So how does Epyc (Naples) compare to Intel's new Skylake

You'll probably never know because nobody tests $10000 CPU in games

>>60543248
How are we suppose to know, you dumbass... There is no magical way to test processors without having them.
Fuck off now and lurk more.

>>60543248
>So how does Epyc (Naples) compare to Intel's new Skylake server processors?
It absolutely stomps them where it actually competes (1S and 2S)?

>>60543296
bitches are gonna complain about single core 1080p scores when AMD has always kicked intel's ass in multicore.

single core tests should have died in 2004 after 64 bit operating systems were released.

that would have forced software vendors to offer real x64 softwares sooner

>>60543310
Judging by specs, it completely destroys Intel, and on top of that it will most probably be cheaper and have lower running costs.

>>60543341
only projections then, ok.

>>60543331
Zen was always designed for multithreaded performance, so it should do really well for servers.

File: 1488594197699.jpg (59KB, 858x546px) Image search: [Google]

59KB, 858x546px

>>60543343
There's a fuckload of benchmarks out already with Ryzen chips beating Intel SERVER chips for much cheaper, so it's safe to assume Naples will only be better, since it's actually designed for servers.

It uses less power and does more for less(lower cost, more DIMMs, more I/O) so the TCO is no doubt lower, TCO is pretty much the most important metric here.

For every 10 Naples 2S motherboards you have to have 13 Intel 2S motherboards if you want the same memory capacity and 14 Intel 2S motherboards for the same I/O.
This is assuming Intel even offers the Platinum brand for 2S, which it seems like they don't, so there's a 40%+ core count advantage to AMD as well.

For every 10 Naples 1S motherboards you have to have 13 Intel 1S motherboards for the same memory capacity and THIRTY(30) 1S Intel motherboards for the same I/O capacity.

>>60543386
>For every 10 Naples 1S motherboards you have to have 13 Intel 1S motherboards for the same memory capacity and THIRTY(30) 1S Intel motherboards for the same I/O capacity.
Lmao.

>>60543386
So it's safe to assume AMD is going to take a large portion of the server market from Intel?

>>60543386
Guess the days of going for cheap 2S Xeon systems for I/O are over, especially if it can lower power draw and platform costs, and increase density.
Pretty impressive.

The prime market for this is machine learning no doubt.

>>60543386
nice
vpses will be even cheaper

>>60543437
>The prime market for this is machine learning no doubt.

Shit. What's the chances Google or Facebook start using Naples based chips?

>>60543512
You mean Google and Amazon, those are the prime pushers of machine learning.

Anyhow that ML market at both is still relatively small compared to their other servers.
Though both will probably wait a while before adopting new technologies, no large company like Google buys these things on launch day.

>>60543512
Someone from 7+1 will buy fuckton of EPYC soon.

File: 1167088194303.jpg (11KB, 211x251px) Image search: [Google]

11KB, 211x251px

>>60543248
Any word on if there'll be motherboards that support overclocking with these, QuadFX-style?

>>60543331
Well x64 doesn't exactly mean multi threaded but let's hope more devs pick up onto it. It has no downsizes except maybe the difficulty but that shouldn't be a problem unless you're a pajeet or something

>>60543533
7+1?

>>60543611
Google, Amazon, Facebook, Alibaba, Microsoft, I don't know the other 3, but could be Apple or Samsung?

>>60543635
Alphabet, Tencent, Baidu , ebay.

File: Screen Shot 2017-05-23 at 7.18.19 pm.png (53KB, 284x214px) Image search: [Google]

53KB, 284x214px

>>60543248
Their lying seems to be getting better.

>>60543611
http://fortune.com/2015/11/15/intel-super-7/

>>60543655
I'm sure you have a non-retarded reason for your post, mind telling us?

>>60543512
Google has already confirmed they are gonna use Naples, Microsoft told Intel to fuck off to

>>60543671
As you know it's the infinite sign. And right under of you look back at >>60543248 it
>Implying
the transfer is infinite.
Therefore it's misleading.

>>60543248
MOAR CORES

>>60543694
>Google has already confirmed they are gonna use Naples
When? Where?

>>60543700
It's the infinity fabric, anon.

>>60543331
Too bad 99% of shit is not, and cannot, become multicore. This includes: algorithmic where the next step depends on result of the previous one.

>>60543791
That's nice and all but there are more than enough applications that can scale up to insane core count, and even if there weren't servers don't run one fucking application at a time, but thousands of VMs

>>60543791
99% of CPU intensive shit is already multicore at least to some degree.

You don't need 32 cores goy, but take a look at our 28 core Xeons if you're interested!

>>60543850
>only 4S+
But schlomo, what about 1S/2S?

>>60543871
4S hasn't managed to penetrate over 8% of the market, it's pretty much exclusively used by universities.

>>60543812
Meanwhile, JavaScript shit keeps melting CPUs.

>>60544272
>interpreted bloat is bloat

No shit.

>>60543248
it rapes intel

>>60544299
My point exactly. Additional cores mean nothing in daily tasks.

>>60544379
We're talking about datacenter CPUs you fucking inbred fuck.

>>60544379
I'm sure someone will get $10000 Opteron/Xeons to open someone's facebook wall

>>60544386
Oh, it's all good then. Why did the first post I replied to mentioned 1080p at all?

In 1S, totally stomps them at compilation; loses at AVX2 or AVX-512 optimised workloads. Surprisingly, more efficient. Both are ES of course so things could change either way.

If they are inexpensive, and there are rumours that they might be, Intel should start getting worried about the 1S/2S segment for almost all workloads.

Pure HPC workloads with hand optimised AVX should stick with Intel.

>>60544641
Pure HPC usually uses VIDEO CARDZ.

>>60543435

Intel has already said it will concede marketshare to keep margins high.

>>60544671
Makes sense, gaining marketshare back is easier than getting margins up.

>>60543248
The difference between 28 vs 32 is a lot smaller than 12 vs 16.

With the increased IPC and higher clocks, 28C could very well beat the 32C in multithreaded workloads.

>>60544700
>With the increased IPC and higher clocks
It has the exact same IPC as BDW-EP and lower clocks than EPYC. And it's 4S+ only. Skylake-EP is DOA.

>>60544700
Intel's 28 core Xeon Platinum processors are limited to 4S motherboards.

Their 2S and 1S offerings are under 22 cores.
And it would be hilarious if Intel actually had higher clockspeeds than Zen, which seems outright impossible at the same core count, and we already know the 28 core clocks, 2.5GHz, Naples is expected to hit 2.8-3.0GHz, purely depends if its 205W or 180W, still unknown.
Either way AMD has an efficiency advantage, and this doesn't even take into account the far larger I/O count and 33% higher bandwidth/DIMM per socket count AMD has.


Most of all, 32 core Naples isn't competing with 28 core Intel because Intel is Intel, and isn't putting the 28core in 2S systems

>>60543386
Not just that, but in a tech show I watch, they talked about how some Xeons have a hidden NUMA (non-uniform memory access) node, so that some of the cores can't actually directly access memory. That feature really limits their usefulness for some applications and EPYC (it fucking hurts me to call it by that name, worse than Ryzen) won't have that problem, as far as anyone can tell. This is enabled by Infinity Fabric and I have to say, of all the stuff that AMD has been announcing and releasing, I'm finding IF by far the most interesting. I really hope IF performs well so that AMD can start building crazy stuff for us.

Theory: the graphics IP that Intel recently licensed from AMD was actually related to IF so that Intel can scale their processors better. Probably not, but that would be very interesting.

>>60544849

AMD has stated they won't be leasing infinity fabric out. That is theirs and theirs alone.

>>60544849
IF is clearly designed for AMD's designs, there are a lot of commercial fabrics out there that are licensed but they're more one size fits all gig, and Intel is not that kind of company.

Intel's money allows them to still use custom buses at every step, but the problem is that takes a LOT of time, ignoring the money.
And Intel is more strapped on time than money currently.

Also AMD might as well file bankruptcy if they license their most impressive feat, the infinity fabric.
It's literally their lifeline for the next half a decade.

>>60544985
>>60545016
Yeah, that's why I tacked that "probably not" at the end there. I agree that IF is too good to share. I'm very curious about how well it can connect to GPUs, though. Nobody seems to be talking about that. The entire tech press is talking about how Vega seems to sit between a 1080 and 1070 Ti in performance and that's not going to be good enough, but I haven't seen any speculation whatsoever on a two-die IF GPU, and given that AMD recently resurrected dual-GPU cards in the Pro Duo, it seems highly likely to me that we're going to see MCMs on the GPU. If AMD intends to do that and IF performs well, I don't think Nvidia has any answer for that.

>>60545091
IF is totally scalable to the needs of a chip and it was designed to be totally scalable from the start.
Ryzen uses two bi-directional 128b/clock IF units to address the memory channels, one per CCX (one channel linked to each CCX), and one for the southbridge.

The rumor is Vega has a 512b/clock IF mesh network, though absolutely no details are known at this time, and I suspect a full mesh topology network is unnecessary WRT the new L2$ implementation as well as considering power requirements.

We might see a full mesh IF in Navi as a requirement for unconstrained MCM communication.

>>60544470
muh gaymen meme

>>60544641
>Pure HPC workloads with hand optimised AVX should stick with Intel.
That'd be rare as fuck compared to shit like VM/webserver/database/gpu host

The GPU hosting case will be interesting - the extra pcie lanes over intel's offerings might mean we see the unholy abomination of amd servers stacked with nvidia gpus for the machine learning meme

>>60545606
>unholy abomination of amd servers stacked with nvidia gpus for the machine learning meme
AMD also has meme learning GPUs coming.

>>60543706
Why wouldn't they use Naples. You need 10 epyc mobos to equal 20 1s Intel Mobos

>>60545635

CUDA is so entrenched that AMD could have a 20% better chip, bundle gold dust infused blow and the finest underage prostitues, and NVidia would still outsell them. It would take many consecutive years of AMD having the better hardware and software support for people to switch. Brand loyalty does exist among the people making buying decisions, afterall, "noone gets fired for buying IBM".

>>60547581
Yeah, they might as well not fucking try. Right?