Mythic CEO Taner Ozcelik and Videantis Managing Director Carsten Reuter discuss why analog computing needs a digital partner, the potential for 100× efficiency improvements over GPUs, and an architecture designed to scale from data centers to cars and robots.

Mythic CEO Taner Ozcelik (left) and Videantis MD Carsten Reuter. (Source: JPR)
Jon Peddie Research’s David Harold spoke to Mythic CEO Taner Ozcelik and Videantis Managing Director Carsten Reuter about their recent partnership and why they are combining Mythic’s analog and Videantis’ digital computing, the advantages resulting from the partnership, company and product advances, and more.
Let’s talk about how you bring together this Mythic analog very distinctive architecture with the Videantis digital product.
Taner Ozcelik: A match made in heaven.
Our premise in analog is energy. Our architecture is designed to deliver a great deal of benefits, and energy efficiency is the number one benefit. It also has a cost and latency advantage. The drawback of analog, however, is you can’t do everything on it. It’s a matrix multiplier at the moment. Well, that happens to be 95% of the problem, but you need to handle the remaining 5% without sacrificing that energy efficiency.
We had to pick a partner that matches really well. That’s why our solution with Videantis is key—it’s important to have a hybrid computing architecture.
Carsten Reuter: We started with multi-standard video compression. We had a programmable hardware architecture, capable of several standards on the same hardware. That was done at that time on different pieces of specialized fixed-function hardware, but we could do it on a single piece of hardware, let’s say, and with more flexibility. This was something nobody could believe, but we then demonstrated it on a chip.
Then it evolved through video compression and image processing into computer vision. Additionally, there were customer requests to open up our development software, so we also provided an SDK that customers could use for programming. At the beginning, this was just a huge effort for us, so we didn’t open up. It was kind of a black box for the customer.
So for the customer, everything happens at the software level. They don’t have to worry about this whole VLIW, SIMD, unusual architecture?
Reuter: We shielded this from the customer. Still today, for the CNNs and the large language models, we provide highly optimized library functions for the customer, and also the mapping flow.
Because if you start from a network, you have to do some preprocessing, or you have to make informed decisions for the customer to preconfigure things. Otherwise, the efficiency is not good enough. We know our architecture really in detail, and this is based on many years of expertise. We had the same topic with computer vision. Our most successful product is a rearview camera product, which is running in over 30 million cars. The chip was made by one of the largest semiconductor companies , and we licensed our architecture. On top of that, we sold the software.

Figure 1. Mythic demo system. (Source: JPR)
What did that look like when you say you “licensed?” Is that a kind of licensable RTL sort of thing, or did you design the actual chip and then they manufactured it?
Reuter: For this company, we licensed the RTL, but it was not like, ‘Here, use it.’ We did not throw it over the fence. We supported customers closely.
We always provide pre-placed netlists, because then you need to know not only where the processing is, but also where the memories and the connections are.
We did our own sample chips as research products, let’s say. We taped-out, I think, three chips or so over time. For such a small company, we were about 20 people, and this was always a huge effort for us to do. But we were able to convince customers.
And so, in this rearview camera situation, the competition was mostly against GPUs or against other video streamers?
Reuter: The thing is, there was no solution at all. Because nobody could do it at that price tag. We are ASIL-qualified, so this was also an effort together with Continental, the Tier 1 that pulled the camera together. A huge thing. It took us two years, and they contributed significantly. The camera could do AEB (Automotive Emergency Breaking) and consumed less than 1W and is the size of a 20×20x×0mm cube.
Making the hybrid architecture work
While we are still talking about the history with Videantis, what about the software element of that? Originally your architecture was kind of black-boxed, and people just approached it through the software. Does that software history, that established stack and experience, carry on into helping Mythic?
Ozcelik: There are basically two ways to look at it.
Mythic was founded and built this little analog chip. It contains a lot of flash cells that basically do the heavy compute on the matrix side. The digital side had to have some software that can interface with the rest of the world very effectively, frictionlessly.
The company had done some great things, but we doubled down since I came on board, about two years ago. We now have a stack that can take any code that’s written for GPUs or any other architecture in PyTorch or TensorFlow, or a network where the common interchange format is ONNX, and just run it on the device, whatever the CPU is.
Software is something we pride ourselves heavily on, for two reasons.
First is the compiler is insanely efficient. The idea there is that you have this parallel structure in tiles of flash cells that do the matrix compute, and you want to effectively map, at compile time, the whole data, the whole network efficiently, so you utilize it to the maximum.
You want to get to something like 70–80%. That took time. Now it’s there. The compiler is incredibly efficient.
Second, dataflow and how we map the network also changed over time, which is also a crown jewel of the company. That basically means we can take these chiplets, put them side by side, and scale up. We plan to implement 3D stacking in the future too, which will enable even more performance out of the same footprint That will allow us to map the network very efficiently, and then put it into a data center.
We do everything that’s not matrix multiplication on the digital side. Today, it’s on the M1 chip, on a bunch of RISC-V processors with some accelerators around it, approximately 80 of them on the chip. But the architecture on M1 did not keep the energy efficiency and show the chip in the best way possible. There were bandwidth issues, the NoC wasn’t done right. It was never simulated properly, in my opinion, and software was not done before the chip was taped-out.
Fundamentally, the analog tile architecture has to connect with digital, and that digital has to communicate with the analog in the most efficient way, plus communicate with the rest of the world. That’s the challenge we’re trying to solve.
Reuter: It’s quite natural, let’s say.
Ozcelik: Exactly. We’re just another accelerator on Videantis’ architecture. They have a VLIW ISA (Instruction Set Architecture) with SIMD (Simultaneous Instruction Multiple Data) processors hanging off of it. And they have a scalar processor, like a tiny little CPU. SIMD processors comprise a vector processor, and a wide vector processor. We’re like the ultra-wide vector with matmul (matrix multiplication) processor.
Reuter: And the scaling method is really the same.
Ozcelik: Exactly. That’s why it’s a perfect match. It’s like, holy sh–t.
Where are you making them? Who’s the foundry?
Ozcelik: We plan to build it in two separate places. The analog is in Dresden, Germany. It’s 28nm process technology, in GlobalFoundries, and because it is an old technology, the capital cost is already depreciated, so the process is very cost effective. The other one is planned for 5nm. It starts in Taiwan but in the future will be made in Arizona.
What’s the implication for your compute-in-memory as it’s done today? What happens as you move down the nodes? You can’t retain flash, right?
Ozcelik: We do have basically at least two, maybe three generations on the flash side, in geometry at least. Each of these will last us two or three years, maybe four years.
That entire architecture, the analog memory portion of it, is like 15% of our entire die. It’s not a big component. It’s what goes around it. ADCs and DACs, especially ADCs, which are the friction to the digital world, are where we spend a lot of the die size.
On the memory side, we looked at several technologies for what happens after, say, flash. Flash runs into quantum tunneling effect, around mid-teens. The best one right now for us to use next is resistive memory, or ReRAM.
We now have two R&D projects with leading foundries in the world because they’re both investing in this technology and asked for our help. So we’re working with them to pack compute into resistive memory as well. It may not be the same way we do it, but that’s what’s going to happen. Then we can go down to single digits or even less in geometry.
There’s also 3D stacking that we haven’t enabled yet. The reason is our analog wafers are so low power that we can actually put them side by side, vertically. There are many ways to scale this. But even today, our out-the-door advantage is like 100×. And that’s going to last for a while.
We believe our competition is running into a wall because Nvidia touted 4×, 5× at every generation in performance per watt, which was mainly due to the fact that the bit precision required for AI models has been coming down in every generation. AI came Nvidia’s way. They got lucky. It was 64-bit in Kepler, 32-bit in Volta, 19-bit in Ampere, 8-bit in Hopper, and now 4-bit in Blackwell. Maybe you have two more generations after that and you are done, 2-bit and 1-bit. But it is very questionable whether the 2-bit and 1-bit networks will work.
So we have our competitor basically coming to a screeching halt, and we have lots of room to go. We’re starting with 100× better efficiency. I don’t think scaling will be a problem for us, in my opinion.
What does the interface look like between your analog realm and your digital realm today? Are there other ways you can improve that going forward?
Ozcelik: There’s something very interesting there: hybrid bonding. Again, the two wafers are separate. The reason is we wanted to get the best of both worlds on both silicon technology as well as architecture.
If I were to constrain them both to 28nm, we wouldn’t be able to find some of the IPs. And the digital side wouldn’t be as power efficient. It wouldn’t have as much performance.
Hybrid bonding is now in all the sensors. I went to Onsemi after I left Nvidia and I learned that basically every sensor nowadays is hybrid bonded, wafer-level stacked.
Image sensors have the same problem: photodiodes, which are analog, have to be at the older geometry, like 65, 55, maybe 45nm these days, and the digital can be anything, like 16 FinFET, 12, and below. So you can bond them together.
Bonding technology and separate wafers, 3D stacking, have come a long way. Not many people know sensors are built this way. It is a $30 billion industry, and most sensors these days are built stacked.

Figure 2. Mythic in action. (Source: JPR)
You were just saying, out the door, 100× performance-per-watt improvement. What’s the basis for those comparisons? How much of the performance advantage is demonstrated in silicon or simulation?
Ozcelik: The first chip that I showed you now has up to 100× in performance per watt per cost, not performance per watt. That is the real chip, M1, and it is in production. Our 100× for Vanguard, which is our new chip, is performance per watt only. In performance per watt per cost, it’ll become maybe 1,000×. The current one is 100× in performance per watt per cost. A third of it comes from power consumption.
And we can demonstrate that benefit on Jetson, by the way, which was my old product at Nvidia. Comparing without and then with Mythic, we’re getting 10× more inferences on the same code, With Mythic, we also get twice the frame rate, one-third of the power, and our chips are approximately one-tenth of their cost. So 100× and more, actually, today, is real on our first generation.
Simulation is something that I’m incredibly big on. At almost every meeting, I talk about modeling, simulation, emulation, verification. Nvidia basically builds chips today that are so good they don’t have to see the chip come back to order the mass-production wafers. Imagine the risk-taking there. But they have perfected the art of chipmaking.
Videantis have zero returns on 30-plus million chips shipped with their IP, which is impeccable. They have been able to pack the entire compute of AEB, automatic emergency braking, which is basically object detection, into a single 20×20×20mm cube.
So the digital part is also low power, the analog part is insanely low power, and then we’re simulating the hell out of it like Nvidia does.
Currently, for instance, we’re running BEVFormer, which is one of the most challenging models out there. In bird’s-eye-view Former, we are running 6six cameras, HD resolution, 1600×900, running at 30 fps, and consuming only about 4.4 W. [And] 98.3% is in analog TOPS. The rest, 1.7%, is in digital.
But that’s all simulated, basically based on the architecture that we have.
RTL for analog wafer is done. We’re in the last phases of tape-out, so we are doing placement and route, and then will tape-out. Analog wafer is all SPICE modeled, real-number modeled, and simulated. At the full chip level, we do cycle-accurate simulations, functional simulations, UVM, and run emulation. We’re using Cadence Palladium systems to do the emulation.
Why analog for AI?
Why were you convinced that Mythic was taking the right direction, that we should be doing something fundamentally different from a GPU for the problem?
Ozcelik: It’s just this 80-year-old von Neumann architecture… sure Nvidia made it live on. He made an 80-year-old like a muscly Rocky. He brought massive parallelism to a processor architecture, gave birth to new speed and transformation of Moore’s Law. If that didn’t happen, we wouldn’t have AI. We wouldn’t have discovered the power of compute that would translate to these algorithms.
We figured out how the neurons are connected to each other, the layers, and what they do. And the algorithms are getting massively better. But the architecture underneath is 80 years old.
The perceptron [an algorithm for supervised learning of binary classifiers] was built on an analog computer—Rosenblatt, 1956. So the world’s first AI was done in analog. It’s been around for a while.
Why did I do this? It’s incredibly different. I wouldn’t do Cerebras, personally. I would not do d-Matrix. If they asked me to run any of those companies, I’d say it’s just a futile effort, because Nvidia is going to crush them over time, no matter what. It’s all digital. Yes, you can pack more SRAM. Yes, you can go wafer scale. But he knows that too, and they are easily achievable. But this is so different, and it takes so long and so much effort to do this that we have a natural moat. We have patents too, now almost 100 with everything that’s in the filing process as well. This is a result of 10 years of hard work, with lots of scars.

Figure 3. 130 years of Moore’s Law. (Source: Mythic)
The whole compute-in-memory analog architecture is really, really interesting. But analog obviously has its challenges around noise and precision and calibration. What about those limitations? How are you overcoming those? Are there other kinds of AI that care less about those analog limits or something that’s more inherently suited to it?
Ozcelik: To your point, it’s not precise enough that I can replace a Navier–Stokes computation for a computational fluid dynamics algorithm that’s operating at 64 bit with double precision. That’s not going to happen anytime soon, right? Meaning deterministic algorithms that people have built and things like that. But AI is probabilistic computing. It’s not deterministic computing. All you care about all the time is the probability of the next token. It could be a vision token or a text token or whatever token. You’re basically predicting the next token. That’s the number one observation.
Second, in the digital world, like with Nvidia’s FP32, you effectively end up optimizing the network before you fit it into a basic footprint. For instance, you do quantization, post-training quantization, as everybody does. And every inference now runs at 8-bit or 4-bit, because the accuracy is just as good, and it is a lot more efficient. That’s why they have come down in precision.
Plus, you effectively prune the networks. Then you do distillation to make it even more compact. In fact, there’s a move now in the number of parameters coming down as opposed to going up, up, up.
But because our of performance-per-watt advantage, we don’t have to do those, or at least all of those optimizations. Those optimizations can reduce accuracy. We demonstrated this in a YOLO model using mAP 50:95. And because we have so much performance-per-watt advantage, we don’t have to do those optimizations, our accuracy is now higher in INT8 than Nvidia’s FP32. Effectively, we can reverse the equation because we have so much performance-per-watt advantage.
So, noise doesn’t matter, right? But it’s only for AI today. The rest of the compute problems, over time, I’d say, can analog get there? I think so.
But there’s another massive trend in our favor, which is that everything is going towards AI. Even Navier–Stokes, even crash simulations tests, I think, are going to be all AI—climate is already AI. It used to be all non-AI.
You’ve both got automotive history. We’re here at a physical AI conference. Is that the sweet spot on which you want to be focused? Even your physical implementation, chiplets, is very suited for that kind of area.
Ozcelik: Data centers are the biggest problem, and we believe we can get to revenue fastest in data centers. Robots also have a long cycle; defense, a long cycle; automotive, a long cycle. Having said that, we are going after all of those four markets with the same architecture.
We have revenue right now, by the way. This year we’re doubling our revenue: $17 million versus $7 million and change last year. It’s mostly development money, so there are no chip sales yet.
Moving forward, we are focusing on data centers first as the fastest route to revenue. That’s our primary focus. But we’re building the chiplets in a way that single instances, or a few of these designed together, can build a system to address the robotics world. Our physical AI focus will be all about robotics, whether it’s a robot doing factory automation, a humanoid, or AMRs or anything else, or cars, which are also robots on wheels, or drones for defense, i.e., robots in the air. These robots all use basically the same models: world models, VLMs, VLAs, JEPA, etc.
So we are going after all markets simultaneously. That means we can do both edge and enterprise with this unique architecture at the same time.
What about the physical drivers of this as well? We had this debate with cars, centralized smart versus zonal, whatever. But it feels like you can’t really have that debate with robots. With robots, inherently, you’ll need some intelligence at the sensor level and all these things, joints and fingers.
Ozcelik: We couldn’t have lifted this company without Honda, because our visions were 100% aligned.
They showed me a slide that could pass as a Mythic slide. “Von Neumann is dead. We need a more brain-like architecture, so we need to go neuromorphic.” And as you know, as we announced, we are working on a central computer architecture with them.
However, we also came up with another idea for the robots. Basically, the idea is, Why don’t we put your little, tiny things, processors, into the sensors around the robot? As a result, we have a couple of projects around that idea. So, yes, you can do distributed computing with our technology because we can enable material compute in a very small footprint.
So, you can do some edge compute, create some intelligence, and then do better and more in the central processor. I believe our technology is perfectly suited for those types of architectures. This will particularly be useful for edge computing in the cars, near the sensors.
Reuter: This was the same argument for our rearview camera, because it’s exactly the same. You create more intelligence at the edge, at the sensor side. Because then, you have just this tiny plug and this cable, and everything is fine.
Ozcelik: The other thing is sensors are all pretty much analog because they interface with the world. The world is analog, right? We always transfer or convert them to digital after sensing because compute has been all digital until now.
So, what if you just pack analog compute into a sensor that’s already analog? You could have the world of computing and sensing all in analog someday. Literally, this can happen.
Videantis
So, you bought Videantis. What changes?
Reuter: We didn’t change anything.
Ozcelik: They didn’t change anything. I don’t want them to change anything. We just take their technology and integrate it together.
It’s not like they just give us the IP and we’ll go integrate. Basically, they’re leading the next generation’s digital die. Digital die, in TSMC’s 5nmprocess, basically, is driven by Hanover. We believe that digital die is going to be a product by itself, and we are designing it accordingly.
The beauty of having that as a separate die is I can actually go to market with it.
Our next generation processor, Vanguard, is designed to bring the digital die, called Minsky, and analog die, called Rosenblatt, together. These digital dies will also be produced as chiplets.
Sometimes you’ll need more digital, and other times, more performance per watt, maybe, which is where the analog comes in. You’ll have a mix and match of these chiplets through the same die-to-die interface.
Another thing I want to highlight is that the reconfigurability of Videantis’ architecture is insanely good, meaning you can, at runtime, change the workload very effectively from AI to deterministic computing, and vice versa.
Second, they’re also designed to be very die-efficient. According to their benchmarks against GPUs, performance per cost is about five times higher.
There are other things they are good at, like scalability and performance per watt. It’s a really good architecture. In fact, once we are done with this chip, we’ll launch a separate product. It’ll go toe to toe with GPUs by itself. That’s why I’m super excited about that. That’s why I asked them to do the entire architecture for the digital die. We didn’t take their processor and embed it in some corner of the Mythic chip. They’re doing half of the chip, Vanguard.
If I was smarter, what would I have asked you?
Ozcelik: Where do we think we’re going to be in five years? When the chip comes back, my goal is to show it in three, maybe four, different, completely different, things on day one. That’s the beauty of architecture.
It’ll be in a car, and that’ll be Honda working on that. It’ll be in a robot, and we’re working closely now with a customer on that. Perhaps in a drone. Then in a data center. The discussions are getting to an advanced stage.
If we can achieve that, it will be, I think, shocking to the world of AI compute. One chiplet in many diverse applications.
If we do our job right, it should feel very different in five years. Analog computing should be a thing that people talk about. APUs should become, hopefully, a little bit of a household name. That would be my dream. One little chip powering four completely different machines, on day one of its birth. That will send shockwaves to the industry.
What do we think?
The AI processor market is crowded, with companies exploring architectures ranging from GPUs to analog computing. Mythic stands out for its flash-based analog compute-in-memory technology, which promises substantial improvements in energy efficiency by reducing the movement of data between memory and processing elements.
It is a high-potential but technically challenging architecture. Analog computing is particularly efficient at matrix operations central to AI inference, but it cannot handle every aspect of a workload. An efficient bridge to the digital world is essential, which makes Mythic’s acquisition of Videantis strategic. Its proven digital architecture and software expertise should help Mythic overcome some of the limitations of its first-generation products.
The proposed chiplet architecture also offers flexibility, potentially scaling from intelligence embedded directly into sensors to much larger data center systems, but its ability to support massive generative AI models remains to be demonstrated in silicon. Mythic’s ambitious performance claims for its next-generation Vanguard processor are currently based on simulation.
We see a particularly strong fit for edge AI, where power consumption, latency, and physical constraints matter enormously. Applications in automotive, robotics, and defense could benefit substantially from the technology, especially as more intelligence moves closer to sensors.
The move into data centers is more ambitious, but if Mythic can deliver anything approaching its projected energy-efficiency advantages while solving the challenges of digital integration, scalability, and software support, it should find a place in an increasingly heterogeneous AI computing market.
This acquisition reduced the number of active AIP suppliers to 159.
Mythic acquired German processor IP developer Videantis in May 2026 through a combination of cash and equity. The companies did not disclose the purchase price.
AND IF YOU LIKED WHAT YOU READ HERE, DON’T BE STINGY, SHARE IT WITH YOUR FRIENDS.