PSA: Xeon phi is not a piece of shit

I'm buying an AMD processor for my next build because i hate Intel. Also, i tell all people who asks to buy an AMD laptop because they can play games and because Intel ones are made with shitty parts. 13 AMD laptops and counting

>>57064538
>over 3 fucking TFLOPS of FP64 performance or ~10X the performance of a $1,000 i7-5960X

whoop-de-fucking-do.
A GP100 does 4-5 DP TFLOPS and a 3 year old Hawaii does 2.6-2.9 fp64 TFLOPS.
What you need to show is exactly what kinds of modelling or whatever actually benefit from having Atom cores steering the SIMD execution units instead of GPU-style SIMT schedulers.

>>57064597
Well this processor is not meant for consumers. This shit has a batshit insane amount of processing power and performance-per-watt (~250 watt TDP).

I used to think since they used airmont atom cores they were essentially just like 10 phones with quad-core atom SoCs glued together. They're not. I underestimated the raw power of these things by a very very huge margin.

>>57064772
I don't care. Fuck you. AMD has my money and the money of all the people around me

File: skeleton.jpg (84KB, 600x380px) Image search: [Google]

84KB, 600x380px

>>57064726
>What you need to show is exactly what kinds of modelling or whatever actually benefit from having Atom cores steering the SIMD execution units instead of GPU-style SIMT schedulers.
Video encoding and very accurate ray tracing that don't suck ass come to mind. Though we may never get consumer grade knights landing processors which is sad since the only way to speed up video encoding is to connect a bunch of desktops together through cables and a server OS which in the end result in like 1/10th of a TFLOP of FP64 performance.

>>57064793
Whoah hold your horses, amdcuck. I'm not even an intel fanboy, just wanted to correct my and other's misunderstandings of xeon phi processors.

I'm sure AMD will release something similar to knights landing soon enough. They have been fucking garbage in the enterprise market so them developing something to compete with knights landing xeon phi processors would do them great good.

>>57064849
>I'm not even an intel fanboy
Of course not. You don't have enough money for 1 Intel cpu. You might be shilling through your $59 Android.

>>57064955
Ironically I have an i3 + Rx 480 and an unlocked Note 4. Planning on getting and AMD Zen processor next year though.

>>57064973
Nobody believes your lies. I will leave the thread. I don't want to get mad

>>57064538
>72-core variant at 1.5 GHz
>$6,200
yeah i'll pass

>>57064818
>speed up video encoding

How important is this really when new mid-to-budget tier GPUs can already do shit like 4k60 HEVC live encoding?

>>57065007
It's actually very economical for the performance it gives you. Not just that but it also means compatibility with x86 software. You could run 72 VMs with each ~40 GFLOPS of FP64 (quad-core desktop AMD A8 APU), 4 threads, 1 physical core, 2-4GB of RAM from a single machine. Or put a bunch of these in server blades and those in a server cabinet and have your very own at home super computer.

>>57065040
>How important is this really when new mid-to-budget tier GPUs can already do shit like 4k60 HEVC live encoding?
Because while GPUs can do video encoding the results are huge fucking file sizes or very shitty quality. Using a GPU to encode video is pretty much like using the ultra fast preset with the baseline profile with software encoding.

>>57065040
ask disney

>>57064538
My $300 GPU gets 6 TFLOPS of FP64 performance, how much does a Xeon Phi cost?
>$2,000
Ya fuck that Pajet.

>>57065078
>It's actually very economical for the performance it gives you.
No it fucking isn't. You can get the same performance from a W9100.

>>57065126
Yeah but those wouldn't be x86 cores, that's really the only advantage of the Phi.

>>57064538
if I buy this can I run a minecraft server?

>>57064538
But what can I, as a not very rich college student, use one for?

How much do they cost too?

>>57065108
>My $300 GPU gets 6 TFLOPS of FP64 performance, how much does a Xeon Phi cost?
Source: my ass

>>$2,000
72-core variant will probably cost like 4-5K

>Ya fuck that Pajet.
>what is x86 compatibility
>what is performance-per-watt

>>57065150
Thank you.

>>57065151
You could run 72 separate ones each inside their own VM.

>>57065150
>Yeah but those wouldn't be x86 cores
Who cares? How much software is specifically looking for x86 to do this kind of work on? Would they even work with this expansion card? Barely anyone has a Tesla card (that's why they've gimped the Quadros) and the reason is probably nothing supports it and it offers nothing to people who can't get custom software written for them.

>>57065165
>But what can I, as a not very rich college student, use one for?
Not much that would really interest you. HEVC video encoding and accurate ray tracing are about it for an average college student. Maybe also running a bunch of VM for family members and friends with shitty computers.

>How much do they cost too?
Too much for you and I. 72-core variant is rumoured to go for 4-5K a piece.

>>57065170
Wow blunder of the century, I forgot we were talking about Intel's offering.

>>57065170
Nigga this shit would let you run a bunch of noice VMs all at once with how much fucking RAM it supports. This would be fucking perfect for schools or government offices. If someone has IT problems, you just go into their VM and fix it without getting off your ass.

>>57065200
As far as I know there are no VM programs that will allow you to use the x86 cores of a Xeon Phi for a VM.

I might be wrong, but I haven't really kept up with Phi since ~2014.

>>57065200
>Nigga this shit would let you run a bunch of noice VMs all at once with how much fucking RAM it supports
So would a half decent dual socket server for the same price as this add in board.

>>57065253
These are fully fledged x86 bootable x86 cores none of that co-processor gay shit that the first gen of xeon phi shitbirds had.

>>57065264
True but they would not provide even a quarter of the raw performance of this thing or even allow you to run 72 separate VMs at the same time with each one having access to 4 threads.

>>57064538
Why would anyone buy this when even moderately priced current gen GPUs will easily thrash it in performance?

>>57065185
Well I've been thinking of making my own YouTube channel for a while out of boredom.

What benefit does this offer over something like a used 7970 or 680 that I can get for ~200aud last I checked?
Is it better at encoding videos?

I don't mind the price if it's worth saving up for.

>>57065253
>>57065200
https://software.intel.com/en-us/forums/intel-many-integrated-core/topic/389511

aperently the only hypervisors with phi support are xen and something called scalemp. so it definitely exists

>>57065328
Mostly x86 software compatability and the ability to run a shitload of VM at the same time. Turns out these things have legit 100% x86 cores unlike the first gen xeon phis.

>>57065098
HEVC is easily capable of 4k60 4:4:4 at less than 100MB/s. That's a pretty low bit rate for the quality.

>>57065346
That was about accessing the cores to use THROUGH a VM with PCI pass through, not using the Phi cores themselves to run the VM.

It would just let you use the Phi inside of a VM running on a normal CPU.

>>57065299
>True but they would not provide even a quarter of the raw performance of this thing or even allow you to run 72 separate VMs at the same time with each one having access to 4 threads.
You realize this is ADD IN board, right?

>>57065299
For the price of two of these you can buy a 1U rackmount server with dual Xeon 12 cores.

>>57065404
>You realize this is ADD IN board, right?
For now. This ain't no manlet co-processor anymore.

>>57065167
>performance-per-watt
Worse on the Xeon Phi than on most GPUs

>x86
Which is not used very often for compute-intensive tasks.

This is a product without a market.

Anyone else getting a strong shill vibe from this thread?

>>57065426
>For now. This ain't no manlet co-processor anymore.
No, forever. This is just Xeon-D by another name, really, and look how that hasn't caught on.

>>57065437
Yea intel always shills their $2k+ super niche co-processor to poor NEET weebs.

It's an untapped market for sure

>>57065462
Well then why the fuck is the response to every argument "muh x86, muh VMs"?

>>57065409
Will that provide 288 threads, 384GB of RAM, and over 3,000 GFLOPS of FP64 performance?

>>57065462
How else are they going to interpolate their anime to 144Hz in real time?

File: 1474381649787.png (577KB, 480x476px) Image search: [Google]

577KB, 480x476px

>>57065495
> counting HyperThreads as if they mattered

>>57064597
>AMD laptop
>Good

>>57065495
Nope, but you can run enough VMs on it for the average office.

File: 1471463190481.jpg (12KB, 241x230px) Image search: [Google]

12KB, 241x230px

>>57065520
>>57065495
> thinking 384GB is a lot in a rackmount server

File: 10829374.jpg (25KB, 536x536px) Image search: [Google]

25KB, 536x536px

>>57065462
>>57065499

>>57065520
>I'm literally retarded: the post
And of course you would post a frog.

>>57065548
HT is garbage for anything besides running identical processes in parallel, mate.

If you are the one trying to peddle KL as a VM solution, you are genuinely retarded.

>>57065573
Each of those physical cores on the xeon phi output like 40 GFLOPS of FP64 computer power which is what your ~3GHz AMD A8 APU can output. Cut that into 4 threads and you have a very decently powerful VM right there.

In fact this would be especially good for VMs because of how little electricity would be used. 72 Desktop computers with A8 APUs: ~4.5kW/h. 1 server with a 72-core Xeon Phi: ~0.3kW/h.

>>57065642
Who needs FP64 compute in a VM though? I'm sure there are some use cases but it's not normal desktop use

>>57065642
Help me out here, I must be missing something...how are users supposed to access their VM without some sort of machine in front of them?

Besides, this is what a terminal server is for. Why the fuck would you set up a bunch of individual VMs to have to manage when you could just set up one server with Terminal Services installed and manage all of them at once?

>>57065688
>>57065702
ignore him, that guy's a fucking retard.

The peak TFLOPS are only achievable with AVX-512 instructions, for which there's virtually no software except custom applications.
Similarly, you only have 72 Atom cores running at 1.5 GHz or whatever, which is complete mobile phone tier integer performance, which would be made even worse under HyperThreading for most loads.
Finally, the 384 GB limit is absurdly low for a so-called beefy VM host, and only achievable with absurdly priced 64GB DIMMs.

A 2S Xeon system gets you 24 DIMM slots for 768 GB of affordable memory or 1.5 TB if money is no object, and system memory matters infinitely more for bulk VM hosting than throughputs in flavor of the month SIMD.