Engineers at Habana Labs this week unveiled their second -generation processes designed for AI Deep Learning. Take the form of Havana Gaudi 2 Training and Habana Greco Inference processors are the new hardware as it is intended for AI deep learning applications using 7nm technology. The company also created the Habana SynapseAI Software Suite, which is designed for deep learning model enhancement and to facilitate the migration of existing GPU-based models to hardware on the Gaudi platform.
“Habana Gaudi2 significantly increases training performance, building on the same high-efficiency first-generation architecture that allows customers to have up to 40% better AWS cloud price performance with Amazon EC2 DL1 Instances and on-premise Supermicro X12 Gaudi Training Server.In addition to its 7 nm jump from 16 nm, Gaudi2 also has 24 Tensor Processor Cores, an increase from the eight cores of the first Gaudi and designed obviously for large deep learning workloads.The new Gaudi2 AI training processor integrates on-chip media processing, triple on-board memory up to 96 GB and double SRAM up to 48 MB. This contributed to the performance of Gaudi2 – up to three times the training progress of the first generation Gaudi.
AI Deep Learning
“The launch of new deep learning processors in Habana is a prime example of Intel implementing its AI strategy to provide customers with a wide choice of solutions-from the cloud to the point – addressing the growing number and complex nature of AI workloads.Gaudi2 can help Intel customers train larger and more complex workloads in deep learning with speed and efficiency, and we look forward to the inference efficiencies that Greco will bring. ” – Sandra Rivera, Intel executive vice president and general manager of Datacenter and AI Group”
“Compared to the A100 GPU, implemented on the same process node and roughly the same die size, the Gaudi2 provides a clear performance of leadership training as shown in the apple-to-apple comparison of the main workloads, ”said Eitan Medina, chief operating officer of Habana Labs. “This in-depth learning of accelerating architecture is more efficient and supported by a solid road map.”
Source: Habana Labs
Filed Under: Technology News, Breaking News