Da is co-founder and CEO of Expedera. Previously, he was cofounder and COO of Memoir Systems, an optimized memory IP startup, leading to a successful acquisition by Cisco. At Cisco, he led the Datacenter Switch ASICs for Nexus 3/9K, MDS, CSPG products. Da brings more than 25 years of ASIC experience at Cisco, Nvidia, and Abrizio. He holds a BS EECS from UC Berkeley, MS/PhD EE from Stanford.
Tell us about Expedera?
Expedera has developed deep learning accelerator (DLA) IP that has the industry’s best performance per watt—18TOPS/watt. Our solutions are scalable to 128 TOPS with a single core and to PetaOps with multicore. We started from the ground up with a hardware/software codesign approach that enables us to deliver the most power efficient and scalable deep learning accelerator (DLA) for AI inference.
Our design outperforms other DLA blocks from leading vendors such as Arm, MediaTek, Nvidia, and Qualcomm by at least 4–5x. We’ve validated this using our 7nm test chip.
We are targeting AI inference, particularly for edge applications. We have at least one customer in this space currently in production—a top smartphone manufacturer.
My cofounders and I founded Expedera in 2018. Our office is in Santa Clara.
What problems/challenges are you solving?
We provide a highly efficient AI inference solution. If a customer needs deterministic performance or a guaranteed level of performance with the best possible power and area efficiency, we can do that. If they need a solution that doesn’t require off-chip memory, we can do that. If they need a flexible, future-proof solution that can handle mixed models, we can do that. We also bring efficiency to model deployment because our co-designed platform reduces software complexity dramatically and ensures predictable performance.
What markets does Expedera address today?
We have announced a top-10 smartphone customer, so it’s fair to say mobile and edge AI are a sweet spot for us. Because of our scalability and determinism, we are a good fit for automotive and industrial automation. In fact, we are engaged with customers from GOPS to PetaOPS.
What are the products Expedera has to offer?
Our Origin deep learning accelerator IP platform addresses a wide variety of AI inference applications. The platform includes silicon IP and a comprehensive SDK built on TVM that provides a Compiler that achieves out-of-the-box high performance. The platform allows us to easily support different precisions and features—it’s very flexible.
What keeps your customers [architects, system designers] up at night?
The reality is that most AI processors are underperforming and stall at around 30-50% utilization or less —wasting most of their potential TOPS. So system architects and designers overdesign their SoC to address unpredictable performance. Expedera provides predictable, deterministic performance with 90% utilization. Greater utilization results in better throughput for customers. Our platform gives architects the end-to-end visibility needed to right-size their AI-accelerator solutions early in the development cycle.
Another issue is the difficulty, the delays, and the uncertainty in model deployment. Data scientist can spend tremendous amounts of time to achieve minimal performance improvement. With Expedera, engineers can deploy trained models without further changes. That increases confidence in the design, and avoids difficult development tradeoffs, bottlenecks and product uncertainty.
What added value do you bring to your customers?
Confidence in their solution. Efficient operation. Ease of deployment. Reduced BOM costs.
Providing an AI-solution as an IP has huge implications for both our business and our customers. The IP licensing approach allows us to address a broad set of edge-AI markets, and potentially license to leading vendors that already hold large shares in these markets. At the same time, we can enable startups and new market entrants that may not have the in-house expertise to design their own AI hardware and would otherwise be unable to participate or compete with incumbents.
What makes Expedera unique?
We’ve taken a fundamentally different approach to AI acceleration, in part, because we come from a networking background. We’ve taken a network-centric approach—rather than the CPU-centric approach—to neural network processing. We are able to segment the neural network into packets which are essentially command streams that can be efficiently scheduled and executed by our hardware in a very fast, efficient and deterministic manner. Additionally, our co-design approach enables a simpler software stack and a more productive and system-aware design and development experience.
What’s driving the company’s global expansion/growth?
The market expectation of AI-everywhere is driving growth and creating a competitive necessity for ODMs to provide increasingly intelligent and autonomous products. We are still in the hockey stick of AI deployment.
CEO Interview: Pradeep Vajram of AlphaICs
CEO Interview: Charbel Rizk of Oculi
CEO Update: Tuomas Hollman, Minima Processor CEOShare this post via:
There are no comments yet.
You must register or log in to view/post comments.