Xeon Phi

Big question I have is does it run x86 or is it like the NV/ATI and it needs to run on special code.

It be quite nice if you can just use these cards and insta add "normal" CPU power.
 
Hmmmm..... interesting concept. I would be nice to see some more details on it. Especially what Kendrak brought up in that it runs standard x86/x86_64 instructions or specialized ones like most GPUs.
 
So if it runs x86..... is this the future of (mega) folding?
 
Kendrak said:
So if it runs x86..... is this the future of (mega) folding?
Click to expand...

Tis something to ask Stanford...given the rather pitiful performance in PPD/Watt of GPUs relative to 4p, these critters will have their work cut out for them. That said the cards don't look to be the 12" long monster GPU cards.

NOTE this part, you need Xeons to use them it looks like. Intel wants to sell more LGA2011 workstation boards/procs it seems:

XPhiIG_575px.jpg


////Skripka cringes at the use of "synergistically"
 
Have we come full circle? My 486 (or was it the 286) had a math co-proc. :p
 
I think the math co-processors were 8087, 80187, 80287, 80387 and 80487. Starting with the 5 series, all CPUs had the x87 architecture included in the x86 die.

wikipedia disagrees with my memory (but I'm old) and talks of a 80587 as well, but I don't know anyone who used one. I installed a 80487 into my 80486 along with replacing all the L2 cache chips doubling the cache with faster cache. It started life as a SX-25, I OCed it to 33, upgraded to DX-50 overclocked to 66 and finally dropped in a Pentium overdrive. The first computer I did major upgrades and overclocking. I remember it well.

P.S. my 8088 had an 8087 as well.... never owned a 8080, but I used several. To put it in persepective my first HDD was a Full height 5.25 (what you now would call double height, 2x5.25 bays) and was a staggeringly large 5 MB.
 
Last edited:
This is a pretty different product from a GPU. It runs it's own operating system and simply uses the mobo for I/O. If smp will run on these I don't think it could use multiple cards at once unless intel provides some magic wrapper that makes that transparent to the client. The memory bandwidth should be pretty insane with GDDR5 and a 512 or 256-bit bus. From what I can tell, a quad opteron system can hit around 500Gflops under benchmarking software so an 8101 crunching an a 1Tflop system would be approaching 1.5Mppd.
 
Zink said:
This is a pretty different product from a GPU. It runs it's own operating system and simply uses the mobo for I/O. If smp will run on these I don't think it could use multiple cards at once unless intel provides some magic wrapper that makes that transparent to the client. The memory bandwidth should be pretty insane with GDDR5 and a 512 or 256-bit bus. From what I can tell, a quad opteron system can hit around 500Gflops under benchmarking software so an 8101 crunching an a 1Tflop system would be approaching 1.5Mppd.
Click to expand...

Presuming you can even stick the Phi in any old motherboard. If I had money to bet, I'd wager this card requires a Xeon with 2xQPI. Just my own personal hunch though.
 
Skripka said:
Presuming you can even stick the Phi in any old motherboard. If I had money to bet, I'd wager this card requires a Xeon with 2xQPI. Just my own personal hunch though.
Click to expand...

Well, there isn't exactly a shortage of systems like that on the team :p
 
Zink said:
This is a pretty different product from a GPU. It runs it's own operating system and simply uses the mobo for I/O. If smp will run on these I don't think it could use multiple cards at once unless intel provides some magic wrapper that makes that transparent to the client. The memory bandwidth should be pretty insane with GDDR5 and a 512 or 256-bit bus. From what I can tell, a quad opteron system can hit around 500Gflops under benchmarking software so an 8101 crunching an a 1Tflop system would be approaching 1.5Mppd.
Click to expand...

I'm not sure if this thing can run an OS by itself. From the slides, it looks like it's dependent on a Xeon CPU. I'm not even sure if you can use this in a system other than an Intel Xeon one... They sure would want to control and maintain their Xeon habitat, so it just makes sense from a corporate pov...
 
Yeah and i bet it will only run with the latest and greatest SB based xeon's, us poor soul's with lga 1366 xeon will get left out yet again:(
 
Nah it will probably just take any old pcie x16 slot.

The pervious version of this was knights corner, 32 cores, this should be ~64cores @22nm, 300W. Let's see what 300W worth of cpu power can do!

EDIT: OK< If I am reading this anand article right... it can probably fold now with the existing client! Apparently the card runs's it's own distro of linux, which would be an x86 based distro, and you could just put the existing client on and run it..
 
Last edited:
I'd agree it'll probably use pci-e x16 slots, I think the issue here for folders will be the cost. Intel is targeting this squarely at the enterprise level and that usually means big bucks.
 
If it can replace two 4p systems then many folders have the bucks to spend. The top end card could replace $5000 in hardware and save $1000 a year in electricity depending on where you live. They aren't much ahead of AMD's new FirePro W9000 in double precision FLOPS so unless they are aiming for a niche market they have to keep prices competitive. The last gen AMD top end card cost $3500 and big organisations can hire the right people to code for GPUs if they remain cheaper than Intel's cards.
 
Last edited:
Zink said:
If it can replace two 4p systems then many folders have the bucks to spend. The top end card could replace $5000 in hardware and save $1000 a year in electricity depending on where you live. They aren't much ahead of AMD's new FirePro W9000 in double precision FLOPS so unless they are aiming for a niche market they have to keep prices competitive. The last gen AMD top end card cost $3500 and big organisations can hire the right people to code for GPUs if they remain cheaper than Intel's cards.
Click to expand...

Replacing 2x 4p systems would require > 100 cores per card, most likely. Even if you assume intel will have the better performing architecture, these will likely be lower clock speeds for power reasons. They're only quoting "50+" cores at this point, so if we assume core for core equality between one of these and a 4p, it basically is a 4p on a card. Then you factor in the theorized requirement to be running a xeon system, possibly a 2p one, and you basically come out with... the cost of an AMD 4p. There very well may be power saving benefits, but I don't see it being >= 2x 4p.
 
I'm going to go out on a limb here and say that Intel will have their own VM environment that runs on the CPU but accesses Phi using VT-d. Which is why it's Xeon chips only.
 
Yeah I don't know how that works, they claim 1 TFLOPS real world perfromance but with only 50 or 64 cores. An i7 or G34 chip only hits approximately maybe about 125 GFLOPS so I don't see how they hit that performance with a simplified CPU architecture and low power cores.
For 1.5Mppd I was basing it only on doing twice the FLOPS of a high end 4p. I don't think it's a valid comparison and I don't see how they have that much performance in 300W of pentium cores.
 
When?

how much?

how many points?

who is buying?

I want more pics!
 
I'm sure we'll see one atleast within a year or two, or atleast someone burning one in on the web
 
The reason it is exciting is because of F@H client support. Stanford gives huge bonuses for bigadv WUs completed on the SMP client and it seems possible that this chip could run those WUs without modifying the software at all. For whatever reason, Stanford would rather have a bunch of powerful x86 hardware instead of more GPUs which is why they give such high ppd to CPU rigs. It would probably be technically possible to solve the same problems faster on GPUs but it would take more software development that Stanford doesn't have the resources to do.
An HD 7970 gets about 10kppd while a 4p AMD rig with 48 cores gets around 500 - 700kppd. If this card is able to deliver twice the performance of a 4p server then it would be something like 1.3Mppd.
 
Nathan_P said:
Not even in production yet, which is good as it allow me time to get my Z9PE-D8 WS and some cpu's to go with it
Click to expand...

They already built supercomputer using this HW.

from Anandtech article:
Intel hasn’t put any hard date on availability but they have said they expect Xeon Phi co-processors to go into full production later this year, and in the meantime Intel has already produced enough co-processors to build a MIC based supercomputer that’s ranked #150 on the new TOP 500 list. Given the typical gap between volume production and when a product is available for purchase it’s likely that Xeon Phi co-processors won’t be available until the end of the year – if not next year – but regardless the timing is such that Intel will be going up against NVIDIA’s GK110-based Tesla K20, which is similarly expected by the end of the year. Meanwhile given AMD’s HPC ambitions with GCN we’re also not ready to rule them out, so all 3 parties may have major compute products out by the start of 2013.
Click to expand...

so it might be available during Xmas :)
 
crazjayz said:
I'm going to go out on a limb here and say that Intel will have their own VM environment that runs on the CPU but accesses Phi using VT-d. Which is why it's Xeon chips only.
Click to expand...

AnandTech said:
Meanwhile on the software side of things in an interesting move Intel is going to be equipping Xeon Phi co-processors with their own OS, in effect making them stand-alone computers (despite the co-processor designation) and significantly deviating from what we&#8217;ve seen on similar products (i.e. Tesla). Xeon Phis will be independently running an embedded form of Linux, which Intel has said will be of particular benefit for cluster users. Drivers of course will still be necessary for a host device to interface with the co-processor, with the implication being that these drivers will be fairly thin and simple since the co-processor itself is already running a full OS.
Click to expand...

I bet these cards will run on any platform. No reason they can't.
 
If each card has a linux OS and if the client and cores work on it you have just added extra SMP clients to a rig without taking up more space, Heat would be a concern but imagine putting some extra SMP grunt onto a 4p rig
 
So wait... does that mean I can put one of these on my G34 4p? Cause that would be fricking epic... assuming they work as they said they will... I might have to upgrade my PSU in the 4p afterall.
 
Would be even more epic if you have the version of the board with more than 1 PCIe slot, until they have been released though we are all merely speculating
 
You must log in or register to reply here.
Back
Top