Amazingly Powerful Architecture for GPU-based ML Tasks
February 12, 2019

Amazingly Powerful Architecture for GPU-based ML Tasks

Anonymous | TrustRadius Reviewer
Score 8 out of 10
Vetted Review
Verified User

Modules Used

  • Power System AC922

Overall Satisfaction with IBM Power

We’re currently using the AC922 across the organization in many of our research groups, prototyping and implementing workflows for ML-driven image analysis. We also use the system for genomics and statistical genetics applications on high-throughput sequencing data.
  • Fast access to in-memory data, speeding up GPU-based computation as compared to commodity hardware.
  • Easy to manage, with a powerful web-based management interface.
  • Integrates remarkably well with other offerings from IBM, including the ESS storage system.
  • Currently, we’d like to see more support for the Power9 (ppc64le) architecture among popular ML and bioinformatics packages, and lessen the need to compile packages ourselves to run on Power9.
  • We are fairly new to the AC922, but so far we have seen a speed-up on some of our image analysis tasks. Also, the integration with ESS has made it easier to migrate in data.
The AC922 is significantly faster for many of our tasks, and the management interface lowers the overhead of ownership. Also, the option of using IBM’s scheduler is very compelling for streamlining use in the future.
Operationalized AI pipelines, with high throughput and repeatability, are the ideal scenario for the AC922. Also, GPU-heavy workloads that require large in-memory data sets are what this server does very, very well. Somewhat less appropriate is running libraries that are heavily or exclusively x86-focused.