[QUOTE=fivemack;431715]The machine comes with a free one-year license for the Intel compilers, but the licensing is for a single named user which isn't terribly helpful on a multi-user computer.[/QUOTE]

If the named user has admin access, would it be possible for other users to use the compiler, or does 'named user' imply 'only user'?

[QUOTE=Madpoo;431730]Unfortunately, Prime95 isn't "officially" an open source product... doesn't that mean it has to follow some guidelines about where the code is checked in and all that jazz? I don't know much about that aspect.[/QUOTE]

In case it helps, Mlucas is fully open source.

[QUOTE=ewmayer;431772]In case it helps, Mlucas is fully open source.[/QUOTE]

Good point... and I wonder about GMP-ECM, is that OSI compliant? I imagine there could be some interesting improvements if that were made multi-threaded to begin with (it isn't now, that I'm aware of), and with SO many threads on a Xeon Phi, and lots of RAM, well... it would be a pretty decent ECM factoring beast. Probably closer to what the NSA uses to crack all of our encryptions. :smile:

How is this different from a GPU?

[QUOTE=bgbeuning;431793]How is this different from a GPU?[/QUOTE]

It has nothing to do with graphics, it's really a parallel processor system.
I suspect the workstation system shown at Colfax is pretty much like
the one shown in Madpoo's very first post in this thread. It might not even
have a regular Xeon to drive it as a co-processor but run by itself.
That would mean any windoze support is unlikely ( unless WINE would work or
some VM ).
Intel will probably release co-processor boards later, but I think these stand-alone
systems come first to seed development efforts.
Here's a page that has lots good links abiut Knight's Landing stuff at Intel:
[URL="http://dap.xeonphi.com/#implinks"]http://dap.xeonphi.com/#implinks[/URL]

[QUOTE=tServo;431801]It might not even
have a regular Xeon to drive it as a co-processor but run by itself.
That would mean any windoze support is unlikely ( unless WINE would work or
some VM ).

Intel will probably release co-processor boards later, but I think these stand-alone
systems come first to seed development efforts.[/QUOTE]

The offering from Colfax was the processor model of the Xeon Phi, so yes, it would be the CPU on the system. Since the Atom cores are x86 compatible, Windows, Linux, etc shouldn't have any problems running on these. I guess you might have to consider the CPU count support of some OS's to make sure it can handle that many.

I just looked it up, and Windows 10 x64 supports up to 256 cores. Hmm...if I understand it right, a 72-core Knights Landing would present itself as 288 cores to the OS. Windows Server 2012 R2 (both standard and datacenter) should have no problems... something like support for up to 640 cores.

[QUOTE=Madpoo;431806]should have no problems... something like support for up to 640 cores.[/QUOTE]

640 ought to be enough for anybody!

[QUOTE=VBCurtis;431818]640 ought to be enough for anybody![/QUOTE]

I suspect that this post could be amusing in 20+ years.

[QUOTE=Madpoo;431806]I just looked it up, and Windows 10 x64 supports up to 256 cores. Hmm...if I understand it right, a 72-core Knights Landing would present itself as 288 cores to the OS.[/QUOTE]

I seriously doubt any memory-intensive program would be able to benefit from using anywhere near that threadcount/physical-core ratio, so that OS limit may mean little in practical terms.

It will be very interesting to see the kind of threadcount at which a memory-nonlocal algorithm like FFT saturates the memory pipeline. If that is anywhere near the number of physical cores, it will be damned impressive. (Note that the calculus here is *very* different than for GPUs, due to the massive-bandwidth and vector-processing capacity of the KL cores, each of which may end up being as much as twice as data-hungry as a Haswell or Skylake core on a per-cycle basis, though the underclocking needed to keep such a manycore system from melting will likely offset the per-cycle throughput gains due to the combination of AVX512 and DDR5.)

This thing is rather interesting as seems to run as a standalone processor and intel's notes mention it is Binary compatible with Xeon. Whatever that means. It should run Windows (with the right support and drivers).

[QUOTE=ewmayer;431823]I seriously doubt any memory-intensive program would be able to benefit from using anywhere near that threadcount/physical-core ratio, so that OS limit may mean little in practical terms.

It will be very interesting to see the kind of threadcount at which a memory-nonlocal algorithm like FFT saturates the memory pipeline. If that is anywhere near the number of physical cores, it will be damned impressive. (Note that the calculus here is *very* different than for GPUs, due to the massive-bandwidth and vector-processing capacity of the KL cores, each of which may end up being as much as twice as data-hungry as a Haswell or Skylake core on a per-cycle basis, though the underclocking needed to keep such a manycore system from melting will likely offset the per-cycle throughput gains due to the combination of AVX512 and DDR5.)[/QUOTE]

A virtual server *host* machine can easily take advantage of that many cores though. That's where a lot of the market for beasties like that is these days. That and massive SQL compute clusters.

When you think about it, even with today's high-end servers with 8 sockets... the Haswell Xeon E7-8xxx v3 has 18 cores. With hyperthreading that's 36 cores x 8 sockets = 288 cores visible to the OS.

Broadwell E7-8xxx v4 is looking like it'll top out around 24 cores on the E7-8890 v4, so 48 x 8 = 384 cores.

Knights Landing with 288 cores visible to the OS on a single CPU is pretty impressive, and I imagine it won't be too long before they actually get a dual socket version of something like that (or just up the core count on a single CPU?) Two KNLs on a single server... whew, 576 cores of fun is what that would be! :smile: And not 20 years away, more like 4-5?

Dual socketing a KNL would, I'm sure, need a LOT faster interconnect that QPI for sure. What I've seen about UPI (ultra path interconnect) seems like only marginal speed improvements, like 10.4 G but maybe with 6 memory channels instead of 4 it's able to transfer more bits in the same time.

Whatever the case, Intel bought all that know-how from Cray and they ought to be able to put it to good use, eventually.

[QUOTE=ewmayer;431823]I seriously doubt any memory-intensive program would be able to benefit from using anywhere near that threadcount/physical-core ratio, so that OS limit may mean little in practical terms.[/QUOTE]

Don't compare Hyperthreading, Hyperthreading and Hyperthreading, they all work in a different way. For the Xeon Phi before [I]Knights Landing[/I] we know it is impossible to achieve full performance of one core with just a single thread. Remember this mlucas benchmarks we did on Xeon Phi? There are architektures which do threading with a (more or less) fixed timeslot per thread just to hide latencies.

Oliver