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

[phoronix]

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

There's a crazy discount right now for those wishing to buy an Intel Xeon Phi MIC card that's equipped with 57 cores running at 1.1GHz, 8GB of onboard memory, passively cooled, and over 1 TeraFLOP of double-precision compute power. All of this on a PCI Express card for less than... $200 USD!..

http://www.phoronix.com/vr.php?view=MTgzNjY

[jonnor]

You need to write special code to make use of that hardware anyway. So this only has a benefit over a GPGPU if you need double-precisions support. A ~ 1Teraflop GPGPU costs 100-150 USD and these systems are widely deployed already.
No wonder Intel has to throw these at people to get them to be used. Wonder how much they are losing per unit...

EDIT: the Xeon would likely have a large advantage in integer-based number crunching, which is important to some areas - like the finanical.

[Del_]

I am afraid the GCC reference is just about operating system support for the card. I believe you can only use the Intel compiler for code. Pretty much makes it a no-go for me regardless of price.

[Michael]

I am afraid the GCC reference is just about operating system support for the card. I believe you can only use the Intel compiler for code. Pretty much makes it a no-go for me regardless of price.

Ah, great to know. Was going to ask you about that as I heard the GCC support might be limited to building the Phi's Linux kernel and since I think you mentioned trying Xeon Phi before.

[Gumix]

For quite some time now most of the Xeon Phi Linux enablement has been working its way upstream within the kernel ("AVX-512 CPU Support Added To Linux 3.15 Kernel")

No, these Phi's are KNC, which have their own 512-bit ISA, not compatible with AVX-512, which would be available in KNL.

in order to run some benchmarks on it, but with not having clear idea how well the Xeon Phi would work with OpenACC/OpenMP/MPI workloads, etc, for our common computational benchmarks or other computational workloads would benefit from the Intel MIC, I've decided to hold off for now...

Without ICC it would perform very poorly, since GCC for KNC supports only basic codegen, without 512-bit vectors, etc.

[Brane215]

First, this being essentialy x86 has very limited value for me. How should this be of any benefit for a user when s/he needs special compiler for the thing ?

Once I need special environment, I don't really care that much about familiarity with cores, compiler has to take care about those partiularities anyway.

Second, this thing has only 57 cores, swhich is not that much for a massively parallel unit.

Third, one of its merits compared to GPU was connecting matrics between cores and their ability to run code independently of each other and still achieve coherency. But peat FLOPS data suggests heavy useage of AVX or similar vector instructions. With them, it doesn't make much of a difference if I use GPU computing units.

All in all, no that interesting.

[Michael]

No, these Phi's are KNC, which have their own 512-bit ISA, not compatible with AVX-512, which would be available in KNL.

That link was just referring to the general Linux enablement of Xeon Phi rather than this particular model per se.

[rabcor]

And of what use would this device be exactly?

[zxy_thf]

Second, this thing has only 57 cores, swhich is not that much for a massively parallel unit.

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.

[gigaplex]

I followed the Amazon link provided and found a single hit at $499.99 plus postage. The Australian distributor mentioned on the Intel website doesn't even stock the discounted model. Colour me unimpressed with this sale.