Intel MIC (aka Xeon Phi)

[P5-133XL]

No. The Knight's landing is an add-on card. Software, like the folding cores, will not be able to access It's cores in the same direct way as a computer's native processor so some modification would need to occur to make folding work.

[RipD]

Not following. Video card processors are available. Why would an Atom processor variant not be? I'm also reading that Knights Landing will available as a stand alone processor.

[7im]

And video card processors have been around for decades and all operating systems and motherboards are designed specifically to use them by default. They also have mature and widely accepted drivers and programming languages for GPUs like CUDA and OpenCL to access those GPUs.

Phi is by no means mature or widely accepted.

It's great if it will fold, but they'll be lucky to sell a comparative handful of these things, none likely to ever see fah software.

I love that people dream big, but it's kinda of like asking if the new Maserati Ghibli makes a good soccer-mom car. Absolutely. Never late for a game.

[P5-133XL]

Not following. Video card processors are available. Why would an Atom processor variant not be? I'm also reading that Knights Landing will available as a stand alone processor.

The video card processors do not use the same folding code (software) as the CPU does. They required new folding cores just like the Xeon Phi will also. That being said, the changes needed to run on a Xeon Phi are probably relatively minor.

If the Xeon Phi ever becomes a single processor that replaces the main CPU on a MB so that the MB/OS merely thinks there is a single CPU with lots of CPU cores then maybe the standard CPU folding cores could run on it. However, currently the Xeon Phi cards are plugged into an external bus like PCI-e which forces all software to use some form of access method, different from the main CPU, to run a program on it. The folding software needs to be Xeon Phi aware which is a necessary change for the current folding software knows nothing about loading software on to an external device, running from it, and then moving the data off it.

Hopefully, that makes my statement clearer.

[btupsx]

Any advancements concerning client development? While instruction set and integration problems do remain, 3100 series Phi's are now dirt cheap. I'd gladly run one or two.

[empirebuilder1]

Knights Landing is Intel's new supercomputer chip, looking to bite into the market share of Nvidia and IBM in the massively-parallel-computing market. It has "more than" 60 cores crammed onto one chip with 16gb of onboard memory. The estimated clock speed of the cores is about 1.2 to 1.3ghz. The entire chip is aiming to be a drop-in replacement for nearly all x86 workloads that can be thrown at it, since it can natively run most x86 OSes.

The pricepoint of this chip will probably be astronomical, but for those looking for multiple compute tasks (video rendering, graphics, physics simulations and more, anyone?) this could be a boon for fast encode/render/processing times without the need for temperamental OpenCL or CUDA programming. I don't know how well it would stand up to a multi-GPU setup in F@H, but the benefit is in the versatility. It looks like it's going to be pushing the 3Tflop mark on double-precision floating point math (6Tflop in single-precision), which is more than Nvidia can say for it's own high end cards.

You'd probably have to be a millionaire to casually drop the kind of cash this system will demand, though.

[bruce]

Knights Landing ... has "more than" 60 cores crammed onto one chip with 16gb of onboard memory...

I don't think you appreciate what they're up against. GPU with over 1000 cores at under $700 and if that's too much for you, there are lots of slower devices at lower prices. Admittedly, you still need a separate CPU to run the OS, but FAH is already programmed for them.

[wdethoma]

When will the Xeon Phi be able to crunch? They have gone down in price.

[bruce]

FAH has a firm policy of not pre-announcing future developments. [The software industry has made pre-announcements in a bold half-truth]
The closest we can come to it is a choice between {today, soon (a very, very loose word), and non-soon] I'd have to call it non-soon unless Core_a7 already supports it. People with AVX are able to run some projects with 64 CPUs.

What does the log preamble show if one is present when FAHClient is started?

[rwh202]

There are posts on here with people reporting running the Nacl chrome native client and CPU slots on xeon phi - not sure if they're beta or more generally available or need any special effort to run.
Might be worth PM'ing them to find out more?

[Nathan_P]

Core A7 runs on Phi, not sure how he got it to run but I saw PPD of around 200k on project 11920

[toTOW]

A7 core runs fine on Xeon Phi ... it's seen as a SMP CPU, not a GPU ... PS3EdOlkkola could help on this.

[wdethoma]

thanks

[PS3EdOlkkola]

@wdethoma - responded to your PM.

I have the Intel Phi model 7210 running at 1.3 GHz, and is recognized by FAH as having 256 processors. I purchased the system completely built from a systems integrator, Colfax. It was roughly $5,500, has 108 GB of main memory, an SSD, large HTC hard disk, and is water cooled. There are at least two x16 PCIe interfaces. In reality, there are 64 physical cores (IIRC), but with hyper threading registers higher in FAH. It supports at new A7 AVX core. It runs Linux CentOS 7.x.

Performance wise, it runs best using 192 of the 256 cores and does about 300K ppd on most A7 projects. It's about equivalent to a GTX 980 at standard clock rates (no factory overclock). Needless to say, it's not a barn burner of performance. Another way to look at it: My Titan X pascal's put out over 4x the performance (1.2+ million ppd) for $1,200.

[bruce]

Performance wise, it runs best using 192 of the 256 cores...

The traditional systems on which FAH has been tested are typically in the range of up to 64 cores. Phi opens up a new horizon for CPU processing which has had, at this point, limited testing. It's really good to hear that it runs well as it expands into systems of this nature.

I'm going to guess that the 192 limitation you're reporting means that hardware-wise, something else in your saturates or the benefits of HT peter out before all the cores are saturated. It's also possible that the version of GROMACS used in FAHCore_a7 has some unknown limitations.

I have not checked gromacs.org to see if such a limitation is being discussed. You'll probably find the real parallel processing experts on their site. If you do, let us know.