FPGA Achieves Up to 28x Better Occupancy vs. GPUs for Edge AI
Using an Agilex™ 5 D-Series device, Altera partner Heronic Technologies achieved 1.9× greater performance, 48% lower latency, and comparable power efficiency, but utilized 28× more of the available TOPS versus the Jetson AGX Orin 64GB for an edge AI inference design example.
1.9×
Better performance FPS
vs.
AGX Orin 64GB
1.7×
Better performance FPS
vs.
Orin NX 16GB
28×
Better occupancy (%)
vs.
AGX Orin 64GB
5.9×
Better occupancy (%)
vs.
Orin NX 16GB
Edge AI’s demand for low-latency, energy-efficient hardware challenges traditional inference capabilities. GPUs/NPUs struggle with inefficiencies in any one application due to their generalized ’one-size-fits-all’ design approach, leading to underutilized compute units.
Altera Partner, Heronic Technologies, produces A/I accelerators for Altera FPGA devices with their proprietary design automation tool flow. Combined with consulting services, Heronic uses AI-driven Design Space Exploration (DSE) techniques to find an optimal architecture design for a customer’s specific AI application and FPGA pair. This approach maximizes Altera’s unique FPGA resources while minimizing memory bottlenecks.
What You’ll Learn:
- Insights into why GPU-advertised peak performance is not always usable performance
- How Heronic addresses a key knowledge gap: designing customized AI for FPGA customers who have limited AI or FPGA expertise, while also significantly shortening the time to design and deploy AI accelerators.
- How an Altera Agilex 5 D-Series SoC FPGA AI accelerator IP block was created using only three inputs: the AI model, the target FPGA device, and the desired performance/resources available.
Talk to Sales
Learn how Agilex 5 devices deliver scalable performance, lower power consumption, and low-latency for high-efficiency embedded, vision, and AI systems.
