Toasty
245 watts! All these setups are liquid cooled?
Intel's latest Xeon Phi processors for high-performance parallel computer systems are now, finally, shipping in volume. The 14nm chips, which feature eight billion transistors, were already in the hands of Cray, Sandia National Laboratory in the US and a few other boffinry types. Today, our sister site The Next Platform …
Hey I'm not on the same level as a lot of you lads so can one of you more experienced folks explain to me why you would pick these over an nVidia part ?
The P100 reported on earlier does 5.3 teraflops DP, these do considerably less at 3.46 for the top-end part. I know there's more to a part than that but talking purely metrics, why would these be chosen over the green team ?
This is way WAY over my head, I actually no next to nothing about co-processors, but this has me curious
I'm no expert on this either, but I can give a few shots in the dark why someone would choose Intel.
The coming Phi CPU will however fit into the next gen (Skylake-E) LGA 3647 sockets, and it is just a specialized x86 Xeon with extra cores. It can boot the OS and be used as an ordinary x86 CPU. So it's not a co-processor or an add-on card. If you are ordering a server with only a Phi CPU you can deduct the price of the cheapest possible Xeon (2603v4, $213). The top end Phi is only a 1.5GHz product, so single thread performance will be weak.
The 5.3 TFLOP P100 requires a mezzanine connector which I've only seen in blades. Perhaps the density optimized servers is the target market for it. Of course the more common PCIe x16 card with 4.7 TFLOPs still beats Intel in raw numbers. Bear in mind that both Intel and Nvidia numbers are theoretical maximums, and Nvidia has likely optimized their cards for CUDA, not OpenCL nor OpenACC.
The Phi has its own RAM (slower and smaller than in P100) but it has a much faster connection to the system memory, whereas P100 uses the PCIe bus which is several times slower. Depends on the computing scenario I guess. Also, if you are parallelizing beyond a single P100 or Phi processor, the bandwidth and latency between the computing units is much better between multiple Phi's than with P100's through PCIe.
The Phi processor is also "shipping in volume" (per article), but the P100 will be available in Q4. Nvidia pricing is also unknown at this point.
Consider a simulation of a Sterling interest rate swap, with Brexit (with the expected Sovereign credit rating downgrade from AAA) in 40% of scenarios the trade wouldn’t last the year (downgrade trigger) before being unwound.
A GPGPU kernel would have 40% of the cores waiting while all the non-brexit scenarios calculated risk through to maturity because of hardware threading; whereas a Xeon Phi would schedule work to all threads for the 60% of cases that don’t terminate early. NVidia is much better for physics, while Xeon Phi is better for behaviour.