mem support

 

doesn't xeon and intel ecosystem has better compatibility with virtualization, server scenarions, hardware and software environments for proffesional applications, as opposed to AMD?

 

This is what Intel wants you to believe.
But has some truth innit, because until recently, Intel was giving the trend in data centers

 

Also peculiar, is the fact that intel has the upper hand in gaming performance, which to me is very important.

 

No from what i demand when i was wanting an Athlon PRO 220 GE to use the DDR4 ECC i can have for free (source AMD itself):
-Nearly every motherboard can have ECC memory and work with PRO iteration of the CPU.
-Ryzen doesn't use ECC (it use the ram as normal ram... But remember that ECC ram are slower and expensive)
-Ryzen PRO use ECC (it use the ECC like xeon as an exemple)

On pro segment there is CPU for nearly every use (exemple the Xeon W 2102 4C/4t on 2066 socket or 24 core at 1.8Ghz to work in 1U server) each one are intended for a specialised task and it will be the same for Epyc.

 

The main problem with AMD on Pro segment is the lack of service and lack of corrective if it doesn't work correctely.
AMD platform is relatively young while Intel is more or les the same since long time.
For a company it can be a problem while for a home use it would be a buy argument.
CPU vs CPU each one have pro and con... but it is nearly the same.

Then Epyc is not for you better get a Ryzen (not a TR).
Still yes but AMD get better in the exercice at each gen.

 

In a flash: No.

Unless you do AVX-512 stuff. Even then, EPYC is better or the same, just because of the new AVX256 units.

Security-wise, Intel is in a mess. Hyperthreading should really be disabled on servers, and storage and network I/O is bad after the mitigations.

They need to drop prices, hard.