Intel Phi vs Xeon cpu in the future.

[Kippa]

I was wondering so far as work stations go, could Intel Phi be a game changer so far as workstations go? Lets say you were to build a 3d rendering rig and rather than go with a dual socket workstation with expensive xeon processors, you instead go for a single cpu based xeon or even an i7 rig and pop in an Intel Phi to do take the workload.

Could Intel Phi change the market and make it less viable to go for dual socket xeons, or even xeons at all?

I know that most software isn't yet coded to take advantage of the Intel Phi or gpu processing in general, I am just looking towards the future.

 

[3DVagabond]

I also had this thought. It seems like it should perfect for render farms.

 

[Kippa]

I read that you could use 4 of Phi in one computer. You could go very cheap on the cpu and invest in Phi instead. Interesting times ahead.

 

[Pottuvoi]

It certainly should be a good choice for some tasks, but it does need own code to be fast.

 

[mv2devnull]

It certainly should be a good choice for some tasks, but it does need own code to be fast.

Depends. If your code is written so that the compiler does auto-vectorize using whatever instructions (SSE, AVX, Phi) the hardware has, then a simple recompile with Phi as target will give you good binary.

A code that depends on some specific hardware (explicit intrinsic calls or Cuda, for example) obviously has to be rewritten.


Note: "More cores vs Phi cores" is essentially a CPU-discussion. "Phi vs GPGPU" is closer to graphics.

 

[3DVagabond]

Depends. If your code is written so that the compiler does auto-vectorize using whatever instructions (SSE, AVX, Phi) the hardware has, then a simple recompile with Phi as target will give you good binary.

A code that depends on some specific hardware (explicit intrinsic calls or Cuda, for example) obviously has to be rewritten.


Note: "More cores vs Phi cores" is essentially a CPU-discussion. "Phi vs GPGPU" is closer to graphics.

How about Phi for ray tracing?

 

[BrightCandle]

The current issue I see with Phi is the board runs its own Linux operating system. You can talk to the Phi over a high speed network communication but its not really integrates to the platform beyond that. I want coprocessors so that OS can start to use these CPUs and choose between them depending on some hints from the programmer. Right now you can't just run a Windows program, with some special threads, that would utilise a Phi and I think that is an error on In tels part.

 

[Pottuvoi]

The current issue I see with Phi is the board runs its own Linux operating system. You can talk to the Phi over a high speed network communication but its not really integrates to the platform beyond that. I want coprocessors so that OS can start to use these CPUs and choose between them depending on some hints from the programmer. Right now you can't just run a Windows program, with some special threads, that would utilise a Phi and I think that is an error on In tels part.

Phi doesn't support SSE, so you would need compile for it even if you just want to use it just as x86 processor.

What you can do is use them as processor nodes which are called from your windows program, there really shouldn't be problems.

I read that you could use 4 of Phi in one computer. You could go very cheap on the cpu and invest in Phi instead. Interesting times ahead.

According to Intel the current limit is 8 boards per computer, should give a nice boost.
http://www.intel.com/content/www/us...h-performance-xeon-phi-coprocessor-brief.html

 

[Kippa]

Woah 8 of them on one cpu, that would make one hell of a render farm! It would cost a fair but would be worth it.

So far as coding goes for software developers Intel Phi uses OpenCl yes? So for the programmers they don't need one code for Tesla, another code for AMD gpu and another code for Intel Phi? Is that right?

 

[BrightCandle]

So far as coding goes for software developers Intel Phi uses OpenCl yes? So for the programmers they don't need one code for Tesla, another code for AMD gpu and another code for Intel Phi? Is that right?

No it doesn't use OpenCL. A cpu based OpenCL could be used within the Phi's Linux Operating system, and they may well have done that, but the more direct option of writing a standard program that runs on X86 is the intended route for development.

So rather than having to learn openCL and contort a program into its style of development you can write a many threaded application that can utilise the Phi's cores instead. You can either run it entirely on the Phi or share some work with the host machines processor and talk via network communications between the two. In essence its a mini PC running inside of your current computer using the same instruction set minus some modern instructions.

Which is why its a shame its not supported directly in the OS because in theory if a fast core is in use a task may be better done on a slow Phi core.