Intel Announces Xeon Phi Co-Processors | Hacker News Reader

Intel Announces Xeon Phi Co-Processors

anandtech.com

76 pointsmrb43 comments

Note: the instruction set and open source software stack are available:

http://software.intel.com/en-us/blogs/2012/06/05/knights-cor...

http://software.intel.com/en-us/forums/showthread.php?t=1054...

So this means that each core is running its own version of Red Hat or some other Linux? Does that mean that I can install my own software on those ah, tiny machines?

Someone may probably get mad or think I am stupid for asking this, but can I install node.js on each core?

It's probably going to be a small RTOS in each core

Or like it's done in PS3, where you have one core scheduling the others

"but can I install node.js on each core?" maybe, but that's really not the idea

ok whats the idea then, how am I supposed to program the tiny machines/cores?

Have a look at OpenCL, I think you will use OpenCL or a similar framework to code for this device.

Have a look at http://ispc.github.com/ -

Intel SPMD Program Compiler

"An open-source compiler for high-performance SIMD programming on the CPU"

You can go the RTOS route, but each card looks like a 50-core, 200-thread machine with 8 GB of GDDR5 memory and can comfortably run a Linux or *BSD stack.

In fact, I'd love to see developers using them as their personal machines (Intel would probably have to make ATX motherboards with it) so that more and more software makes use of the more and more available threads on our own machines. An entry-level laptop these days looks like 8 processors to the OS and this number will only go up.

While these are x86 cores, and clearly capable of running desktop software, it wouldn't be a good experience as the primary CPU. The circuit size almost certainly means that these are in-order CPUs, probably with minimal superscalar capability, and likely a shallow pipeline. So scalar code is going to run terribly on them. Desktop usage is dominated by scalar CPU overhead.

That's precisely why it would be interesting - developers would have to work around these limitations. Number of cores will continue to go up and, if the software allows it, they can be simples ones.

With all due respect: yawn. People have been predicting the end of scalar code now for a decade, and it's no closer than it was. Show me a SIMD Javascript interpreter, or DOM layout engine, or SSL implementation, or compression algorithm, or heap allocator, or...

Won't happen. Out of order, heavily superscalar, deeply pipelined cores have won in the market because they're better for 90% of the problems you will see. But there are better architectures for the remaining 10% (and of those, "3D games", "Bitcoin mining" and "Password cracking" seem to make up about 90%), so you'll see lots of churn in this space as people try to figure out the best and most future-proof architeture.

For me, I like this a lot. The Knight's Corner architecture looks mostly identical to the AVX code you can run on Sandy/Ivy bridge right now, just with a few new features and a different SIMD width. From a programmer's perspective, it's much cleaner than the hidden-behind-multiply-abstracted-layers-above-proprietary-secret-hardware nonsense being peddled by AMD and NVIDIA.

But it won't run your desktop. Ever. Sorry.

I agree it won't run current applications (as in Eclipse, Word, Excel, Emacs) much better than current CPUs do. But I can, for instance, imagine a huge number of applications that cannot run well on current architectures that would benefit from having a huge number of relatively simple cores.

My notebook currently has two kinds of cores - 2 amd64 ones and I don't know how many really dumb GPU cores. It would be great if my Gmail experience were snappier when I'm not dragging windows between monitors.

The first multi-processor machine (a 4-processor Pentium Pro) I used didn't impress me by being fast (it was, but not that much). It impressed me by being much smoother under load than my previous computer. It doesn't matter how sure you are it won't happen, it eventually will.

I agree with almost everything you said, but I take issue with a couple of your examples. First, depending on the cipher suite, SSL can in fact get significant speedups from SIMD. Second, while compression is a pretty serial task, if you compress in blocks, then you can parallelize across those blocks; I know that's not SIMD, but it also doesn't require fancy processor cores to take advantage of the parallelism. A lot of server tasks are similar to that: parallel enough that they could benefit from having a bunch of somewhat wimpy but power-efficient cores thrown at them.

From my limited understanding: each card will appear as a 50 core linux box.

Each core has a set of 16 way vector insns, and there is a little api for synchronization.

Think of graphics cards, since they're the most commonplace example of coprocessors. You don't install a game on your graphics card; rather, the system gains the ability to offload graphics processing to the card, in order to improve how well the system as a whole runs the game.

Except in this case, you can install Linux on it.

Each card (not core) is running an embedded version of Linux. Binaries are almost all linked to busybox. You can mount NFS shares to share data.

The embedded systems need specially compiled binaries. With the Intel compiler, this is usually achieved by turning on a flag at compile time. So yes, you can install and run your own software on the nodes.

Seems like the tl;dr is: 50+ x86 cores with RAM and CPU soldered to PCIe boards.

In other words, a mix of CPU and GPU. A standard architecture, as opposed to GPU private architectures, many cores, and faster memory pipelines.

Now, where did I put that functional programming book..

Hmm, the Mac Pro has Xeon processors and PCIe slots. People were complaining just recently about the lack of Mac Pro updates.

I assume some of Apple's professional video tools, at least, are CPU-bound. They'd probably love to have an "add 50 more processors" option for Mac Pros.

The Phi cores are really slow unless you use the SIMD unit; the typical way to do that would be OpenCL, but you can run OpenCL code faster on your GPU.

Okay, I'm a little confused. What exactly is a Co-Prossessor? It looks like a graphics card. Will it replace the graphics cards or work alongside it? Does it only handle highly parallel tasks, with the processor handling serials tasks?

http://en.wikipedia.org/wiki/Coprocessor

Yes, the idea is that you will offload highly parallel tasks to it while the main CPU handles serial tasks.

While we're adding relevant Wikipedia links,

http://en.wikipedia.org/wiki/Heterogeneous_computing

If Xeon Phi is going to run linux out of the box, then that's Intel basically declaring that they want this card to be supported by the opensource community from day one, correct?

Correct me if I'm wrong, but it seems to me that the Xeon Phi is only going to be useful to open source developers early on, and then as libraries are built that make use of the Xeon Phi, application developers will be able to build on top of those libraries. So we're probably a long way off (and likely a few different shifts in co-processor technology) before a card like this is useful to the typical, end user.

Realistically Intel is going to provide the libraries like the OpenCL runtime.

Little low on technical detail. "Works synergistically with Intel(R) Xeon(R) Processor", honestly!

Interesting though: will it be developed into a graphics board, returning to Larabee? What will NVidia and AMD (ATI) do in response?

Looks like Intel is trying to attack the CUDA/Stream market. I'm curious to see how NVidia and ATI will respond.

s/ATI/AMD/

Nvidia has Project Denver in the making, based on ARMv8, the 64 bit architecture - custom CPU and Maxwell GPU.

Now, I'd love to see one with hardware-assisted transactional memory like the newest POWER machines. A standard HTM implementation for x64 ISA machines would be a huge leap forward.

The future "Haswell" chips are going to have support for transactional memory.

No news on these guys though. Give it time. In a few years these things will be getting up to speed in features/RAS and such.

Xeon works "synergistically" with Xeon Phi ;)

Any word on what the cost of this card will be?

The 80s called. They'd like their co-processors back!

I'm quite sure my Apple II didn't have 50+ cores.

That would be a funky looking apple...

You know that GPUs are coprocessors, right? This is analogous. The cores on here are small, power-efficient, and have a lot of SIMD.

Reply on news.ycombinator.com