EdgeCortix introduces the SAKURA-II, an AI accelerator for the edge with impressive performance of 60 TOPS (INT8) and low power consumption of 8 watts. The chip is designed for complex AI tasks, such as processing large language models (LLM), large vision models (LVM), and transformer-based multimodal applications, as well as for edge applications such as IoT devices and autonomous vehicles.
It offers hardware flexibility, allowing integration into PCIe expansion cards with one or two SAKURA-II chips , or into M.2 2280 modules (PCIe x8 or x16 interface) if preferred. Therefore, performance can reach up to 120 TOPS with INT8 or 60 TFLOPS with BF16.
The AI ​​platform also offers a cutting-edge software component, featuring a MERA suite for programming and optimization, a heterogeneous build platform, advanced quantization techniques, and model calibration capabilities. It also stands out for its integration with popular development frameworks such as PyTorch, TensorFlow Lite, and ONNX, as well as access to a comprehensive library of state-of-the-art transformational and convolutional models.
Furthermore, EdgeCortix has also considered taking its AI accelerator designs further, making them possible for integration into SoCs from other companies, such as AMD.
IA SAKURA-II technical specifications
The technical specifications of the EdgeCortix SAKURA-II are as follows:
- NPU with DNA-II or second-generation Dynamic Neural Accelerator architecture.
- Performance up to 60 TOPS with INT8 or 30 TFLOPS with BF16.
- DRAM memory with dual 64-bit channel type LPDDR4x (8GB, 16GB, 32GB on-board) with bandwidth up to 68 GB/s.
- Integrated 20MB SRAM memory.
- Efficiency of up to 90% utilization, with energy consumption of 8W.
- BGA packaging.
If we refer to the SAKURA-II module with M.2 format , we have:
- DRAM memory
- 8GB (2x banks of 4GB LPDDR4)
- 16GB (2x banks of 8GB LPDDR4)
- PCIe Gen 3.0 x4 interface
- Maximum performance of 60 TOPS on INT8, 30 TFLOPS on BF16
- 10W module power
- Dimensions M.2 2280 (22x80mm)
For the PCIe expansion card , the specifications for the AI ​​accelerator are as follows:
- PCIe Gen 3.0 x8 interface
- For SAKURA-II single chip model:
- 16GB DRAM memory (2x banks of 8GB LPDDR4)
- Performance of 60 TOPS on INT8, 30 TFLOPS on BF16
- 10W power.
- For the model with two SAKURA-II chips:
- 32GB DRAM memory (2x banks of 16GB LPDDR4)
- Performance of 120 TOPS on INT8, 60 TFLOPS on BF16
- Power of 20W
- 1x Slot
- Includes heatsink
As for the prices , if you're wondering, they will arrive from the second quarter of 2024 with:
- M.2 8GB: $249
- M.2 16GB: $299
- PCIe 1xSAKURA-II: $429
- PCIe 2xSAKURA-II: $749