Solving AI Hardware Challenges

Print Friendly, PDF & Email
PCIe Gen 4

In this week’s Sponsored Post, Katie Rivera of One Stop Systems explores some of the AI hardware challenges that can arise, as well as the new tools designed to tackle these issues. 

Katie Rivera, One Stop Systems

Katie Rivera, Marketing Communications Manager, One Stop Systems

GPUs are the brain of artificial intelligence, deep learning and machine learning. The more GPUs, the “smarter” the brain. So it follows that users would want to use as many GPUs as possible to train AI and deep learning systems. For many deep learning startups out there, buying AI hardware and a large quantity of powerful GPUs is not feasible. However, GPUs are crucial because they provide a powerful tool for these applications, so many of these startup companies are turning to cloud GPU computing to crunch their data and run their algorithms.

[clickToTweet tweet=”Katie Rivera – Utilizing GPUs in the cloud, users quickly process their data in parallel.” quote=”Katie Rivera – Utilizing GPUs in the cloud, users quickly process their data in parallel.”]

SkyScale, a provider of cloud-based ultra-fast multi-GPU hardware platforms, leases time on the latest generation compute engines in increments as small as an hour. SkyScale’s model, termed “bare metal”, provides dedicated, secure hardware so users can save time and money without compromising performance. Utilizing GPUs in the cloud, users quickly process their data in parallel, reducing the amount of time it takes to process data for applications such as image and video recognition.

Facial recognition is nothing new. Every time you upload a picture to Facebook and go to tag your friends, Facebook has already tagged them for you. Several years ago cell phones could unlock your phone by scanning your face, and the technology is getting better rapidly. Earlier in 2017, Google’s new machine learning API recognized objects in videos. The next logical step is recognizing actions in videos.

In July, the Computer Vision Pattern Recognition (CVPR) event took place. In conjunction with CVPR, Activity Net holds a half-day workshop called the ActivityNet Large Scale Activity Recognition Challenge, where participants compete in various tasks. The Trimmed Action Recognition (Kinetics) task evaluated the ability of algorithms to recognize human actions in a large dataset of 10 second short video clips. The data set contained approximately 300,000 videos from YouTube with 400 different types of actions. SkyScale’s customer, twentybn, recently placed third in this task.

“[twentybn] trained deep neural networks that captured different statistical modalities present in the data and ensembled the obtained predictions to get the final result.”

To accomplish twentybn’s impressive feat, SkyScale rented them a One Stop Systems (OSS) GPU Accelerated Server, OSS-PASCAL4. The OSS-PASCAL4 features four NVIDIA Tesla P100 SXM2 GPUs, which provides 21.2 TeraFLOPS of double precision performance and 84.8 TeraFLOPS of half precision performance for deep learning and machine learning applications. The system uses the NVIDIA Tesla P100 GPU accelerators; the most advanced GPU ever built for AI and deep learning applications. Mature and start-up AI companies can dramatically increase throughput while also saving money by using fewer higher performance, lightning-fast nodes made available for rent in the cloud.

Looking ahead, OSS has designed a new SXM2 GPU Accelerator with eight NVIDIA Tesla P100 SXM2 GPUs to allow AI, deep learning and machine learning users to add additional compute power to their existing servers. There are not many SXM2 capable servers on the market today, so the OSS SXM2 GPU Accelerator allows customers to add the most powerful GPUs to existing infrastructure. Visitors to GTC Europe in Munich in October will be able to see the SXM2 GPU Accelerator at the OSS booth #E05.

This guest article was submitted by Katie Rivera, marketing communications manager at One Stop Systems.