I’ve heard it said that “60 is the new 40,” which means that people in their 60s today are often as active, healthy, and youthful as people in their 40s used to be in previous generations. This must be true because, a few days ago, as I pen these words, I celebrated the 47th anniversary of my 21st birthday, and I don’t feel a day over… I’m sorry, what was it that we were talking about?
One of the (few) advantages of growing old is having been around for a long time, seeing a lot of things, and knowing a lot of things about how a lot of things came to be, if you know what I mean.
Take AMD (Advanced Micro Devices), which was founded in 1969, for example. In the early days, AMD’s main products were memory chips and other components for computers. In 1975, the company entered the microprocessor (MPU) market with the Am9080, a reverse-engineered clone of the Intel 8080. The following year, when Intel started to include microcode in its MPUs, the two companies entered into a cross-licensing agreement, as part of which AMD was granted a copyright license to use Intel’s microcode in its MPUs and peripherals.
Starting in 1978, Intel released a series of processors whose names ended in “86” (8086, 80186, 80286, 80386, and 80486), which resulted in people referring to the family collectively as x86. This nomenclature is also used to refer to the underlying x86 instruction set architectures (ISAs).
Back in the 1980s and 1990s, Intel was the dominant player making x86 processors for personal computers (PCs). These were the brains of most home and office computers. At the same time, AMD made x86-compatible processors—essentially chips that worked just like Intel’s but were often cheaper or had slightly different performance characteristics. This made AMD a major competitor to Intel in PC space (where no one can hear you scream).
In hindsight (the one exact science), the competition between Intel and AMD has been a major driver of innovation in the microprocessor industry for decades. The two companies have often leapfrogged each other in performance, architecture, and efficiency. Each time one pulls ahead, the other pushes harder to catch up or surpass them. This rivalry has driven down prices and increased performance per dollar. As a result, consumers like you and me benefit from more cores, higher clock speeds, and more efficient chips at competitive prices.
But we digress… as the PC market matured and started to plateau, AMD began exploring embedded systems—smaller, more specialized computers used inside things like cars, industrial machines, networking gear, and consumer devices.
This shift wasn’t a full exit from the PC market, but rather a diversification. AMD took its expertise in making powerful CPUs and began creating processors tailored for embedded applications, which need reliability, low power, and long product lifespans.
In 2022, AMD acquired Xilinx, a company famous for its FPGAs (Field-Programmable Gate Arrays). These are highly flexible chips used in everything from telecommunications and aerospace to industrial automation and AI at the edge. This was a huge move. By acquiring Xilinx, AMD gained a significant foothold in embedded systems and edge computing and entered the programmable logic market (something Intel had done earlier by acquiring Altera in 2015), which put it into a position to offer heterogeneous computing platforms: CPUs + GPUs + FPGAs, all “under one roof,” as it were.
As a result of all this, today’s AMD is one of the top players in both traditional computing and embedded systems.
The reason I’m waffling on about all this is that I was just chatting with Yingyu Xia (market lead for Smart Cities, Security, and Retail within the Adaptive Embedded Core AI group) and Michael Zapke (Senior Product Marketing Manager, Industrial), both at AMD.
The topic of our conversation was AMD’s family of Kria System-on-Modules (SOMs), where “Kria” is intended to convey the concept of “creativity.”
As an aside, SOM is one of those abbreviations (specifically, an acronym, since it’s formed from the initial letters of a phrase) that can either be spelled out as “S-O-M” (the most common pronunciation), in which case it’s considered to be a non-pronounceable acronym (also called an initialism), or it can be pronounced like “SOM” rhyming with “bomb,” in which case it behaves more like a true acronym, which can be treated as a word. If we say it as a word, then we would also say “a SOM.” If we spell it out as “S-O-M” (ess-oh-em), then we would say “an S-O-M.” The main thing is to be consistent, so pick one and stick to it.
I was just bouncing around the AMD website, as you do, and one SOM that caught my eye was the Kria K26.
The Kria K26 SOM (Source: AMD)
The big silver device in the middle is a custom-built Zynq UltraScale+ MPSoC device in a small form factor targeting vision AI, robotics, and industrial applications. A few of the features this bodacious beauty boasts are as follows:
- Application Processor: Quad-core Arm Cortex-A53 MPCore
- Real-Time Processor: Dual-core Arm Cortex-R5F MPCore
- Graphics Processing Unit (GPU): Mali-400 MP2
- Video Codec Unit (VCU): Up to 32 streams
- On-chip SRAM: 26.6Mb
- Too many other functions and interfaces to cover here
Moving off-chip but still on-SOM, there’s also 4GB 64-bit DDR4 and 16 GB eMMC. The scary thing to me is that, compute-wise, just one of these tiny SOMs would totally overpower one of the humongous mainframe computers I was involved in designing circa the early 1980s.
Enterprise-level companies may prefer to use a chip-down approach, which means taking a Zynq UltraScale+ MPSoC and creating a custom board around it. By comparison, smaller companies will greatly benefit from using an SOM in which all the tortuously tricky design has already been done. These SOMs can be mounted on carrier cards created by AMD, or users can create their own custom carrier cards as required.
Kria SOMs are being used in a wide range of machine vision, artificial intelligence (AI), robotics, and industrial applications. For example, Yingyu and Michael were telling me about a company that is using them in an underwater computer vision fish farming deployment. Another company is using them to observe the tables in casinos, watching for any jiggery-pokery or sleight of hand.
A really good example of the sort of thing you can achieve here is provided by the folks at Solectrix GmbH, an industrial solutions provider, who have successfully deployed an AMD Kria SOM to create an innovative vision processing system for mobile industrial machines. This system allows machines to “see” and react to their environment in real-time, even in harsh industrial settings.
Some machine vision applications require only a single camera. Others can demand twenty or more. The guys and gals at Solectrix identified a range of use cases requiring four cameras. Furthermore, they determined that a Kria SOM provided them with just the capabilities they required.
Just to make sure we’re all tap-dancing to the same skirl of the bagpipes, GMSL (Gigabit Multimedia Serial Link) is of great interest for automotive and industrial applications because it’s a long-distance, high-bandwidth data link for transmitting video, audio, control signals, and power, all over a single coaxial cable or shielded twisted pair (STP). It’s designed for environments where you need to transmit uncompressed HD video with low latency, handle long cable runs (up to 15+ meters), and ensure reliability in noisy conditions.
The reason I mention this here is that, rather than using a carrier board from AMD, the chaps and chapesses at Solectrix created their own custom board, allowing them to support up to four GMSL2 interfaces.
The result is the SXVPU (Smart eXtensible Vision Processing Unit). This little beauty can be used to provide machine vision intelligence to a wide range of applications. Cast your orbs over this video and accompanying case study, for example.
My understanding is that the dudes and dudettes at Solectrix currently export the captured video over Ethernet to be processed externally, but there are many other options that could be employed, such as mounting an M.2 accelerator card with a PCIe interface in the same box as the Kria SOM, for example.
All I know is that the power and performance of AMD’s Kria SOMs leave me chortling in delight. I could never have envisaged anything like this back in the dark days we used to call the 1980s. How about you? Do you have any thoughts you’d care to share with the rest of us?
