AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_01
At Hot Chips 2024, we will be talking about the AMD Versal AI Edge series Gen 2. We first covered the Xilinx Versal AI Edge in 2021, and now there is a Gen 2 offering.
Please note that these articles are published live during the presentation, so pardon any typos. Yesterday, I created 2 weeks worth of STH content.
AMD Versal AI Edge Series Gen 2 for Vision and Automotive
A big part of the value proposition of AMD FPGAs/SoCs is the ability to run the complete workflow, from pre-processing the data to running inference to post-processing the results, with low latency.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_03
The Versal AI Edge Series Gen 2 has a similar value proposition to its predecessor: replacing multiple chips with one.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_04
Here's the block diagram: AMD is upgrading its Xilinx-derived AI Engine to deliver better performance per watt.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_05
Here is the product listing for AI Embedded Systems, some SKUs include additional Arm Cortex-A78AE cores and video/image processing tiles.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_06
AMD has a new generation of AI engine, this one called AIE-ML v2.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_07
This is a compute upgrade to the AI Engine, where AMD is doubling the performance for formats like INT8 and bfloat16, but it also supports new memory formats.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_08
AMD also has MX9 and MX6 data types.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_09
On the processing side, AMD has more powerful CPU cores, real-time cores, and Arm GPUs. If you hear anyone say that AMD doesn't make Arm CPUs, remember this slide: AMD sells Arm CPUs for the embedded market.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_10
The new processing system provides improved performance.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_11
Security in embedded systems is critical because embedded devices have long life spans and are often deployed in locations that are less secure than data centers.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_12
This is the new video processing IP, which is important if you have a camera feed where you're trying to do edge inference, for example.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_13
Functional safety is necessary in some embedded applications, such as automotive, where chips must not fail, or at least cannot be allowed to fail in certain ways.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_14
While automotive AI often gets attention for external features like self-driving, some modern vehicles are also equipped with interior cameras that run AI inference to check on the state of passengers.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_15
AMD has an AI and vision processing pipeline that starts by taking sensor data and pre-processing that data.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_16
The next stage in the pipeline is AI inference.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_17
Once the sensor data is fed into an AI model, the output can be used for some processing – for example, if the cabin camera sees the driver, AI inference can determine that the video feed shows the driver dozing off and a decision can be made to encourage the driver to take a break.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_18
It is also FPGA-based, allowing you to create custom sensor pre-processing pipelines.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_19
AMD also says that multiple modules can run simultaneously on the AI Engine.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_20
Alternatively, you can context switch the entire array between models.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_21
CPU cores can also be used for post-processing.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_22
One example of this is automated parking.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_23
This is the overview slide.
AMD Versal AI Edge Gen 2 Hot Chip 2024_Page_24
For those looking at other areas of application, AMD also has other Versal product lines.
Final Words
Overall, this progress is exciting. The edge is probably the biggest gain in the AI race, with so many applications and use cases that we'll likely see a much more diverse mix of chips than data centers, the PC space, and phones combined.
