Super fast data crunching for deep learning inference workloads
15,488 Images-per-second inference throughput with .8ms latency
AI InferenceGOYA INFERENCE PERFORMANCE
BERT – Base SQuAD Performance
Goya performance on BERT-Base, a popular natural language processing model, is based on the following configurations:
Hardware: 1x GOYA HL-100; CPU Xeon Gold [email protected] GHz.
Software: Ubuntu v-18.04; SynapseAI 0.11.0-477
Performance on ResNet50
Goya delivers super fast throughput and ultra low latency on ResNet50, a popular image recognition model. These performance metrics are based on:
Hardware: Goya HL-100 PCIe card; host CPU Xeon Gold [email protected] GHz; and
Software: Ubuntu v-18.04, SynapseAI v-0.11-447.
Inference performance: All AI modelsInference published MLPerf results
GOYA INFERENCE PRODUCTS
GOYA HL-100 PCIe CARD:
• PCIe: Gen4 x 16 lanes
• Form Factor: Dual-slot
• Memory: 16GB with ECC
HABANA SynapseAI® SOFTWARE Suite for Inference
Habana’s software platform provides efficient and flexible development and deployment of a wide array of inference workloads and ease of customization to suit user-specific solution requirements. Designed for flexibility, Habana’s inference software leverages foundational and flexible blocks: Habana’s programmable Tensor Processor Core (TPC) SDK, optimizer, graphic compiler and runtime software, a rich and extensible kernel library and data center deployment and management tools.
TPC KERNEL LIBRARY & SDK
- Programmable TPC with SDK
- Extensive TPC kernel libraries support user customization
- TPC tools (simulator, compiler, debugger) for custom kernel development
OPTIMIZER, COMPILER & RUNTIME
- Seamlessly integrates with existing frameworks
- Can be interfaced with C or Python API
- Supports development flow from graph to optimized recipe
- Enables quantization, pipelining, layer fusing
PERFORMANCE MANAGEMENT & DEPLOYMENT TOOLS
- Deployment validation toolset: test and monitor functionality and performance
- Libraries and tools for run-time detection and reporting
- Run-time and orchestration plug-ins
HABANA INFERENCE SOFTWARE PLATFORM
SUPPORTS CUSTOMIZATION AT EVERY LAYER
