# Hardware configuration
Studio uses advanced technology to support up to a max of 128 neural engines and/or GPUs.
The limitation of 128 neural engines is caused by network configuration. In our roadmap, 1024 engines will be supported.
# Capability matrix
Each neural engine and GPUs have its own detection capabilities.
|Apple M1 Pro/Max||Classifier||CoreML||20|
|Apple M1 Ultra||Classifier||CoreML||40|
|NVIDIA GeForce RTX 2080 Ti||Classifier||Tensorflow||25|
|NVIDIA GeForce RTX 3080*||Classifier||Tensorflow||65|
Note: * The FPS is estimated based on the relative performance of TFLOPS FP32 to NVIDIA GeForce RTX 2080 Ti. Actual performance may vary.
Training capabilities are different for different cluster configuration.
|Apple M1/M1 Pro/M1 Max||Classifier||CoreML|
|Apple M1 Ultra||Classifier||CoreML|
|Azure Custom Vision||Classifier||CoreML, Tensorflow|
Currently we only support on-device training on Apple M1 processors. Training on NVIDIA processors can be performed using other third-party software systems such as JupyterLab, but it cannot be directly trained in Studio.
# Starter design
To start with a small team configuration of 5 developers, you can subscribe to Studio Business. Each developer can install EdgeAI App in the local machine for ML acceleration.
In Business version, Studio does not support on-premise storage. However, it supports using storage engine from Amazon S3 and Azure Blob Storage by simple configuration.
# EdgeAI App
EdgeAI App is an application which can be run in iOS, iPadOS and macOS. The purpose of using the app is to accelerate the machine learning operations in Studio. The app can also be executed standalone as an edge device to run pipelines in guest mode.
# On-premise cluster design
With all capabilities, you can design the cluster which consists of a pair of load balancers and manager nodes and a group of neural nodes from Apple M1 and NVIDIA.
# Component view
# DMZ and load balancers
A set of load balancers should be in placed to balance the incoming requests to the manager nodes. At least 2 nodes are required for resilience.
# Manager nodes
Manager nodes are responsible for cluster management. At least 3 nodes are required to form a quorum.
# Neural nodes
Neural nodes can be a mix of different processors. Manager nodes will automatically discover the capability of each neural node.
# How many neural nodes shall be provisioned?
The number of nodes depends on how many frames the platform shall be processed in a second as well as what capabilities it should support.
For example, you want to provide a platform with a capacity of 4 x 25-FPS streams of object detection. In total, the platform shall be able to process 100 FPS (= 4 x 25). Since one Apple M1 provides a neural performance of 20 FPS, 5 neural nodes (= 100 / 20) are needed.