Intels Xeon Phi looks to demote GPGPU

[the54thvoid]

An article popped up at S|A that announced Intel's former Knights Corner (Larabee) project was now in fruition of sorts in Xeon Phi cards. Due to the coding simplicity of the architecture it may do less work than a gpgpu but it is very much easier to code for. As such CD at S|A said it was the death knell for Nvidia's HPC dominance (and kyboshing AMD too).

I thought it was a CD rant and paid no attention.

But i just read a second article there from Anandtech and it kinda summarises the same thing. In other words, the raw power of gpu compute is no match for the coding simplicity of the Intel coprocessor Xeon Phi (which is no slouch).

This could dent Nvidias market quite a bit.

S|A article:
http://semiaccurate.com/2012/11/13/what-will-intel-xeon-phi-do-to-the-gpgpu-market/
Anandtech article:
http://www.anandtech.com/show/6451/the-xeon-phi-at-work-at-tacc

 

[FordGT90Concept]

Yeah, I'd believe it. x86 is far more flexible than CUDA and Streams. It also has the advantage of better error checking and that's very important for scientific computations. There's a lot of scenarios where GPGPU couldn't be used because of error checking that Xeon Phi can do.

There may be a few applications where GPGPU will still be used but I think the TOP500 will be flooded by AMD CPUs (because they're cheap) + Xeon Phi cards in the coming years.

 

[Cortex]

Hopefully compute Maxwell (Tesla) will be a homogenous GPU - all shaders can execute 2 prim/clk FP32 or FP64 (or 2 SP or 1 DP). I think intel cannot match that kind of performance (up to 4xDPof GK110 [or 2x] - 2880 FP32 SP and 2880/3 FP64 SP), even with smaller production process. Also a special GPU only for compute without TMUs (GPU version with TMU for visualizations).

 

[Protagonist]

Hopefully compute Maxwell (Tesla) will be a homogenous GPU - all shaders can execute 2 prim/clk FP32 or FP64 (or 2 SP or 1 DP). I think intel cannot match that kind of performance (up to 4xDPof GK110 [or 2x] - 2880 FP32 SP and 2880/3 FP64 SP), even with smaller production process. Also a special GPU only for compute without TMUs (GPU version with TMU for visualizations).

I don't think Intel will sit back with only the current line of Xeon Phi. They will offer something even more powerful at that time to compete.

Honestly I think things will go more as the OP has stated. Intel Killing them with simplicity,...

 

[FordGT90Concept]

Flexibility would probably be the better word. There's not much an x86 processor can't do in reasonable time. It's what they've been largely using for the past two-three decades.

 

[Steevo]

I want to see a F@H benchmark compared to new GPU cards.

In Intels own words they only expect 2-2.5X increase in performance. Its just this can happen in standard CPU tasks instead of GPGPU code.

 

[FordGT90Concept]

F@H is one of the rare instances where x86 will not be faster because there's very little logic and it is almost entirely composed of physics calculations (formula). The only benefit F@H would see from Xeon Phi is a reduction in errors.

F@H was always intended to be distributed computed though. Xeon Phi will see virtually no deployment to home computers so it makes no sense for them to target them anyway. GPUs are far more common in desktop computers than co-processors and that will remain the case for the forseeable future.

 

[LAN_deRf_HA]

because they're cheap

Don't they make those decisions based on factors other than initial cost? One look at the power consumption and they'd figure out Intel is more cost effective.

 

[FordGT90Concept]

Like most of us, it likely comes down to preference/who offers the better deal.