It’s a bit embarrassing to realize that you’ve been wrong about something for several years, but that’s the position in which I currently find myself (don’t tell my dear old mum or I’ll never hear the end of it).
Ever since I was first introduced to the folks at SiMa Technologies (SiMa.ai), I’ve been under the mistaken impression that they are a hardware company focused on making system-on-chip (SoC) devices to power artificial intelligence (AI) and machine learning (ML) applications. These Modalix MLSoCs are purpose-built for vision, perception, and autonomous systems. In a crunchy nutshell, they are the silicon behind physical AI (see also Everything You Wanted to Know About AI But Were Too Afraid to Ask).
I’ve previously written columns on SiMa, all focused on its hardware roots. For example, cast your orbs over the following: How to Build a Multi-Billion-Transistor SoC, Who Needs a Network-on-Chip (NoC), and There’s Exciting News on the Multi-Modal AI SoC Front.
So, you can only imagine my surprise when, in a recent chat with Krishna Rangasayee, CEO, and Manuel Roldan, Software Product Manager, they opened the discussion by saying, “We are an AI software company that builds its own silicon.”
It wasn’t long before we tracked down the source of my confusion. All my previous conversations had been with Srivi Dhruvanarayan, who is the VP of Hardware Engineering at SiMa. As a hardware design engineer myself, I fully understand why Srivi focused more on the hardware side of the fence. In fact, to be fair, I’m pretty sure that most of my questions would have been hardware-focused. But now it’s time to hear “The rest of the story,” as Paul Harvey might say.
SiMa.ai’s focus is to scale physical AI (Source: SiMa)
It soon became clear that the guys and gals at SiMa can summarize their entire strategy in three simple words: Any, 10×, and Pushbutton. The first reflects the company’s determination to support any computer vision application, any neural network, any AI model, any framework, any sensor, and virtually any resolution.
The second refers to the company’s long-standing emphasis on delivering performance-per-watt that is around an order of magnitude better than conventional GPU-based edge AI solutions.
The third—the newest and arguably the most important—is making physical AI development almost “pushbutton simple” through its new agentic software environment. Everything else we discussed during our conversation can be viewed through the lens of these three ideas.
The Hardware Foundation
Before we dive into SiMa’s latest announcements with gusto and abandon (and aplomb, of course), let’s briefly remind ourselves of what lies beneath SiMa’s software wizardry. As I discussed in my earlier columns, Modalix isn’t simply another CPU with an AI accelerator bolted on. Instead, it’s a heterogeneous computing platform in which different processing engines tackle the jobs they perform best.
There’s an ARM processor subsystem for operating systems and application code, a vector DSP for signal processing, a dedicated computer vision pipeline for image manipulation, and SiMa’s proprietary Machine Learning Accelerator (MLA), capable of delivering 50 TOPS while consuming only around five watts.
Rather than forcing every workload through a single type of processor—as GPUs essentially do—Modalix distributes work across these specialized engines. CNNs, transformers, LLMs, multimodal models, traditional computer vision algorithms, sensor processing, and conventional application software all execute where they make the most sense. It’s this heterogeneous architecture that allows the chaps and chapesses at SiMa to claim around an order-of-magnitude advantage in performance-per-watt over GPU-based edge AI solutions.
The important thing to remember, however, is that none of this hardware exists in isolation. Even the most sophisticated heterogeneous architecture is only as useful as the software that allows developers to harness it. Which brings us neatly to the first of SiMa’s announcements.
Palette Neat
One of my favorite comments from Krishna came later in the discussion when he observed, “English is the new C++.” At first glance, that sounds like one of those throwaway soundbites that companies love to sprinkle into presentations. As it turns out, however, there’s more substance behind it than I initially appreciated.
Palette has long been SiMa’s software development kit (SDK). Palette Neat is its new agentic extension, but calling this “an AI coding assistant” would be doing it a tremendous disservice. Lots of companies are currently bolting large language models onto existing development environments and calling the result “agentic.” The folks at SiMa have taken a different approach.
The company has optimized Palette Neat specifically for its own silicon and, just as importantly, trained it in the skills required to build physical AI applications. Rather than trying to know something about every processor ever created, Palette Neat knows one architecture and environment exceptionally well. It understands the Modalix hardware, compiler, SDK, libraries, optimization kernels, debugging tools, and deployment flow. This constrained problem space makes it far less susceptible to hallucinations than a general-purpose coding assistant while making it dramatically more useful for engineers developing physical AI applications.
It’s important to note that Palette Neat isn’t trying to force everyone into a single way of working. Developers who prefer to interact with an AI agent can simply describe their application in plain English. Those who prefer writing conventional C++ or Python are equally well supported. And for fuddy-duddy old engineers like me, who still like to “see” what’s going on, Palette Neat also provides a graphical representation of the application and the data flowing between its various processing stages.
Palette Neat empowers all application developers (Source: SiMa)
What really impressed me was hearing how far the automation extends. This isn’t simply a matter of asking an AI to generate some C++ or Python code and then hoping for the best. Once developers have described the application they want—or written the initial code, depending on their preferred workflow—Palette Neat takes over. It generates the code, cross-compiles it, deploys it to the target hardware, runs it, measures performance, identifies bottlenecks, and iteratively optimizes the implementation. If it determines that a computation currently running on one of the ARM cores would perform better on a dedicated accelerator, it moves the kernel, rebuilds the application, and tries again.
Even more impressively, if the application crashes with a segmentation fault (for example), Palette Neat automatically launches GDB on the target, analyzes the various execution threads, diagnoses the fault, and incorporates what it has learned into the next iteration. It performs the sort of repetitive build-test-debug cycle that every embedded software engineer knows only too well—and it does so without requiring constant human intervention.
As Manuel observed, the whole process can feel almost like black magic. You type a prompt, wander off to make a cup of coffee, and twenty minutes later the system calmly reports, “It’s done.” Not only has it generated the application, but it has also compiled, deployed, executed, optimized, and packaged the result, ready to run.
Perhaps the most compelling aspect of Palette Neat isn’t that it can generate code—plenty of tools now claim to do that—but that it dramatically reduces development time. SiMa boasts examples ranging from a robotics application completed in just two hours to a transportation system developed in a single day, with other customer projects that previously took weeks or months now routinely measured in hours or days.
Dismantling the GPU Moat
The second announcement is just as interesting, although for entirely different reasons. Krishna jokingly described this as “a little eye poke at our GPU friends,” but his underlying point deserves serious consideration. The reason GPUs dominate so much of today’s AI landscape isn’t that they were originally designed for AI. Quite the contrary. As their name suggests, GPUs were created to render graphics. It simply turned out that devices boasting thousands of relatively simple processing elements could also perform neural network calculations rather well. History took over from there.
As a result, countless edge AI systems are built around GPU-based platforms, such as Nvidia’s Jetson system-on-module (SOM). Engineers have designed custom carrier boards around these SOMs, validated those designs, written software, and shipped products. Replacing the GPU itself isn’t difficult. Replacing everything built around it is.
Jetson-based Nvidia SOM (Source: SiMa)
Typical carrier board for Nvidia Jetson modules (Source: SiMa)
All this led the folks at SiMa to ask themselves the question, “Why make customers redesign their hardware at all?” Their answer is the new Modalix SOM. Rather than introducing yet another proprietary form factor, SiMa produced a module that’s pin-compatible with NVIDIA’s Jetson Orin NX family. Existing carrier boards can accommodate the Modalix SoM with little or no hardware redesign, allowing developers to evaluate and deploy SiMa’s heterogeneous MLSoC without discarding years of engineering investment.
Pin-compatible Modalix-based SOM (Source: SiMa)
The SOM itself presents the proven Modalix device in a compact, production-ready format, complete with LPDDR5 memory, storage, high-speed interfaces, and industrial-temperature support. Under the hood, it’s still the same heterogeneous architecture I discussed before; it’s simply been wrapped in a format that makes adoption dramatically easier.
What struck me during the discussion was Krishna’s observation that customers are no longer asking primarily about benchmark numbers. A few years ago, everyone wanted to compare TOPS. Today, he says, the first question is much more likely to be, “How quickly can you get me into production?” That’s a rather telling shift in priorities. Performance still matters, of course, but development velocity has become just as important.
And that’s really the common thread connecting both announcements. Palette Neat removes much of the software complexity associated with building sophisticated physical AI applications. The new Modalix SOM removes much of the hardware complexity associated with adopting a new AI platform. Together, they attack the two biggest obstacles standing between engineers with good ideas and the deployment of products.
All of which brings us back to Krishna’s opening remark. After spending the last few years thinking of SiMa primarily as a silicon company, I now understand why he insists that it’s “an AI software company that builds its own silicon.” The silicon remains every bit as impressive as it always was. But increasingly, the real competitive advantage lies not in the transistors themselves, but in making those transistors easy to use.
If SiMa can genuinely reduce physical AI development from months to days—or even hours—while simultaneously allowing developers to drop its technology into hardware originally designed for GPU modules, then it may well have found one of the shortest paths yet from idea to production.
You must be logged in to post a comment.
Leave a Reply