AMD Embedded Tour – Massy – June, 11th
The “Embedded” Business Unit of AMD comes along Gaming, Data Center & AI business units.
During the event, the focus was on Adaptive SoC and FPGAs, but embedded processors solutions Ryzen and Epyc were mentionned. It was emphasized that embedded here refers to extreme environment tolerance of these devices.
A review of portfolios
The AMD adaptive SoC & FPGA offer stretches from the “cost optimized” portfolio to the “versal” portfolio. It is worth noting that there are “long term support” families and “medium term support”. Especially the Ultrascale family should is not long term whereas Ultrascale+ is. Long term implies approx 20 years long support.
The cost optimized portfolio comprises Zynq, Artix and Spartan, with many subvariants, depending on the fabric volume and the number of IPs. The Zynq is the CPU+FPGA SoC. The Ultrascale+ improvement in performance comes mostly from etching footprint of 16 nm, associated with FinFET 3D technology. This comes with gains in quantity of hardware elements and consumption.
The versal portfolio is the latest currently available generation in production, and was released in 2021.
The versal product in intended to offer the best signal processing capabilities. A region of special processors called AI Engine (AIE) is added on top of conventional DSP. They support vectorial operations required for instance for efficient Neural Network computations. But other important change in the architectures are introduced. A NOC (Network on Chip) is defined in the hardware to ease the connection of different blocks in now very large chips. A special attention has been paid to the high throughput low latency exchanges between PL and AIE blocks.
Introducing Versal Gen2 & Versal RF
The Versal Gen 2 inherits the architecture of gen1. However, the processing system has many more CPUs and the AIE are upgraded. Some connectivity of the PS that were withdrawn from Ultrascale+ to Versal are reintroduced.
The hardware encompasses high performance bricks for :
- Preprocessing (PL) : Sensor processing / Sensor Fusion / Data conditionning
- AI Inference (AIE) : Analytics / Perception
- Postprocessing (PS) : Decision making / Control / Feedback
On top of the hardware product, the Vitis IDE is adapted to couple efficiently with PyTorch, TensorFlow or ONNX.
