Yes, this is what genuinely puzzles me the most. I get that mainframes have their place for highly stable IO-heavy processing, but I've never heard anyone claim that they are particularly effective for compute-bound workloads. Especially when machine learning does not require much robustness - you just reboot the cluster nodes when they crash (and they will, because GPU drivers are crap).
Re:IBM z14 mainframe
by lostboys67 on Monday July 17, 2017 @09:49AM
(#14788444)
Yes I thought it was commodity systems with dual xenon's and a MFT of memory and 24-36 TB DASD per node and for networking your choice of (infiniband / 10g/ 40g) for hadoop style HPC
you can of course add GPU's there are some workstation boards specifically designed for this
Re:IBM z14 mainframe
by microtherion on Monday July 17, 2017 @10:36AM
(#14788823)
> ok what ML platform would one run on it?
Not sure, but it will definitely be called "Watson"
It more about the fact that tons of customer data and transaction history are already on the platform. Why move the data when you can do the science locally.
Or does it only apply to companies that doesn't want its data to get out to the cloud?