Announcing our newest W&B SDK performance enhancements
We've made wandb-core much more performant. Learn what's new and how you can take advantage of our latest improvements.
Created on July 10|Last edited on September 23
Comment
Updated Sep 23: The new wandb-core SDK is now generally available and enabled by default for all users. You can learn more about it here.
💡
We are excited to announce that a completely redesigned wandb-core is now public preview.
wandb-core is an advanced backend for the Weights & Biases SDK. The new wandb-core reduces memory and CPU usage, delivers faster operations—like startup, shutdown, and offline syncs—and provides more stable operations, especially under high resource contention.
Some of the most significant enhancements include:
- Lower system resources footprint: Reduced memory and file descriptor usage, improving overall efficiency.
- System robustness: Enhanced stability on systems with high resource contention.
- Startup / shutdown performance: Faster startup and shutdown times, particularly for short experiments.
- Parallel run performance: More efficient CPU usage and higher performance for parallel tasks.
- Higher performance file logging: Faster and more efficient logging operations.
- Network utilization of filestream: More efficient data management, reducing network data transfer.
- Offline sync speeds: Accelerated offline sync speeds for long-running experiments.
Benchmarks
We wanted to establish quantifiable benefits of the redesigned wandb-core, so we executed a set of benchmarks focused on startup / shutdown performance, parallel run performance, and logging (for reference, the SDK supports online and offline modes). Offline logging allows our users to decouple running experiments from syncing the data to the backend server. We ran the benchmarks for each mode.
Startup/shutdown performance
When initiating an experiment there is a setup cost to make sure the experiment is reliably tracked. The delays in starting and stopping an experiment is SDK overhead which takes away from compute resources which could be used for model training. This overhead is mostly noticed with very short experiments (measured in seconds). We wanted to understand the reduction in overhead using the new SDK.
Parallel run performance
The SDK has the ability to track multiple experiments in parallel (like using python multiprocessing, for example). We wanted to understand how much more efficient parallel experiments are using the new SDK.
Logging tables
W&B Tables are an important datatype that allows detailed analysis in the wandb UI. We wanted to understand how much more efficient logging of tables is using the new SDK.
| Metric | Mode | Improvement with wandb-core |
|---|---|---|
| Startup/Shutdown Time | Offline | 36% improvement |
| Startup/Shutdown Time | Online | 23% improvement |
| Parallel Scalar Logging Performance | Offline | 83% improvement |
| Parallel Scalar Logging Performance | Online | 88% improvement |
| Table Logging Performance | Offline | 18% improvement |
| Table Logging Performance | Online | 40% improvement |
Artifacts
Another part of the redesign was to help improve artifact upload/download performance based on feedback from customers. wandb-core is now optimized for faster data transfer, significantly reducing the time required for uploading and downloading artifacts.
For example, downloading a 10MB artifact now takes 2.81 seconds compared to 6.67 seconds with the previous backend. This means datasets, models, and other large artifacts are managed more efficiently, reducing delays and improving workflow. We consistently see 33% faster artifact upload times and 28% faster artifact download times.
Get started
wandb-core is now enabled by default for all users. Check out the notebook to start using it and reach out to the support team at support@wandb.com with any questions.
Add a comment
Iterate on AI agents and models faster. Try Weights & Biases today.