The Biscotti E1.S uses up to 2 Hailo-8 Edge AI processors, featuring up to 26 tera-operations per second (TOPS) each. With an architecture that takes advantage of the core properties of neural networks, the Hailo-8 neural chips designed onto the E1.S allow edge devices to run deep learning applications at full scale more efficiently, effectively, and sustainably than other AI chips and solutions.
Biscotti
Features
- Up to 1600 TOPs of inference performance with air-cooled systems
- Inference for Visual Analysis or Generative AI
- Plug-and-play and hot swappable
- Easily upgrade from a Biscotti E1.S AI module to the next generation, doubling performance instantly
- Use 20% of the Watts with TPUs compared to training GPUs
- 20% of ASP and TCO compared to training GPU servers
- 12-week lead times, significantly less than the typical for GPU servers
- Generative AI models with up to 1.8 billion parameter
Specifications
| Biscotti AI Module | ||
| AI Processor | Dual Hailo-8 | |
| Form Factor | E1.S (EDSFF) | |
| Power (max) | 10 Watts | |
| Performance | 52 TOPS | |
| Interface | 4 Lane PCIe Gen 3 | |
| Supported Frameworks | TensorFLow, TensorFlow Lite, ONNX, Keras, Pytorch | |
| Supported OS | Linux, windows | |
| Certification | CE, FCC, VCCi, KCC, WEEE | |
| Dimension | 33.75mm x 118.75mm x 9.5mm (5.9mm w/o Heat Sink Enclosure) | |
| Operating Temperature | 0°C ~ 70°C | |
| Biscotti Performance (With Atom X6414RE) | ||
| Network (AI) | Frames per Second (HW) | |
| Resnet_V1_50-Imagenet | 2692.61 | |
| Yolov3_Gluon | 50 | |
| Yolov3 | 50.4 | |
| Yolov4-leaky | 56.59 | |
| Centerpose_regentx_1.6gf | 4633.13 | |
| Eight Biscotti Module Performance (w/AMD Genoa System) | ||
| Network (AI) | Frames per Second (HW) | |
| Resnet_V1_50-Imagenet | 21.5K | |
| Yolov3_Gluon | 400 | |
| Yolov3 | 403.2 | |
| Yolov4-leaky | 452.72 | |
| Centerpose_regentx_1.6gf | 37.1K |
Ordering Information
| Part Number | Note |
| UCT201-021-NC | Dual AI TPU w/12 lane Switch |
| UCT201-021-HS | Dual AI TPU (9.5mm Heatsink) w/12 lane Switch |