How complicated computing applied sciences change into industrial options is commonly unclear. In most situations, launch bulletins mark the primary time the general public hears about such merchandise. However, that isn’t at all times the case. For instance, the looks of IBM’s not too long ago launched Bayesian Optimization Accelerator will be tracked over different firm breakthrough improvements. Following that trial presents some insights into how what some individuals would possibly contemplate an obscure know-how can change into a commercially viable product.
What is Bayesian optimization?
Originated by the Rev. Thomas Bayes in 1763, the Bayes theorem determines the likelihood of an occasion primarily based on information of situations which may affect or be associated to the occasion. As famous in Stephen Meserve’s submit (hyperlink is above) on IBM’s new answer, Bayesian strategies are commonplace in arithmetic, however making use of normal options, similar to Monte Carlo search, to product design issues is commonly difficult or impractical.
With that in thoughts, IBM designed the Bayesian Optimization Accelerator to seek out optimum options for real-world design challenges in much less time and with fewer assets than different options. It can scale to orders of magnitude bigger in quite a lot of dimensions and sort out extremely complicated issues. Plus, IBM’s answer can decide design factors with a smaller variety of samples than different strategies require, delivering outcomes quicker and extra cost-effectively.
However, growing the Bayesian Optimization Accelerator as a industrial answer required the efforts of a number of IBM groups and enterprise items.
Hardware basis: IBM’s POWER9-based AC922
Introduced late in 2017, Power Systems AC922 was the primary industrial IBM server to make the most of the corporate’s new POWER9 processors, in addition to NVIDIA Tesla Volta 100 (V100) GPUs. The AC922 additionally integrated options developed for the Summit and Sierra supercomputers deployed on the U.S. Department of Energy’s Oak Ridge Laboratory. The Summit system additionally led the High500.org top-performing supercomputer lists from June 2018 till June 2020.
As famous in a weblog by Ron Gordon on the time of the launch, the “AC” designation within the AC922 stands for Accelerated Computing due to the efficiency of two POWER9 CPUs, I/O bandwidth and reminiscence bandwidth and as much as 4 NVIDIA GPUs. The AC922 is especially properly suited to Artificial Intelligence functions, together with machine studying and deep studying, utilizing Linux and frameworks like Torch and Caffe.
The AC922 continues to ship very good efficiency for AI workloads and gives a {hardware} basis for the brand new Bayesian Optimization Accelerator.
How IBM Research’s Bayesian optimization helps drive Power improvements
In June 2020, IBM mentioned advances achieved by a group of engineers in its High-Speed Bus Signal Integrity (HSB-SI) group. Implementing IBM Bayesian optimization software program, a machine-learning device developed by IBM Research, the group was capable of cut back the variety of simulations required to succeed in the optimum configuration for chip-to-chip communications.
…
