phoronix.com
An EPYC 7601 vs a 2S Xeon Gold 6138 in the Photonix Linux test bench. If you want to skip the read, 64 Zen threads more or less = 80 Intel threads. With less TDP, more I/O, more RAM capacity, and less money. Intel only has AVX left, so if you aren't using AVX EPYC is the better chip to get in every way 98% of the time. AMD is going to take Intel to Poundtown for the next couple quarters.
>only multithreaded tests
kys poojet
Oy-vey.
Oy gevalt.
>64 Zen threads more or less = 80 Intel threads. With less TDP, more I/O, more RAM capacity, and less money.
AMD employees CONSTANTLY post on shitronix, the site is compromised
The EPYC he's using is also sent by AMD and not bought himself, it's always gonna be shown in a good light (with fancy compiler changes and settings)
OY VEY
The definition of holocaust.
Most of those apps don't support dual socket, AMD EPYC is barely faster at 32 cores than Intel's 20 cores
Should I buy 7551p box?
Oy vey to you too, rabbi.
dellid this
>apps don't support dual socket
Intel shills are now really getting desperate
...
And just when i think those apple shill is getting closer, you guys raised the bar again.
>180W AMD vs 120W Intel
>AMD still loses
>$4000 AMD CPU
>vs $2000 Intel CPU
>2S
Oy.
>2S
Vey.
>apps not multi socket aware
>apps effectively run only on one CPU
>apps
>p
>p
>s
>apps
The absolute state of MACACO
...
...
Slower than Intel at higher TDP, I didn't expect much more, AMD is still a few years behind Intel technologies
Pretty much agree, just look Vega, lower performance at higher power.
I can almost imagine him crying into his pillow after writing the nonsense ITT.
its two 125w Intel processors vs a single 180w AMD processor nerds
Unfortunately for you, one doesn't work most of the time due to app incompatibility
So it's really 32 AMD cores vs Intel 20 cores
>tfw
>app incompatibility
lets just do some linear numbers here, I know it doesn't scale like this but it'll show you're full of shit
the E5-2687W v3 is a 10 core 20 thread with 3.1ghz clock
the 6138 there is 20 core 40 thread with 2ghz clock
Skylake is roughly 5% faster per clock than Haswell, so the 6138 should be 35% faster than the 2687W with perfect scaling, while two 6138's would be 171% faster. In real life it'll be lower, than these numbers.
When you do percentage calcs on all the benches, and average the percentage, the dual 6138 system is 131% faster than the 2687W.
Breaking it down into further categories, we can assume if the 2687W is faster then the application does not thread well, but if the 6138 is faster, but not MUCH faster, then it's likely the multiprocessor isn't working as you claim.
So the benchmarks where its not much faster, ie where multiprocessor failed, are Rodina Streamcluster at 21% and Darktable Server at 32%. The benches where it was slow, ie threading didn't benefit as much as clockspeed, are H.264 encoding and Darktable Boat/Masskrug.
Basically you're a shill grasping at straws.
Objectively wrong, you can't compare CPUs with pure numbers off the specsheet
>app
Amd suck at marketing, real price high end EPYC is $2100 and $4200 dual socket.
Mostly people believe EPYC single CPU cost $2100, AMD really suck at marketing.
>marketing
>datacenter
>CPUs
?
>better for a bunch of meme shit the average user doesn't care about
then why is the 6138 consistently getting above 100% scaling
>implying average user spends $300+ on a CPU
$3400 per unit for dual socket, $2100 per unit for single socket, same clocks and all. The $4200 unit has faster clocks than the $3400 model.
Now post Epyc vs Xeon gaymen benchmarks, goy
Yeah like not stuttering.
> Most of those apps don't support dual socket
Isn't socket count transparent to software (and to some extent the OS), showing up only as available cores... or am I missing something?
He's a braindead intlel shill.
>on Linux
K.
The current state of shilltel.
>Running windows servers
LUL
AMD is now selling 32c/64t in one single chip.
Starship will feature 48c/96t.
2S starship platform will feature 96c/192t.
Cannonlake is 100% more dense than Skylake.
Skylake = 28 high IPC cores
Cannonlake is 28+100%+10%IPC = 56 cores
>32/64 for $2100
wtf I love AMD now
32 AMD cores = 20 Intel cores
Ugh, it's the opposite. 1 epyc > 2 xeons on non-avx loads.
Basically, 32 AMD cores > 40 Intel cores.
Are we gonna see a sudden surge of AVX optimized program now?
If we suddenly decide doing gpu loads on cpus, yes. Otherwise, nah.
Cannonlake server was cancelled, and 10nm has performance regressions.
And 1.7% (Jensen approves!) yields. And that atrocious mask count.
No since there's two companies capable of producing FP64 capable GPUs.
>that 6T SRAM cell size
One of the reasons I decided against a dual xeon pc a while back was because many of the programs I would use don't like dual socket.
Cannonlake is a joke. 7nm will be the real deal.
Yes. 7LP DUV is denser than Intel's 10. 7LP EUV will tear Intel apart.
Nah, there are always programs that refuse to use dual sockets, the video renderer I use/use to use was one of them.
Intel's 7nm is somewhere 2021.
Right when Common Platform *might* get bored and would launch 5nm GAAFETs.
Ugh, that's not how dieshrinks work.
As far as I know, and my knowledge may be a bit out of date, but 6nm was a hard limit due to quantum tunneling. Did they ever figure out a work around?
Yes, they (Common Platform) went back to 2D and started stacking silicon nanosheets on top of each other.
You can also use doped III-V FinFETs.
the 7nm he is talking about will take amds cpus from 3.2-4ghz up to 5gz+ for normal use, as in this is where the cpu is optimally clocked for voltage and power use, where as intel can clock to 5ghz, but fuck me is that a bad idea for power use
Intels 10nm will be behind their current 14nm process for 2 generations by their only predictions, real world may be even less kind in this regard. while gf 10nm is ahead of schedule and on track.
Man IBM selling their fabs along with 16k patents and staff were the golden ticket for GloFo.
when you do that, do you gain anything or is there any performance regression? I could see low power who gives a flying fuck about performance applications jumping at smaller dies, but 5nm at the cost of performance or even just equal performance would be a sidegrade at best, any thing you can tell me in that regard?
I know nothing of node characteristics.
IBM only recently managed to fab a test chip (think of some dieshrunk POWER9) on 5GAA using EUV.
Node discussion is so comfy.
Almost no shitposting.
IBM's research is fantastic but the company is a dumpster fire in terms of doing anything with a good idea.
t. Former IBMer
That's why the only thing they are doing in Common Platform is precisely R&D.
Ibm is more about the long game in terms of their research opposed to short term profits.
This
Where's the 320x240 cs:go bench
3/10 bad bait made me reply
This desu
I want my 32core Epyc vs Xeon Platinum 320x240 gaymen benchmarks.
POWER9 will have 12c/96t.
they're the ones that discovered and publicized the ryzen segfault bug, pretty strange for an AMD shill site
Geez the Intel damage control in here, though understandable. When Intel starts laying people off the shills will be first to go :^)
While EPYC is pretty dam good I think adoption will be slow for the first few quarters while early adopters keep putting it through its paces. Once the good word gets around though by next year, Zen2 servers might be a thing, and that's when the real silicon holocaust begins.