While cloud computing made hardware commodity, AI is putting the focus back on hardware. Hardware has always been a key driver for HPC, however for business workloads hardware was increasingly taking a back seat.
The following article on technology requirements for Deep and Machine Learning, touches upon some of the hardware aspects like cache and memory capacity, memory bandwidth requirements and accelerators (GPUs and FPGAs).
Increasingly we are seeing benchmarks on different hardwares, comparing training times, which is key for data scientists. It’s very likely that benchmark results for AI frameworks might soon become very common among different hardware vendor and who knows we might soon have a SPEC GPU benchmark.
With that said, choice of software platform is equally important to provide the needed agility for data scientists and developers. Data scientists are increasingly adopting containers to improve their workflows by realising container benefits like dependency management, reproducible artefacts etc. In an enterprise AI setup you will also need functionalities like optimal resource utilisation, self-service, resource controls, CI/CD pipelines, governance among other things. For example allowing GPUs to be used only by data scientists, setting up resource quotas – CPU/MEM/GPU for different groups, automatically provisioning AI jobs in a cluster of nodes etc.
The key aspect here is that traditional IT infrastructure solutions cannot deliver the required capabilities for AI. There is a need for hardware and software solution that is designed for AI.
One such hardware and software solution that you can evaluate is IBM Power9 processor based servers with GPUs and IBM Cloud Private – an application platform built on Kubernetes:
Checkout the following references to learn more:
- Faster training times with Power9 and NVIDIA GPUs
- Enterprise AI with IBM Cloud Private and Power servers
- Setting up CI/CI pipeline using IBM Cloud Private
- Why hardware matters in the cognitive enterprise
As always, please feel free to share your inputs in the comments section.