INTRODUCING HABANA® GRECO®
2nd-GENERATION INFERENCE PROCESSOR.
BUILDING ON HABANA’S FIRST-GENERATION INFERENCE ACCELERATOR TECHNOLOGY, GRECO INCREASES PERFORMANCE AND EFFICIENCY WITH A VARIETY OF ADVANCES.
Memory: With 16GB LPDDR5 memories, Greco offers a 5x boost in memory bandwidth relative to Habana’s first-generation inference processor, previously known as Goya, and an increase in on-chip SRAM from 50 to 128 MBs, helping to boost inference performance on all model types.
Compute: To enable greater inference speed and efficiency targeting computer vision deployments, Greco integrates an independent media engine for managing compressed media natively, supporting media formats HEVC, H.264, JPEG and P-JPEG. In addition, Greco will support new data types, Bfloat 16, FP16, INT8/UINT8, INT4/UINT4, giving customers more alternatives and flexibility in balancing inference speed and accuracy.
Form factor: Greco form factor is reduced from our first-generation inference processor’s dual-slot PCIe card to single-slot, half-height, half-length (HHHL) PCIe Gen 4 x 8, packing the performance of the full PCIe card format into the compact HHHL to deliver improved compute density and system design efficiency and flexibility.
Lower power: Greco significantly reduces power from 200W TDP Goya to 75W TDP, enabling lower cost of operations for inference deployments.
Developing on Greco: Greco is paired with the Habana SynapseAI Software Suite, the Habana Developer Site and Habana GitHub, and features integration of TensorFlow and PyTorch frameworks.
Habana will sample Greco with customers in Q3 of this year; mass production is scheduled for 1H 2023.
For more information on the soon-to-arrive Habana Greco processor, contact us >