-
Phi
-
[Walter Soyka] “Just when you thought threads on general computing trends were safe!”
Must. Not. Post……………………………………………………………………………………………………………………
_______________________
David Lawrence
art~media~design~research
propaganda.com
publicmattersgroup.com
facebook.com/dlawrence
twitter.com/dhl -
“… it is a pretty standard accelerator card for HPC and generic number crunching. It looks like a cluster of x86 servers that are on a TCP/IP network, and takes almost no programming expertise to port to. Optimization is very similar to a standard x86 CPU, and you can use the Intel tools that are very well established in the HPC world.”
basicly takes x86 code and after recompilation in intel software its able to run on many more parallel x86 CPU cores
more to read: https://www.theregister.co.uk/2012/11/12/intel_xeon_phi_coprocessor_launch/
https://newsroom.intel.com/community/intel_newsroom/blog/2012/11/12/intel-delivers-new-architecture-for-discovery-with-intel-xeon-phi-coprocessors?cid=rss-258152-c1-278335
https://download.intel.com/newsroom/kits/xeon/phi/pdfs/Intel-Xeon-Phi_Factsheet.pdf
https://download.intel.com/newsroom/kits/xeon/phi/pdfs/Intel-Xeon-Phi-Coprocessor_ProductBrief.pdf -
[Michael Gissing] “So we basically agree then. In the short term big players that have custom software can make best use. We can all potentially benefit but my question was whether software companies would see sense (ie $$s) in re writing software for this sort of co-processor over GPUs which they are currently doing.
If they do then we can all benefit. However when the trend is to cheap NLE software in mobile devices, I see perhaps less use for the average editor than say CUDA or OpenCL.”
None of this could be handed to grand central dispatch?
-
[Richard Herd] “To hell with Persistent In and out points.”
Did you know persistent in and out points spelled backwards is stniop tuo dna ni tnetsisrep?
-
[Jakub Vomacka] “basicly takes x86 code and after recompilation in intel software its able to run on many more parallel x86 CPU cores”
Thanks Jakub.
In my limited understand of the inner workings of a computer, doesn’t this mean that, with a fairly simple exchange this type of processing is simply added on? It goes from the bigger Xeons to the smaller 60(!) cores?
Can’t something like Grand Central Dispatch manage the traffic efficiently?
Or am I dreaming too much?
-
[Jeremy Garchow] “stniop tuo dna ni tnetsisrep”
Ah ha! We now know what he meant in Temple of Doom while pulling the heart out of the poor victim.
https://www.youtube.com/watch?v=mk2E1CoGe98
-
I see your youtube and raise you a youtube
https://www.youtube.com/watch?v=-_bJx6NX9f8
Some contents or functionalities here are not available due to your cookie preferences!This happens because the functionality/content marked as “Google Youtube” uses cookies that you choosed to keep disabled. In order to view this content or use this functionality, please enable cookies: click here to open your cookie preferences.
-
[Jeremy Garchow] “In my limited understand of the inner workings of a computer, doesn’t this mean that, with a fairly simple exchange this type of processing is simply added on? It goes from the bigger Xeons to the smaller 60(!) cores?”
Well, memory access is still an issue — it’s not clear to me that Phi has direct access to main system memory (and even if it does, it will be relatively slow over the PCIe bus). Hearing that it’s addressable via TCP/IP suggests it runs almost as a separate computer within its host.
[Jeremy Garchow] “Can’t something like Grand Central Dispatch manage the traffic efficiently?”
I don’t think this would work, as GCD currently stands. Wikipedia says GCD is built for symmetric multiprocessing [link], which means multiple identical CPUs (or cores) attached to the same pool of memory.
I could be mistaken, but it sounds more like Phi cores are different from mainboard Xeon E5s, and it sounds like the Phi system has separate memory. That’s why I had wondered about OpenCL, which allows execution of parallel tasks on different hardware.
This is all still pretty speculative on my part, so hopefully as more details emerge on Phi we’ll get a clearer picture of what’s possible (assuming the next Mac Pro has a PCIe slot to stick this bad boy in).
Walter Soyka
Principal & Designer at Keen Live
Motion Graphics, Widescreen Events, Presentation Design, and Consulting
RenderBreak Blog – What I’m thinking when my workstation’s thinking
Creative Cow Forum Host: Live & Stage Events -
[Walter Soyka] “Well, memory access is still an issue — it’s not clear to me that Phi has direct access to main system memory (and even if it does, it will be relatively slow over the PCIe bus). Hearing that it’s addressable via TCP/IP suggests it runs almost as a separate computer within its host.”
It’s still faster than not having it, right?
[Walter Soyka] “I don’t think this would work, as GCD currently stands. Wikipedia says GCD is built for symmetric multiprocessing [link], which means multiple identical CPUs (or cores) attached to the same pool of memory. “
As it currently stands, sure. What I don’t know is that it seems this truly is an extension of the Xeon code, if that’s true, then perhaps GCD could be updated fairly easily since it already seems to be addressing intel’s code.
[Walter Soyka] “This is all still pretty speculative on my part, so hopefully as more details emerge on Phi we’ll get a clearer picture of what’s possible (assuming the next Mac Pro has a PCIe slot to stick this bad boy in).”
We’ll see, but that’s why I found the release date to these and Tim Cook’s proposed “something special” date to be a convenient coincidence.
Who knows.
-
[Jeremy Garchow] “It’s still faster than not having it, right?”
Well, sure, but I don’t think popping a Phi in your computer instantly supercharges it as if the main computer were a 64-core monster.
This would certainly be awesome for throughput-light, computation-heavy work like 3D rendering where you can load up the onboard RAM and let the CPUs have at it, and it should fare at least as well as GPUs at throughput-heavy work.
[Jeremy Garchow] “As it currently stands, sure. What I don’t know is that it seems this truly is an extension of the Xeon code, if that’s true, then perhaps GCD could be updated fairly easily since it already seems to be addressing intel’s code.”
It’s not just an issue of different instruction sets; it’s an issue of physically different architectures. I don’t think you can’t arbitrarily shuffle instructions around from the mainboard to the Phi without also shuffling the data it needs to operate on from main memory to Phi memory. If I’m understanding all of this correctly (a decent-sized if!), that’s creating a lot of overhead just to move thread execution around.
Phi sounds more like clustering, which GCD is not built for. Apple’s Advanced Computation Group used to be active in clustering research (XGrid), but their interest seems to have waned in recent years.
But of course I could be wrong.
[Jeremy Garchow] “We’ll see, but that’s why I found the release date to these and Tim Cook’s proposed “something special” date to be a convenient coincidence.”
This would be very cool, but I’m not sure I see the business case for Apple pushing the envelope in performance computing anymore. I’d love to be wrong.
Walter Soyka
Principal & Designer at Keen Live
Motion Graphics, Widescreen Events, Presentation Design, and Consulting
RenderBreak Blog – What I’m thinking when my workstation’s thinking
Creative Cow Forum Host: Live & Stage Events
Reply to this Discussion! Login or Sign Up