Soul Machine Stokes Next-Gen AI Engines

Published on 12 May 2022

Soul Machine, AI Engines

Soul Machines' Human OS technology revolutionizes customized service. The firm intends to democratize the service sector by producing AI-based Digital PeopleTM for smooth consumer interactions. Soul Computers need strong machines without memory or compute limits to train machine learning and AI models.

The business employed the HP Z8 G4 workstation with Nvidia Rapids software package to set the path for huge models like an in-house version of Facebook's BlenderBot, a machine-learning chatbot on one workstation.

Human-machine cooperation's benefits

Soul Machines created Digital PeopleTM by merging hyper-realistic CGI and autonomous animation to create human-like interactions for global marketing experiences. “Soul Machines wants to build Digital People to help democratize the service industry,” says Shane Blackett, Vice President of Software and Technology.

Soul Machines, based in San Francisco and Auckland, New Zealand, works with visionary customers that seek happy end-users. “We want to go farther, where the digital person, the user, and the site content co-exist in a three-way interaction,” Blackett explains. "We're constructing a scalable platform for Digital People and conducting an AI-based simulation to add fidelity and realism."

Soul Machines' Human OS Platform has a patented Digital Brain enabling human-machine cooperation. The firm uses AI, machine learning, and other complicated algorithms. Soul Machines serves financial institutions and consumer products companies. A digital person for a worldwide CPG in Japan sells cosmetics in Japanese and English online.

Digital personas run models that identify learning, sensing, and behaviour. To train and operate AI-based and machine learning models and generate pictures takes a lot of CPU, video memory, and graphics. HP's Z8 Workstation. Alireza Nejati, Senior R&D Software Engineer at Soul Machines, says his team's training models are memory restricted. "GPU memory was a huge motivator for us"

Memory-powered

New Zealand adopted COVID-19 when Soul Machines chose the HP Z8 in early 2020. Nejati's staff had to wait until June to utilize the workstation after the corporation transferred its IT operations outside. “The Z8 is our most powerful machine,” adds Nejati. It enabled us to run several models, which wouldn't be feasible on any system due to memory requirements.

He mentions the Z8's configurability, Dual Xeon Gold CPUs, up to 3TB DDR5 ECC memory, 4 NVMe M.2 ports for high-speed storage, and up to 96 GB of GPU RAM utilizing two RTX 8000 graphic cards. Z8's Intel® OptaneTM DC Persistent Memory helps with big, complicated data sets.

Soul Machines ran the Z8 manually as a server. The business used the Nvidia Rapids software libraries packaged with the z8 to better manage resources. Soul Machines' biggest model previously trained on the HP Z8 was Facebook's BlenderBot, an open domain chatbot trained in online discussion from diverse sources using selected datasets. While BlenderBot contains 9 billion parameters, less than some machine-learning models. 

Nejati believes training the model on one system is difficult and requires 96 GB of video memory. "You can't do 50 gigabytes." When models were smaller, a single GPU could train one. “Modern models are so huge that even one instance of the model must be distributed over numerous GPUs,” Nejati explains. When GPUs are on various computers, software engineering becomes difficult. You must correctly spread the model over different GPUs since it executes sequentially.



Download Intel's whitepaper to learn more about Soul Machine Stokes Next-Gen AI Engines only on Whitepapers Online.