Fabless chipmaker EdgeCortix Inc. has launched its latest edge computing artificial intelligence (AI) accelerator designed for manufacturing, security, robotics, aerospace and telecommunications.
Called the Sakura-II platform, the accelerator is designed to manage a range of tasks like large language models (LLMs), large vision models (LVMs) and multimodal transformer-based applications. The device is based on EdgeCortix’s second generation dynamic neural accelerator architecture that includes low latency, high accuracy, compact form factors and high memory bandwidth.
The AI accelerator can deliver up to 60 trillion operations per second (TOPS) of 8 bit integer performance and 30 trillion 16 bit brain floating point operations per second (TFLOPS). This allows the accelerator to deliver power efficiency and real-time processing capabilities while simultaneously executing multiple deep neural network models with low latency.
Additionally, the Sakura-II platform comes with the company’s MERA software suite, a heterogeneous compiler platform with advanced quantization and model calibration capabilities. The suite supports frameworks like PyTorch, TensorFlow Lit and ONNX. With the MERA Model Library, it allows users to access a range of transformer models to aid in a transition from training to edge inference.
Other features include:
- Tailored for generative AI workloads
- Managing multi-billion parameter models
- Compatible with EdgeCortix’s MERA software suite
- Up to four times DRAM bandwidth
- Real-time data streaming
- Software-enabled mixed-precision for near FP32 accuracy
- Spare computation support
- Arbitrary activation computation support
- Dedicated Reshaper engine
- On-chip power-gating and power management capabilities
