遥遥领先Renesas瑞萨有源晶振常用于人工智能

遥遥领先Renesas瑞萨有源晶振常用于人工智能，对于瑞萨来说，人工智能正在成为游戏规则的改变者，尤其是在边缘领域。事实上，尽管有关于生殖人工智能和ChatGPT-4的所有传言，但2025年前创造的所有数据中约有75%将来自网络边缘——而不是云。与此同时，企业收集的90%的数据如今都被丢弃了，这为找到这些数据的生产性用途提供了一个令人振奋的机会。

瑞萨适合做什么？当我们想到计算时，我们会想到MCU、MPU、CPU和GPU。

创新的硅架构是快速处理数据、低延迟和低功耗的关键。瑞萨OSC晶振提供一系列计算设备，从电池供电的MCU到基于Linux的MPU。密集型人工智能应用被卸载到具有高功耗比性能的板载加速器上。对于大规模和MLDevOps的模型部署和管理，瑞萨还提供云连接堆栈。

除了计算，瑞萨还为我们的GPU客户和合作伙伴提供内存接口、高效率电源和定时芯片，所有这些在处理大型语言模型所需的大量数据时都是关键。

欢迎人工智能生产力列车

不管具体的重点是什么，一个成功的人工智能结果取决于你的客户对他们自己的数据集的了解程度。如果客户需要大量的指导，我们为他们提供从培训模型到实施的一切。另一方面是非常精明的客户，在这种情况下，我们更多地采用“免费增值”模式。还有一些客户介于两者之间。

也就是说，在整个企业中，只有54%的项目包含了人工智能。另外46%的人没有这么做的原因是因为与部署人工智能相关的挑战和复杂性。想想看，如果我们让用例变得足够容易理解，我们会有多高的生产力。这就是TinyML这样的东西发挥作用的地方。

正如Evgeni所说，“TinyML生态系统是一个完美的雷达屏幕，因为我们有100多家成员公司和15，000名员工在世界各地工作。五年前，这还是一个概念验证，但五年后我们将看到人工智能技术掌握在消费者手中，帮助他们解决各种问题，有源晶振产品更加适合用于人工智能。”

新思科技的Shankar指出，人工智能将在帮助公司降低设计复杂性和更好地管理工程预算方面发挥作用，因为从5纳米到3纳米工艺节点的成本增加了一倍以上。Shankar表示，Synopsys计划通过将人工智能应用到EDA堆栈的每一层来优化其客户知识产权，从设计到验证和测试，并已经提供了一个软件即服务模型，允许客户访问云中的人工智能服务。

“我们看到人工智能技术的采用速度非常快，这主要是因为我们的客户看到了这个生产力问题，他们正在将人工智能的采用作为一个优先事项，”他说。“我们已经完成了超过250个生产设计，在过去的12个月里，这一增长速度真的很快。作为工具供应商，我们必须解决‘让人工智能更容易被采用’的问题，但设计界对它有一种渴望。”

“人工智能正在以多种方式推动我们的技术向前发展，”艾拉说。“过去，作为测试设备供应商，我们必须比正在测试的设备更快、更准确。现在，随着人工智能芯片的出现，我们也必须变得更聪明。这导致我们以多种方式使用人工智能，包括边缘计算和我们在测试期间实时运行的机器学习算法。人工智能将成为许多事情的核心，我们的工程师将不得不接受这一点，并学习如何将它作为一种基本工具来使用。”

“人工智能正在以多种方式推动我们的技术向前发展，”艾拉说。“过去，作为测试设备供应商，我们必须比正在测试的设备更快、更准确。现在，随着人工智能芯片的出现，我们也必须变得更聪明。这导致我们以多种方式使用人工智能，包括边缘计算和我们在测试期间实时运行的机器学习算法。人工智能将成为许多事情的核心，我们的工程师将不得不接受这一点，并学习如何将它作为一种基本工具来使用。”

专题小组成员从左至右:高德纳公司的高拉夫·古普塔；瑞萨的赛莱什·奇蒂佩迪；新思公司Shankar Krishnamoorthy艾拉·莱文塔尔，阿德万斯特；叶夫根尼·古塞夫，TinyML基金会

For Renesas, artificial intelligence is becoming a game changer, particularly at the edge. In fact, despite all of the buzz circulating about generative AI and ChatGPT-4, about 75 percent of all of the data created by 2025 will come from the edge of the network – not the cloud. At the same time, 90 percent of data collected by enterprises gets discarded today, which opens an exciting opportunity to find productive uses for it.

Where does Renesas fit in? When we think of compute, we think of MCUs, MPUs, CPUs and GPUs.

Innovative silicon architectures are key to processing data quickly, with low latency and low power. Renesas supplies a range of compute devices, from battery-powered MCUs to Linux-based MPUs. Intensive AI applications are offloaded to onboard accelerators with high TOPS/watt performance. For model deployment and management at scale and MLDevOps, Renesas also provides cloud connectivity stacks.

In addition to compute, Renesas also supplies memory interfaces, high efficiency power and timing silicon to our GPU customers and partners, all of which are key when it comes to processing huge amounts of data needed for large language models.

Welcoming the AI Productivity Train

Regardless of the specific focus, a successful AI outcome depends on how educated your customers are about their own data sets. If the customer requires a lot of hand-holding, we provide them with everything from the training models to implementation. On the other side are the customers who are extremely astute, in which case we work with more of a “freemium” model. And then there are customers that fall in between.

That said, across the enterprise, only 54 percent of projects that incorporate AI see the light of day. The reason the other 46 percent don’t is because of the challenges and complexities associated with deploying AI. Think about how much more productive we could be if we made the use case easy enough to grasp. That’s where things like TinyML come into play.

As Evgeni noted, “The TinyML ecosystem is a perfect radar screen, because we have over 100 member companies and 15,000 people working all over the world. Five years ago, this was a proof of concept, but five years from now we will see AI technology in the hands of consumers to help them solve all kinds of problems.”

Synopsys’ Shankar called out the role that AI will play in helping companies reduce design complexity and better manage engineering budgets, as the cost to move from a 5nm to a 3nm process node more than doubles. Shankar said Synopsys plans to optimize its customer IP by applying AI to every layer of the EDA stack, from design through verification and test, and already offers a software-as-a-service model that allows customers to access AI services in the cloud.

“We are seeing extremely fast adoption of AI technology, largely because our customers are seeing this productivity problem, and they are driving the adoption of AI as a priority,” he said. “We have over 250 production designs already completed, and that ramp has really shot up over the past 12 months. We have to solve the ‘making AI easier to adopt,’ problem as a tool vendor, but there is a hunger for it in the design community.”

In the realm of semiconductor test, Advantest must ensure that the equipment it builds is able to outperform the devices they’re testing. This includes using AI to test AI-based chips, which demands additional computing capabilities, and more recently, machine learning algorithms that help make immediate decisions across the entire semiconductor value chain.

“AI is driving our technology forward in a number of ways,” Ira said. “It used to be, as a test equipment vendor, we had to be faster and more accurate than the devices we’re testing. Now, with the advent of AI chips, we have to be smarter, too. That’s led us to use AI in a number of ways, including edge compute and machine learning algorithms we run in real-time during test. AI is going to be at the core of many things, and our engineers are going to have to embrace that and learn how to use it as a fundamental tool.”