[SamuelB]

Hello.

I'm looking in a new machine for 90% structural(non-linear+contacts) and 10% CFD. Models ranging from 50k to 1M dofs.

Limitations: No cluster. 12 HPC Licenses (may increase in the future). Operated with RSM.

Due to limited amount of licenses it needs to be a fast machine. I'm looking at 3cores+1gpu per job in a total of 3 jobs, or 2cores+1gpu for 4 jobs. 32GB per job.
A 2x8cores with lots of RAM is already present for big jobs.


CPU questions:
Significant improvement from sandy to ivy to haswell in Strutucral/CFD? Haswell has AVX2 and that may provide an increase from 0% to 50% in floating point. But haswell has half the memory bandwidth of Ivy,

What about AMD 6328(8C) and 6308(4C)?

GPU questions:
Are there benchmarks for different nvidia cards? (k10, k20, k5000)?
Do I need 2 Nvidia GPU? I have seen CUDA applications that the CUDA card is only unlocked with a companion card.
Are AMD card supported in 15? Are there benchmarks?
What about the Intel Phi cards for structural? How do they perform?

Options evaluated so far (RAM not included, GPUs are K5000):

Assuming I can use a single videocard for computing:
(My choice) 3x $2500 1GPU + 4C @ 3.6GHz Dell T3600 - E1620(sandy bridge).

3x $3200 1GPU + 4C @ 3.7GHz from HP. Same as above but with ivy bridge.
2x $7000 2GPU + 2x4C @ 3.5Ghz Dell T5610
1x $12000 3GPU + 2x8C @ 3.5Ghz.Dell Rack Workstation (Very nice option)

Side comments:

This 4C @ 3.5Ghz Haswell-Xeon with a GPU (if supported). This is my choice if a GPU could be added. Verifying.
This 4x4C @ 3.5Ghz 6308 Opterons with GPUs would be awesome for a larger deployment.

Thanks.