Intel's Xeon Phi Is Being Sold For An Insanely Low Price Right Now

[woodbird]

Passively cooled 270W cards kind of require a rack server chassis to work. These things have to dissipate 270W of heat, and they have no fans of their own, so they basically require a rack mount enclosure with lots of air flow = fast and noisy fans.

So, no, basement nerds aren't the target audience. People like sysadmins for banks (very much ECC) and oil companies (subterranean imaging and shiny realtime display of said 3D data for geologists) are.

Couldn't agree more.

[wizard69]

They are not meant for general purpose computing. Phi's direct competitor is NVIDIA Tesla card. It is wrong to say that there's no software support. There are simply no software support for casual users like you. I do parallel computing on a daily basis and Phi's has the best software support among competitors (well, I might be bias, but I still think CUDA is not the future, OpenCL? after using it for two years I hate it even more than CUDA).

The thing with OpenCL is that it is supported on a wide array of hardware. That gives in a distinct advantage over CUDA.

Just like Tesla, the target market is not your desktop, not even high end workstations.

Actually Intel might have the intent here to drive adoption in high end workstations. There are users out there right now that can't buy enough computing power.

The target market is HPC servers, TOP500. The workstation cards are just there to wet the aptitude of developers, so they might convince their institutions to build their next cluster using them. For the same reasons, I received Tesla K20 samples before. These cards are not, and never will be for general purpose. And I don't think that is the reason to be doomed.

Never say never. Seriously Intel is improving Knights Landing to better support operating systems, I could see these chips making interesting work stations.

If you don't do HPC on a daily basis, then don't pretend you know what you are talking about. Don't start with ~200 dollars GPU. No, they are not usable. They have tiny ~2GB memory at most, inefficient double precision, and no ECC, and many differences that does not make a difference in graphics or even casual use of GPGPU, but make a big difference in HPC.

You might be right that Intel is trying to seed developers in the HPC world but I think it is equally possible that they are going after or maybe better said, trying to create an expanded market. The HPC world is a small space that can't realistically support the development and manufacturing of PHI. This is where the GPU manufactures have a big advantage, much of their development effort is spread across a number of products using similar technology. I don't see much of a future for Phi if Intel only sees HPC as only target market.

[rdnetto]

To clarify one thing, GPU vendors count every ALU as ONE CORE.
The actual number of cores which can execute instructions independently is way much less that they advertised.

A fair comparision needs also count the number of ALUs within a Xeon Phi processor, as the equivalent of "cores" in GPUs' terminology.

Even cores and clock frequency are pretty meaningless when you start to take architecture differences into account. e.g. there are Atoms today that outperform decade-old Pentiums while running at a lower clock rate. That's we have FLOPS.
You can get GPUs with up to 4 GB of memory and 2.6 TFLOPS of computational power for a similar price point, so unless you have a problem that is pathological on GPUs, there's little advantage to this.

Passively cooled 270W cards kind of require a rack server chassis to work. These things have to dissipate 270W of heat, and they have no fans of their own, so they basically require a rack mount enclosure with lots of air flow = fast and noisy fans.

Ah, that explains it. I was wondering how on earth they had 270W TPD without any fans.

[darkblu]

It's been already stated a few times, but for emphasis:

1. To use the SIMD instructions (which are _not_ SSE/AVX) of KNC/MIC you need Intel's own compiler, which is far from being free. So don't be fooled - Intel are not selling Phi's for any 'insanely low price'. Unless you plan to use them as a paperweight, of course.
2. '270W with passive cooling' is an oxymoron. You'd need serious chassis temperature control to get those to work stably at peak.

[Ardje]

First of all: the card is worth the $200 just to get the biggest vpenis.sh
Second: the card is probably capable of running a crappy wordpress site (wordpress sites do not have to be crappy, as long as you can see php as not crappy, but wordpress sites are often crappy) which usually needs more than 16 real cores to do 8 requests per second. The host cpu can then be used for real stuff.
I know it sounds nasty and stupid, but so do the crappy wordpress sites.

[Zetbo]

Even cores and clock frequency are pretty meaningless when you start to take architecture differences into account. e.g. there are Atoms today that outperform decade-old Pentiums while running at a lower clock rate. That's we have FLOPS.
You can get GPUs with up to 4 GB of memory and 2.6 TFLOPS of computational power for a similar price point, so unless you have a problem that is pathological on GPUs, there's little advantage to this.


Ah, that explains it. I was wondering how on earth they had 270W TPD without any fans.

Please tell me where can I buy 1TFlops of double precision compute power for ~$200?

This Xeon Phil offer is very attractive to developers who needs a crap ton of double precision computing power. Obviosly, there are few problems like that it has a TDP of 300W without active cooling. It needs a 64bit EFI support to boot which is not very common in motherboards, except some high end Xeon motherboards.

[Jabberwocky]

Please tell me where can I buy 1TFlops of double precision compute power for ~$200?

This Xeon Phil offer is very attractive to developers who needs a crap ton of double precision computing power. Obviosly, there are few problems like that it has a TDP of 300W without active cooling. It needs a 64bit EFI support to boot which is not very common in motherboards, except some high end Xeon motherboards.

TLDR; The Powercolor Radeon R9 280X comes close in terms of price and performance and it has awesome cooling . It uses a lot of power .

I got my Powercolor Radeon R9 280X for $200 USD albeit it's going for $250 on newegg at the moment. It has dual BIOS so I was not afraid of flashing it to test under OSs without proper over-clocking software. By dropping the voltage from 1.15V to 1.10V, over-clocking the GPU with 100MHz (1100MHz) and left the ram stock at 1500MHz x4 = effective 6GHz. I reached over 4TLOPS single- / 1TLOPS double-precision. I mined for 2 months with that BIOS. By undervolting to 1.05V I reached over 1.1 TFLOPS double-precision but it was not stable. The OS never locked up just caused the driver to crash causing the desktop environment to restart. AFAIK some Powercolor cards came with Elpida ram others with Hynix ram. My cards had Hynix ram. Temperature wise the GPU reached 75 Celsius with ambient temperature of 25+ Celsius (my room, I live in Africa). The voltage regulator was usually about 3 degrees warmer than the GPU reaching 78 Celsius. With proper ambient cooling / case airflow the card runs much cooler at max workload. Although the TDP is around 250W my card easily hits 400W. PS: second hand sapphire toxic cards are under $200 USD and even faster.

[Zola]

it's a 499$ & not wort a ?.

As it's actually 499$ it's not even wort of discussing.
How ever just to mention the advantage of many core sys is in unparalleled tasks execution as GPUs are only for high parallel workloads, FPGA's however are for good for with moderate number of threads & have best efficiency but are little harder to program. Intel is learning his mistakes with first two gen of PHI's from Althera Cavium & Tilera. Current PHI gen is useless & old. It can't operate in self stand mode & so on.
Now enough about this nonsense!
If you want something to discuss about I find this as a interesting example of how will future of computing really look a like.
It still isn't a SoC in a right meaning of the world as some of used blocks are not in a form of licensable IP. Sadly their is no player in the whole industry that is capable to delivery it on a single chip.

URL: http://coloradoengineeringinc.com/20...-cuda-support/

I find this interesting & so did a computing algebra department in Edinburgh, as I have a good friend working there.
At the end name dose also bring back memoryless.

[Yorgos]

I think that this piece HW would me same fate like Agea dedicated physX cards. Only HW without proper software support is worthless - even Bill Gates knew that 30 years ago..

Maybe some of you - linux geniuses - can drive ultimate kernel driver to use if as coprocessor for general usage, but i dont believe in it. If it would be possible, Intel would do it itself.
I think that this HW is damned from first design, same piece of crap like Larabee..

In fact, nVidia killed those cards because you could use them with the competitor's HW, so they woke up one morning and decided to stop the support for ever.

[Markore]

So it is card only for Server/Rack motherboards with Very good (and loud) ventilation, or Workstations, both already having ECC and that are compatible with card.
You also need Intel compiler that is non-free and will probably cost you more then the card(s).

As I understand from previous nice comments,
you would buy this card if you want to put it in your compatible Workstation to try it out with your specialized software, willing to buy Intel support software
and you know what you are doing with ECC supported specialized high precision floating point hardware, while deciding if buying many more of them for your company's crunching farm.