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

[Hugh]

replying to myself, I learnt what the specific requirement is. The motherboard and BIOS need to support I/O adressing mapped over 4GB (which feels weird and obscure. I guess systems with support for IOMMU / Vt-d are safe, though it is probably not a requirement)

This is where I found a fairly clear explanation: http://www.pugetsystems.com/blog/201...-Xeon-Phi-490/
It sounds as if it is purely a matter of BIOS support (assuming that you have a reasonable processor, not one crippled like the Atom).

[grok]

I thought it should run in any mediocre PC,

replying to myself, I learnt what the specific requirement is. The motherboard and BIOS need to support I/O adressing mapped over 4GB (which feels weird and obscure. I guess systems with support for IOMMU / Vt-d are safe, though it is probably not a requirement)

[grok]

I think not. It should use SSE, AVX which are not available on this Xeon Phi.

[torsionbar28]

I'm not a developer so pardon my question if it sounds silly:

Would / could this Xeon Phi be useful for video transcoding? Like with Handbrake? Handbrake does a nice job of natively loading up all the available cores in a system when transcoding video, so its authors have written it to be well parallelized already.

[grok]

There will soon be ARMv8 server chips with 16 cores and more, too. Possibly made of several ~quad core systems isolated but all on a very fast on-die back bone.
If you master a KNC Xeon Phi you will have some nice learned lessons and experience to use such systems

[grok]

[*]I would need to buy a new workstation, multiplying the cost by an order of magnitude

Thanks for the informative and constructive comments.

I thought it should run in any mediocre PC, e.g. in my current one I'd put it on the PCIe 16x slot, use an old PCI graphics card (I've used it before, good for browsing and video playback..)
Screw a couple steel bars to the open ATX case and attach many 80mm to 120mm fans. It would be ugly but it'd work.

If you feel you need a workstation (i.e. new desktop PC) you might get a C202 chipset motherboard, Pentium or i3, 500W PSU and 8GB ECC : that is cheap. If I don't miss something.
A steel, if that is for Computer Science academia (you get a use case and access to the compiler) and you don't mind the limitations (it would be awesome for experimenting with concurrent languages and concurrent functional languages, but can you get them to run or build?)

It is made from short pipeline CPUs and so it's probably good with branching and pointer-chasing code. It can most probably be used for classes of problems that won't make sense on a Tesla, Titan, geforce etc.
Hacking on a Kaveri and then a Carrizo is another way. (and pairing Carrizo with Tonga or Fiji)
Most generic will be a dual socket Haswell-EP, now that's pricy but that's what you want if you'll run "regular" albeit parallel code.. Next-gen Xeon Phi will be about similar, either lone on a motherboard or paired with a recent Xeon (not sure if Haswell-EP or Haswell-EX) on a dual socket motherboard.. $$$

It's a trap. You'll have to decide to be trapped either by Intel, AMD, nvidia or even nvidia + IBM POWER (the simliar thing to Xeon + KNL Xeon Phi or AMD HSA) ; possibly Qualcomm, Samsung and the like after that.

[chuckleb]

You can usually get coprocessor cards in a couple of configurations, with or without a fan. The ones with a fan can be used in a workstation and are just like a video card, etc. The ones without the fan are designed to be run in servers that have a lot of fans pulling in air from the front and shoving it through the case. They are often 1U or 2U, so the air has nowhere to go but over all the components. This is similar to the fact that most 1U CPU heatsinks don't have fans attached because the case airflow is sufficient to cool the fins.

When you start to get into 4U or desktop cases, you 1) don't have enough fans moving air, and 2) have such large volumes inside that you don't get adequate localized cooling. This is why you need CPU cooling fans.

One thing that I was thinking of is to adapt an impeller fan or something to force air through the back end of the card. That would give it enough cooling and you can get fans for $5-10. You'd need to make a bracket or something though.

In the points about the obsolesce of the card, I think there is still merit in the card. There are thousands of these cards installed and people are not ripping them out because a new technology came out, they will use them as long as they can. It's a good price to learn how to use coprocessors, but you could also do that with a cheaper Nvidia card with less RAM and cores. It'll be a good learning environment if you wanted to go that route.

I think the most valid cons against the promotion are the hardware requirements to run the card and compiler issues.

I was really excited when I read the original article. The comments were very useful in convincing me that the deal isn't for me. I was just going to dabble.

• this card is passivly cooled, requiring appropriate external airflow (that I don't understand)
• the card with a fan seems to be harder to source cheaply
• I would need to buy a new workstation, multiplying the cost by an order of magnitude
• I would have to buy a compiler license
• It's not clear how fast this would run my integer code (with pointer chasing)
• this card and architecture are obsolescent

Thanks for the informative and constructive comments.

[Hugh]

Thanks

I was really excited when I read the original article. The comments were very useful in convincing me that the deal isn't for me. I was just going to dabble.

• this card is passivly cooled, requiring appropriate external airflow (that I don't understand)
• the card with a fan seems to be harder to source cheaply
• I would need to buy a new workstation, multiplying the cost by an order of magnitude
• I would have to buy a compiler license
• It's not clear how fast this would run my integer code (with pointer chasing)
• this card and architecture are obsolescent

Thanks for the informative and constructive comments.

[grok]

The next one will be much more PC-like : cores based on Atom Baytrail architecture, vector instruction set from Skylake, able to work fully on his own (not a card that needs a master CPU anymore). It will be more amenable to running 3rd party code.

But you'll probably need a linux kernel with an adequate scheduler, and use OpenCL software or software intended for a cluster (that would probably work very well) if not targetting the next-gen Xeon Phi itself.
You'd be looking at a $5K computer methinks (used to be the cost of a high end PC lol) plus possibly network upgrades.

For current Xeon Phi best use is probably research, self-teaching even if at worst it boils down to i586 integer code only (or you might get to use Intel compilers and stuff anyway)

[Luke]

This card could possibly used as a remote video render farm.

OK, someone said each core is similar to an original Pentium boosted to about 1.1 GHZ and with the ability to run 4 threads added to it, with 57 of these procs. The board has its own IP address, can run generic x86 code, and can be ssh'ed into.

Suppose you get a board that can simply fire up the card, and use it as a video render farm for Kdenlive or another video editor? You would edit on the normal machine but would generate a script for rendering, then pass that script to the Xeon Phi. Even if each core was only as good as a P5 at 1.1 GHZ, with 57 of them it would be about equal to a 13 physical core processor at 4.5 GHZ in the special case of video editing in H264. If the 4-thread hyperthreading is only as good as the AMD Bulldozer twinned module approach (about 1.3* for 2 cores), you'd get the equivalent of 2.6*13 cores, or the equivalent of 4 AMD Bulldozer 8 core procs plus one Bulldozer six core proc on a single board, for the cost of a pair of Bulldozer 6 cores. The 1080p video I use won't fully load AMD Bulldozer in Kdenlive, but I suspect 4K video would be another story entirely. FFMPEG/libx264 itself scales just fine to an arbitrarily large number of threads it seems.

Hell, that render farm could be set up for a whole crew of filmmakers to access over a high-speed network, one box if nothing else could handle 8 video editing jobs at once based on my results with Bulldozer.