Reports indicate a $30bn (€25bn) deal for AMD to buy FPGA pioneer Xilinx is well advanced and set to be announced next week.
Unlike the 2015 purchase of Altera by Intel, the Xilinx deal is very much driven by the data centre and puts AMD on a collision course with Nvidia, which is pursuing its own $40bn takeover of ARM.
AMD has had significant success with its latest EPYC x86 architecture in the data centre and high performance computing. The chips are being used in several massive supercomputers, inlcuding a 2 exaflop system being developed by HP.
But one area that it hasn’t managed to gain as much traction as it would like is in machine learning. Buying Canadian graphics chip designer ATI Technologies in 2006 for $5.4bn gave it the Radeon GPU. But that has been focussed on desktop graphics and integration with the x86 chips, rather than the booming AI market in the data centre that Nvidia now dominates with its GPUs.
Similarly Intel has been using the Altera FPGA technology it acquired for $16.7bn for its push into 5G rather than the data centre.
Xilinx however has created a significant business in HPC and the data centre, from accelerating automated trading algorithms for banks to medical image analysis, as well as 5G open RAN. It has also been driving into the AI accelerator market, both in the data centre for training through Amazon’s AWS F1 instances, and at the edge for inference. Using FPGAs avoids the eye-watering cost of AI chip development at 5nm.
But Xilinx recognised that having to use RTL to implement the logic in its devices was holding it back, so the Vitis unified development environment launched in October last year is a key step in opening up the use of FPGAs. Smart engineers