An Accelerator for Recommendation Systems
Deep learning based recommendation models are one of the most common data center workloads, generating hundreds of millions of dollars in revenue. Use-cases include search, news feeds, adverts and personalized content.
- SEAL™ can halve the cost of the infrastructure required for the peak traffic capability these systems need
- SEAL is the most efficient and practical way to enhance the capability of existing server infrastructure
Get Much More out of Your Infrastructure
Scale Up, not Out
SEAL is a co-optimized hardware/software accelerator, designed to scale up existing infrastructure. SEAL is powered by a PyTorch extension library that:
- accelerates the memory-bound sparse operations in all recommendation models
- delivers large gains in latency bounded throughput
- fully preserves existing model accuracy
- is complementary to existing compute accelerators
SEAL is available initially in a form factor conforming to the OCP M.2 Accelerator Module specification.
Eliminate Memory-Bound Bottlenecks
Recommendation models always contain both dense and sparse features. Up to 80% of time can be spent on the memory-bound sparse features, where existing accelerators and infrastructure give a poor return.
SEAL provides tailored memory bandwidth for these memory-bound features, eliminating the bottleneck. This leads to more rankings per impression, hence better recommendations and increased revenue.
Complementary to Existing Infrastructure
SEAL is easy to install in existing infrastructure, either in available slots or via a carrier card
- Easy to install with existing servers
- Complementary to other accelerators
- Does not require any change to the recommendation model. No model retraining. No degradation in accuracy
- Supports co-location of models with no performance penalty
- Supports concurrent deployment of different versions of a model, and loading/unloading models on the fly to facilitate A/B testing
SEAL represents the lowest power, smallest form factor, easiest-to-deploy method of adding memory bandwidth to existing infrastructure used for recommendation models
SEAL is available initially in a form factor conforming to the Open Compute Project M.2 Accelerator Module specification
To evaluate what SEAL can do for your business