Companies mentioned: AMD 0.00%↑, TLSA 0.00%↑
Every year at CES, the largest consumer focused trade show of the year host in Vegas, we expect companies to come out with the latest and greatest CPUs and consumer devices, depending on when the design and production cycles align. In recent iterations however, there has been a growing focus on other adjacent segments as well. Opportunities to test the first self-driving cars, for example, came to most media in this industry through CES several years ago. But also the push to speak not only direct to market but direct to manufacturers is a key reason why CES regularly pushes the 200k+ a year mark. That’s both registered attendees and others who simply set up shop on a hotel in the Vegas strip and have business meetings simply because everyone is in the same place.
This year at CES, AMD is focusing heavily on the automotive world. In the Automotive world, it’s important to divide compute in automotive into three segments:
In Car Entertainment (ICE) or Digital Cockpit
Advanced Driver Safety and ADAS (Cameras, monitoring, parking)
Autonomous Driving and Self Driving
AMD has had big success already in the digital cockpit market. The embedded processor division has scored big wins with companies like Tesla to enable better in-car graphics and processing power over other alternatives. Back in November 2021, we discovered AMD’s Ryzen Embedded V1000 series being used to deliver that digital cockpit, and since then those chips have been deployed in millions of vehicles.
This year at CES, AMD is announcing a second generation family of automotive APUs, the Ryzen Embedded V2000A, featuring up to six Zen 2 cores, 7 RDNA graphics units, dual gigabit Ethernet, support for four 4K displays, and AEC-Q100 certification. Users more familiar with the consumer hardware may scoff given that Zen 5 is only around the corner, however AMD and competitors typically have to use proven silicon with long life-times for the automotive space, or use knowledge gained over years to harden it for appropriate qualification. AMD is planning 10 years availability for this CPU range, and positions the V2000A family as more of a mainstream offering, whereas V1000 will sit in the entry level.
The idea for such a chip is that a vehicle could arguably support one display for the driver, and three more displays for passengers which could be used for watching films or playing video games. While 7 compute units for graphics (probably Vega) isn’t going to blow the world alight, it is substantially higher performance than a number of other options in the market. AMD describes this as an ‘AAA Gaming’ experience, however I suspect three people playing Rocket League or some emulators might be more in line with the graphics horsepower. These days I find most of my generation and the younger generation have handheld devices like a Nintendo Switch to pass the time on long journeys.
One of the demos at the event will be from eCarX, showing a liquid cooled version powering an ultrawide display as a digital cockpit.
A Wild Xilinx Appears
While promises of AAA Gaming will light up some column inches, since AMD acquired FPGA designer Xilinx, the amount of AMD chips in a motor vehicle has grown exponentially. Xilinx has been in the game for a while with a large chunk of its portfolio:
Forward Looking Cameras: AMD Zynq
In-Cabin Monitoring: AMD Zynq
Automated Parking: AMD Zynq
ADAS: AMD Versal
Radar and Lidar: AMD Zynq
Zynq is a relatively legacy product that fits the job in those areas really well, but it’s the ADAS side of the equation that is getting updates this year at CES. AMD is updating its Versal family with a new automotive grade version of its Versal AI Edge SoC.
What exactly does this mean? Versal is not a traditional FPGA – it’s what Xilinx and now AMD call an ACAP, or an Adaptive Compute Acceleration Platform. That’s about as clear as mud, so let’s go in with a diagram to show the difference.
From left to right:
Artix: A pure FPGA. On this chip is programmable logic, with some small amount of IO.
Zynq: Take an FPGA and add in some Arm cores, and perhaps a dedicated memory interface or PCIe interface. This makes the SoC capable of some directed compute by having a hardened core, and advanced connectivity with the outside world.
Versal: An FPGA on steroids. Alongside the Arm compute cores, of which there are now more, there are also now AI Engines. These AI Engines give ASIC-like performance in dedicated workloads. More of the die space is dedicated to low latency interconnect IP, as well as security, video, and software support.
Now with that breakdown, the new Automotive Grade (XA) Versal AI Edge series processors are meant to bring a lot of that dedicated performance and flexibility to the ADAS world.
The reason you wouldn’t use a pure CPU here like the V2000A above is that over the lifetime of the vehicle – such as new revisions or new techniques – having programmable logic allows the processor to adapt over time and bring a faster time to market without having to rely on hardware limitations. By this I mean keeping latency low, which FPGAs are really good at, whereas a modern core 10 years down the line might not be able to keep up. AMD has the relevant standards required for ADAS on this chip, and with the big focus on how machine learning can help with things like safety or parking, support for developers is also growing leaps and bounds with the focus on open AI model enablement.
Versal is also committed to features like sensor fusion – being able to take multiple inputs from cameras and other sensors to build a picture around the vehicle before deciding what to do next. As you might expect, there are a range of options depending on price, performance, or requirements.
For those interested, here is the specification sheet.
There’s an * there on the 1752 model, saying that the AI engines in that chip are different. In all these chips except one, the AI engine is actually a version called AIE-ML, which has additional cache for machine learning tasks usually suited for video. The VE1752 uses the first generation regular AIE engines without that cache. When I asked AMD why this one exists, they said it was due to customer requests with ongoing engagement. AMD is also listing the VE1752 as the lead device, showcasing that it’s probably the one first out of the fab in volume due to the simpler design.
The dev kits for these devices will be available now, with the VE1752 coming in volume in Q1 2024. The rest of the family are set to be available later this year. That’s pretty quick for an FPGA family! But it does mean with development time cycles, they won’t actually be on vehicles until late 2025. Probably.
The main demo at CES with the Versal chip is the one in the middle here – road and vehicle detection with machine learning inference based on four 3 megapixel cameras at 30 FPS. Three megapixel is 2048x1536, so this means above Full HD (1080p) cameras.
Other AMD at CES
While those are the big announcements, AMD recently launched the MI300X and new Ryzen Mobile 8000 series APUs. While I’m not expecting much in the way of MI300X talk, the new APUs should be in devices to observe, even if they’re not ready for retail yet. AMD doesn’t have a big keynote this year – if they did I’m not sure Automotive would be high on the list - but they do have a press reception, where I expect some of those devices to be.
I actually need to rekindle my relationships with Dell, HP, and Lenovo, and others in order to go visit their showcases – that’s on my to do list before I fly out in a couple of days. (If you’re from the PR teams of a company showing Ryzen 8040 hardware, please reach out.)
For anyone going to CES and wants to talk silicon, let me know! I still have time slots available, mostly on the 7th.
I plan to have more CES content on More Than Moore, and my YouTube channel. Subscribe to get these updates!