Specifications
Balanced inference efficiency and performance
Optimized for the most cost- and power-sensitive designs, Ethos-N57 delivers premium AI experiences in mainstream phones and DTVs. With the highest performance, an open-source software framework and the largest AI ecosystem, the Arm AI platform makes it easy to develop on Arm.
Ethos-N57 mainstream ML inference processor contains eight compute engines
Key features | Performance (at 1GHz) |
2 TOP/s |
MACs (8x8) | 1024 |
|
Data types | Int-8 and Int-16 | |
Network support | CNN and RNN | |
Efficient convolution | Winograd support | |
Sparsity | Yes | |
Secure mode | TEE or SEE | |
Multicore capability | 8 NPUs in a cluster 64 NPUs in a mesh |
|
Memory system | Embedded SRAM | 512 KB |
Bandwidth reduction |
Extended compression technology, layer/operator fusion | |
Main interface |
1xAXI4 (128-bit), ACE-5 Lite | |
Development platform | Neural frameworks | TensorFlow, TensorFlow Lite, Caffe2, PyTorch, MXNet, ONNX |
Neural operator API | Arm NN, AndroidNN | |
Software components | Arm NN, neural compiler, driver and support library | |
Debug and profile | Layer-by-layer visibility | |
Evaluation and early prototyping | Arm Juno FPGA systems and cycle models |
Key features
Balanced Performance
Delivers up to 2 TOP/s of performance using 1024 8-Bit MACs.
Optimized Design
Drives up to 225% convolution performance uplift using Winograd on 3x3 kernels, delivering up to 90% MAC utilization.
High Efficiency
Internally distributed SRAM stores data close to the compute elements to save power and reduce DRAM access.
Futureproof
Supports a wide range of existing Machine Learning (ML) operations and future innovations through firmware updates and compiler technology.
Key benefits
The key benefits of Ethos-N57:
- Supports a variety of popular neural networks, including CNNs and RNNs, for classification, object detection, image enhancements, speech recognition and natural language understanding
- Reduces system memory bandwidth by 1.5-3x through clustering sparsity and workload tiling, with lossless compression for weights and activations on select networks
- Maximizes the number of parameters stored on-chip by storing compressed weights and activations in local SRAM and decompressing them on the fly
- Leverages sparse power gating techniques to reduce power by up to 50%
- Improves performance and extends battery life through intelligent data management techniques to minimize memory movement with up to 90% of accesses on chip
- Supports TrustZone system security to safeguard sensitive data with support for secure and non-secure modes
Ethos-N comparison
Ethos-N78 |
Ethos-N77 |
Ethos-N57 |
Ethos-N37 |
||
Key features | Performance (at 1GHz) |
10. 5. 2. 1 TOP/s | 4 TOP/s | 2 TOP/s | 1 TOP/s |
MAC/Cycle (8x8) | 4096, 2048, 1024, 512 | 2048 | 1024 |
512 | |
Efficient convolution |
Winograd support delivers 2.25x peak performance over baseline | ||||
Configurability | 90+ design options | Single product offering | |||
Network support | CNN and RNN | ||||
Data types |
Int-8 and Int-16 | ||||
Sparsity | Yes | ||||
Secure mode |
TEE or SEE | ||||
Multicore capability |
8 NPUs in a cluster 64 NPUs in a mesh |
||||
Memory system | Embedded SRAM | 384KB – 4MB | 1-4 MB | 512 KB | 512 KB |
Bandwidth reduction | Enhanced compression | Extended compression technology, layer/operator fusion, clustering, and workload tilling | |||
Main interface | 1xAXI4 (128-bit), ACE-5 Lite | ||||
Development platform | Neural frameworks | TensorFlow, TensorFlow Lite, Caffe2, PyTorch, MXNet, ONNX | |||
Inference deployment | Ahead of time compiled with TVM Online interpreted with Arm NN Android Neural Networks API (NNAPI) |
||||
Software components | Arm NN, compiler and support library, driver | ||||
Debug and profile | Heterogeneous layer-by-layer visibility in Arm Development Studio Streamline | ||||
Evaluation and early prototyping | Ethos-N Static Performance Analyzer (SPA), Arm Juno FPGA systems, Cycle Models |