Knights Landing reservations
 

I'm doing some research for my day job and looking specifically at Knights Landing. Mainly just catching up to date on the latest server processors and this was a tangential look.

Based on what I'm reading, the upcoming Knights Landing CPU seems like a...how shall I put this... "friggin' awesome" CPU.

60-72 cores with 4 threads per core / 2 dual-precision FPU's per core. It's obviously a massively multi-cored piece of tech.

What catches my eye specifically, in regards to it's potential for GIMPS, is the near-memory feature.

With 8-16 GB of "near memory" with bandwidth of 400 GB/s and up, it really takes a huge chunk out of the mem bottleneck I've seen on my E5-26xx based servers. Those have just been getting hammered with memory bottlenecks past a certain point, making the multi-threading workers start to throttle in performance pretty drastically.

You really notice it when spanning multiple sockets when you might only have a max 25 GB/s QPI between chips... you might see a little bump adding one extra core on the other socket, but beyond that you actually start losing some ground.

Having all of those cores using a very fast 400 GB/s memory system with that much capacity, I guess I'd have to think it out a bit more and reckon how many transfers/second Prime95 would actually be generating at different ranges, but it's certainly better than under any CPU out there now.

Anyway, thought I'd throw this out there:
[URL="http://www.theplatform.net/2015/03/25/more-knights-landing-xeon-phi-secrets-unveiled/"]http://www.theplatform.net/2015/03/25/more-knights-landing-xeon-phi-secrets-unveiled/[/URL]

I guess part of the beauty of having it integrated into an actual CPU is that it should be somewhat easier in that package to actually just run GIMPS and see how it goes, rather than the coprocessor options.

I'm also guessing that the LL and TF routines themselves might benefit from some tweaking to take advantage of AVX-512 and any of the other fun things.

Thoughts? Anyone out there with any actual experience with any of the current Xeon Phi coproc boards and maybe has something ported to run on it?

It's probably going to be expensive, but it looks nice!

[QUOTE=Madpoo;399131]
60-72 cores with 4 threads per core [/QUOTE]

One thought.

HYPERTHREADING: OFF!

[QUOTE=aurashift;399145]One thought.

HYPERTHREADING: OFF![/QUOTE]

As I understood it, the threads per core of KNL is different than the "virtual" core of current CPUs. Admittedly I've only scraped the surface when it comes to reading the material, but that was my initial takeaway.

Oh, also, I found this older thread talking about Xeon Phi here: [URL="http://www.mersenneforum.org/showthread.php?t=18223"]http://www.mersenneforum.org/showthread.php?t=18223[/URL] and it was mentioned that the upcoming (and as-yet unnamed) Knights Landing would change things up in a good way.

[QUOTE=Mark Rose;399141]It's probably going to be expensive, but it looks nice![/QUOTE]

It sounds friggin' awesome - similar SP throughput as the best of nVidia's current offerings, but the the DP throughput blows anything from nVidia away. (LL needs DP, but the vector-int capacity using the 16 x 32-bit int-compute capability of AVX512 - should also be very impressive, and TF can use either DP or int math, depending on which proves faster in practice).

Hoping that there will be no issues installing a CoPro one of these in the PCI3 slot of my MSI Haswell mobo, and further hoping that slightly-reduced-available-cores versions of these (the article notes the yield issues, and makes it sound like selective disabling of faulty cores is no big deal) will be available for under $1000.

The fact that current x86 code using standard multithreading protocols (thus the main recoding will be going from AVX 256-bit to 512-bit) can be built without having to first plow through a 1000 pages of dense "custom threading and language extensions" as for nVidia is a bonus.

Question: For folks like me (and likely most GIMPSers) for whom 8GB RAM is more that enough, will we even need to buy any extra DDR4, or can we just go with the on-die near-memory?

[QUOTE=ewmayer;399167]Question: For folks like me (and likely most GIMPSers) for whom 8GB RAM is more that enough, will we even need to buy any extra DDR4, or can we just go with the on-die near-memory?[/QUOTE]

Check that the system can boot without RAM. Even if you plan to run a system without a display, some kind of GPU is usually required. The same could happen here.

[QUOTE=ewmayer;399167]Question: For folks like me (and likely most GIMPSers) for whom 8GB RAM is more that enough, will we even need to buy any extra DDR4, or can we just go with the on-die near-memory?[/QUOTE]

I don't remember if it was in that link or one of the others I found about it, but it sounded like there would be 3 different memory modes available.

Oh, yeah:
[QUOTE]The interesting thing about the Knights Landing processor is that it will have three memory modes. The first mode is the 46-bit physical addressing and 48-bit virtual addressing used with the current Xeon processors, only addressing that DDR4 main memory. In the second mode, which is called cache mode, that 16 GB of near memory is used as a fast cache for the DDR4 far memory on the Knights Landing package. The third mode is called flat mode, an in this mode the 384 GB of DDR4 memory and 16 GB of MCDRAM memory are turned into a single address space, and programmers have to allocate specifically into the near memory. Intel is tweaking its compilers so Fortran can allocate into the near memory using this flat addressing mode.[/QUOTE]

Make of that what you will... if this were the CPU model, as long as the OS and any other running things could all fit comfortably in the near-RAM 8 or 16GB, then the cache mode would probably work okay. If Intel is working on compiler options it seems like it might be somewhat easy to specify... assuming that makes it into more than just Fortran compilers right away. :smile:

It might be more comforting to code it so the critical bits are directly addressed to the MCDRAM? Why leave it to chance I suppose?

In the coprocessor mode I suppose it depends on the implementation... it may not hurt to have a single stick of DDR4 on there. I don't know if they'd have the option of running RAM-less and just using the built in stuff. I mean, I don't see why it couldn't... 8 or 16 GB is a lot for some apps (and not nearly enough for others).

I got a kick out of reading about them using these to build that large supercomputer for Los Alamos. Sounded like they might have upwards of 8000 of these plus another who knows how many Haswell cores, all adding up to 40+ petaflops. Holy smokes! I guess it depends on how many cores wind up in the final shipping products.

An interesting point to consider is that Intel has its own fab plants and Nvidia is dependent upon TSMC.
Thus, Nvidia's Maxell is a 28nm part and Intel's Knights Landing is already down to 14nm !!!

[QUOTE=ewmayer;399167]
Hoping that there will be no issues installing a CoPro one of these in the PCI3 slot of my MSI Haswell mobo[/QUOTE]

Possible but unlikely. The current Kights corner boards require a mobo with a feature called "64 bit BAR support".
This spec is usually found only on Xeon mobos. However, there is an ASRock 99M Extreme4 board that handles it.

Another barrier to entry is that the ONLY way to compile code to run on these chips is via Intel's own compilers.
I'd be willing to bet that won't change soon.

We can expect a jump when Nvidia finally starts using the 16nm process. I suspect Pascal could be pretty impressive.

[QUOTE=tServo;399752]The current Kights corner boards require a mobo with a feature called "64 bit BAR support".
This spec is usually found only on Xeon mobos. However, there is an ASRock 99M Extreme4 board that handles it.

Another barrier to entry is that the ONLY way to compile code to run on these chips is via Intel's own compilers.
I'd be willing to bet that won't change soon.[/QUOTE]

Sounds like a game of "find a GIMPSer who gets to play with one of these at work and whose employer is willing to look the other way as we cajole said GIMPSer into letting us remotely beat the crap out of it" being set up.

I've been reading about these as well.

I'd be keen to try one.

As for prices, here's the current generation.
[url]http://ark.intel.com/products/codename/57721/Knights-Corner[/url]

Not much change from $2k for base models.

To me the big speed plus is the mucho memory bandwidth.

I've done enough mucking around now, to say that current multi core CPUs are incredibly memory starved.

What's the core limit for prime95?

-- Craig