SemiWiki.com - Podcast EP291: The Journey From One Micron to Edge AI at One Nanometer with Ceva’s Moshe Sheier

Episode Date: June 13, 2025

Dan is joined by Moshe Sheier, Ceva’s vice president of marketing. Moshe brings with him more than 20 years of experience in the semiconductor IP and chip industries in both development and managerial roles. Prior to this position, Mr. Sheier was the director of strategic marketing at Ceva. Dan explores the history of Ceva with … Read More

Hello, my name is Daniel Nennie, founder of SemiWiki, the open forum for semiconductor professionals. Welcome to the Semiconductor Insiders podcast series. My guest today is Moshe Shair, Siva's Vice President of Marketing. Moshe brings with him more than 20 years of experience in the semiconductor IP and chip industry in both development and managerial roles. Prior to this position, Mr. Shair was Director of Strategic Marketing at Civa.

Welcome to the podcast, Moshe. Thank you, Daniel. I'm very happy to be here. So first, can you tell us how you got started in the semiconductor industry? Wow, so that was more than 30 years ago, actually. Like many electrical engineers, I studied different tracks as part of my studies, looking into communications

and control and so on. But I was always intrigued by the hardware and the need all those big systems, which ended up in VLSI and my first role in this big group, the company that Siva, my current employer, spun off from. So started long ago, basically, a view to the desire to understand how those little chips works underneath. And I have to say that the first chip I worked

on was the manufacturer that one micron technology, believe it or not, which is almost 1,000 bigger technology node than where we are today, getting close to the one nanometer, I guess, very soon. So quite a long way since then. And what brought you to SIVA? So as it happens, I had a friend working there who got me interested in the domain. Basically, I started as a VLSI engineer working on the first DSP calls that DSP group were making back in the days that later evolved to the NPUs that we're familiar with today. And it was all about voice and audio processing to begin with,

moving on to vision later on. So I spent a good 10, almost 14 years in R&D roles, engineering and management, and then moved over to the dark side, which is marketing, where I'm fulfilling my current position. Yeah, it's similar to my experience as well. So Moshe, as smart devices evolve, their demands are rapidly increasing and diversifying, you know, from ultra-low power MCUs to highly complex SoCs. How is SIVA strategically positioned its IP portfolio to meet these current requirements, but also to anticipate and address the future needs of smart edge devices capabilities. Okay. So indeed, our domain is very diversified today in terms of the requirement.

SIPA's strategy is based on solutions for three critical use cases, which we typically see that any smart device requires today. We call these a Connect, Sense, and Infer. Connect includes wireless connectivity IPs of Bluetooth, Wi-Fi, UWB, 5G, anything from massive IoT up to satellite communication and base stations even, to address any connectivity needs really, including multi-protocol platform solutions which we have recently introduced that significantly reduce development and integration efforts. The sense includes

signal processing solutions based on DSPs for applications like audio, vision, radar, embedded software for contextual user interactions for applications such as spatial audio, noise reduction, invoice conversations, motion sensing and sensor fusion. And last, in FAIL, this use case includes, of course, And last, in FAIL, this use case includes, of course, edge AI solutions in the form of scalable NPUs, targeting anything from embedded ML, like MCUs that you mentioned, all the way up to a generative AI at the edge for complex SOCs

in automotive, copilot applications, as they are called, for AI PCs and such, along with a unified AI SDK. So by supporting the latest and also emerging wireless standards, providing highly scalable NPUs ranging from hundreds of GOPs up to thousands of TOPs, we try to ensure that our customers can solve their most critical use cases of wireless connectivity, sensing and AI, those three critical use cases to enable current and future smart edge chips and devices. Right. Well, Siva serves diverse markets like automotive and consumer electronics and industrial IoT. How are you adapting your Edge AI solutions to meet the unique demands of each, especially the contrast between safety, critical automotive systems and low power consumer devices?

So Edge AI indeed involves a very broad set of workloads that you described. The way we address this is by offering our new pro, as we call it, family of NPUs with a broad scalability of performance and power efficient implementations. This starts with our low power andupro Nano NPUs. These are targeted embedded ML applications in consumer and industrial IoT, going into MCU devices, all sorts of smart sensors, and it expands to the high-performance automotive grade Nupro M and PUs that scales to hundreds of tops and even more using a multi-core configurations. So very diverse needs by ensuring one family

with a large level of scalability, we try to address these diverse needs. Right. But, you know, with the growing complexity of AI workloads like transformers, multimodal, and generative AI moving to the edge, what are the main challenges companies like Siva face in meeting power performance and latency demands? And how is your NPU architecture and software stack evolving to address this? So Siva has been developing NPU architectures for almost a decade.

It started with the vision processors that evolved into what we know today as the NPUs, with each architectural generation improving the level of performance and efficiency. To address that growing complexity of AI models, we're using a heterogeneous processing for better efficiency. So combining a multiply and accumulate arrays

along with sparsity mechanisms, compression, acceleration, adding support for low resolution and even floating point weights like FP16, FP8, even FP4 is coming up, is dictated by the model developers. Also introducing new types of networks, like transformers that you mentioned, that is the base of all the LLMs and LVMs that we know today. So each generation improves power efficiency,

reduces cycle count, reduces memory bandwidth requirements. To give some example example perhaps, our latest new program architecture is able to achieve 3,500 tokens per second per watt for a Lama2 7 billion parameter LLM implementation. So it has evolved a lot and is able to reach that level of performance with edge implementation of such complex LLM. One of the biggest challenges of SIVA as well as our customers is future proofing the NPU IP and the AI chips that are based on it.

So we address this by including a programmable processor based on our legacy DSPs as part of the NPU design, which allows even new future networks which are not used today to be mapped to this programmable processor and ensuring future compatibility to new networks and models which our customers may face. Right.

So SemiWiki has been working with Siva for many years since we started SemiWiki actually, and I know you as a customer centric company. So can you share examples of partnerships or customer deployments that illustrate SIVA's Edge AI strategy in action? Sure, so to pick a recent example, we recently announced that Next chip, a Korean ADAS chip company, has licensed our new program, NPU, the one I mentioned earlier, for automotive safety applications, while Renesas, a long-time customer, is already deploying the V4H, as they call it, SOC,

for automated driving, deploying different AI models on the SIVA processor. That's an automotive. Other than that, we have customers in the surveillance market like Novatec from Taiwan, a Boat, which is a big headphone OEM from India, using our processors for vision in the case of Novatec surveillance chips, and audio AI, the case of the Boat headphone. So different applications in automotive, consumer

applications. And on top of that, there is a vast ecosystem of third party partners providing various AI software solutions that are optimized for the SIVA and new pro NPUs. So we've been around for a long time. Also, our NPUs and AI processors are silicon proven, deployed, and few examples that I gave to show that. Beyond raw processing power and efficiency,

the software ecosystem, the tools, models, and support for various AI frameworks is critical for customer adoption. Can you elaborate on SIVA's strategy for building and nurturing a robust software and developer ecosystem around its Edge AI, AI-D? So that's very true. AI SDK is critical, perhaps in some cases even more than the APU hardware itself.

And we provide a unified SDK. By unified, I mean it supports the full SIVA Nupro family, going from the same tools, from the new PoNano embedded ML NPU all the way up to the new program, high performance automotive grade, hundreds of stops NPU. And this SDK supports the industry standards, AI frameworks, TensorFlow and so on. What we call an architectural planner,

which allows our SOC customers early analysis and profiling of their AI workloads to pick the right size, the right amount of hardware mechanisms, as well as memory sizes to fit the application. It includes graph compilers for optimization, quantization of the models mapping to the specific hardware. Also a full model zoo.

So many models that were already deployed, optimized by us or by partners for the new pro architectures to serve as a model or a starting point for their own application development and also the ecosystem offerings mentioned earlier. So indeed, the AISDK is very critical, I think, is very critical, I think, when a chip companies or AI developers in general are choosing an AI solution, they want to see an end-to-end solution, not just the NPU architecture, but the tool chain, the model zoo, the ecosystem, something that will allow them to go all the way from ecosystem, something that will allow them to go all the way from architectural planning to allow their customers to develop AI applications on the chip.

So this is what we try to address with our SDK. Okay, a final question Moshe, and I'm going to be blunt here. The Edge AI semiconductor market is becoming increasingly competitive with a mix of established players such as Siva, so many startups, and even large OEMs developing in-house solutions. What is Siva's key differentiation in this landscape? How do you ensure that your IP solutions offer compelling value proposition

for companies looking to integrate Edge AI capabilities? Okay, so indeed, that's a good question. The field is very busy with many solutions. Or I would say the market is full of different solutions. What Siva offers is what we call a self-contained NPU. It is not an accelerator. If you look at ARM, for example, they offer NPU accelerators,

which you need to combine with an ARM CPU in order for them to work. This is not the case for SIVA. We offer a self-contained, self-sufficient NPU, which is scalable from 200 GOPs up to 400 TOPs for a single core solution. So a very wide range of applications. Low power is built in by architecture, compression,

sparsity mechanisms that I mentioned, and this end-to-end AI software stack. To give some more color perhaps on the differentiation that we offer, it starts with the performance and power efficiency based on our legacy low-power mobile IP. We were born in the mobile industry of the 90s. Low power is very critical for us. High throughput AI, minimal energy, very critical for our MCU customers. Scalability to support all really AI workloads from MCU's up to automotive.

And this fusion of NPU plus DSP to provide the future proofing of the solution for new networks. I think all these elements of the power efficiency, the scalability, and the standalone solution, NPU plus DSP for future proofing is really what puts us apart from other solutions out there. And I gave, for example, the Renesas and Nextchip for Automotive, one of the most critical applications out there and many other markets and segments that we address with our NPUs. So I hope that answers that one. Absolutely. Thank you Moshe for your time. It's always a pleasure to speak with you. Thank you Daniel. My pleasure as well. That concludes our podcast. Thank you all for listening and

have a great day.