Nvidia is marketing enhancements to its AI tech for understanding language that it claimed sets new records for performance on conversational AI. The GPU manufacturer claimed that its AI platform now has the fastest inference, the quickest training record, and biggest training model of its type till date.
By including major optimizations to its GPUs and AI platform, Nvidia is targeting to turn out to be the premier supplier of conversational AI offerings, which it claims have been restricted up to this point owing to a wide inability to set up huge AI models in real time.
Conversational AI, unlike the much easier transactional AI, uses nuance and context and the responses are immediate, claimed Bryan Catanzaro (vice president at Nvidia for applied deep learning research) to the media in an interview. As such, the models for conversational AI require being very huge and real-time so as to operate with human-akin comprehension.
As for its breakthroughs in performance, Nvidia claimed that it has set the world record for educating the popular Google-developed BERT (Bidirectional Encoder Representations from Transformers) language model. Nvidia skilled a huge BERT model on its DGX SuperPOD deep learning server in 53 minutes by using model parallelism, a method to divide a neural network when models are very huge to fit inside the memory of a sole GPU.
“Without this type of tech, it can take weeks to educate one of these huge language models,” claimed Catanzaro.
On a related note, Nvidia this week launched out EGX, a service for pacing AI in edge apps. The platform merges AI tech from Nvidia with storage, security, and networking techs from Mellanox. The software stack of the platform, NVIDIA Edge Stack, is optimized for real-time AI offerings such as speech, vision, and analytics.