AMD is working to establish its growing GPU portfolio as a viable data center solution alternative to NVIDIA.

AMD's $5.4 billion purchase of ATI Technologies in 2006 seemed unlikely to succeed, as the companies operated in separate markets and across opposite coasts, with ATI based in the Toronto, Canada region and AMD in Sunnyvale, California. However, the acquisition proved transformative. The graphics business kept AMD afloat while its Athlon/Opteron CPU business struggled, with many quarters seeing graphics revenue exceed CPU revenue and likely preventing the company from bankruptcy.

That lifeline is no longer necessary. AMD has become a highly competitive CPU company again, with quarterly sales approaching the $2 billion mark. For the second quarter of 2019, AMD's GPU shipments increased 9.8% versus Q1, while Nvidia's remained flat and Intel's decreased 1.4%, according to Jon Peddie Research. While both the CPU and GPU businesses continue to perform well, the central challenge for AMD is translating its gaming popularity into enterprise sales.

In high-performance computing and artificial intelligence, Nvidia maintains clear dominance, and AMD has no answer for Nvidia's RTX 270/280 or the Tesla T4. Nonetheless, AMD has secured notable wins: Oak Ridge National Lab plans to build the Frontier exascale supercomputer in 2021 using AMD Epyc processors and Radeon GPUs. AMD CEO Lisa Su described it as featuring "highly optimized CPU, highly optimized GPU, highly optimized coherent interconnect between CPU and GPU, [and] working together with Cray on the node to node latency characteristics really enables us to put together a leadership system." AMD has also secured deals with Google to power its cloud-based Stadia game console, delivering 10.7Tflops/sec—more than the Microsoft and Sony consoles combined—and a multiyear GPU computing contract with China's Baidu.

The fundamental challenge, according to Peddie, lies in software rather than hardware. Nvidia's CUDA language, originally developed by Stanford professor Ian Buck, who now heads Nvidia's AI efforts, enables developers to write GPU-optimized applications using familiar C++ syntax. Nvidia subsequently established CUDA curricula at hundreds of universities. "The net result is universities around the world are cranking out thousands of grads who know CUDA, and AMD has no equivalent," Peddie said. AMD supports the open-standard OpenCL library and the open-source HIP project, which converts CUDA to portable C++ code, but standards-based approaches face inherent limitations. OpenCL, developed by Apple and now maintained by the Khronos Group, exemplifies this problem—when Microsoft introduced DirectX, it obliterated OpenGL despite the latter's earlier dominance. Standards fare better when backed by a company with financial incentive.

For AMD to gain ground against Nvidia in the data center and HPC/AI markets, it requires a CUDA competitor. Until two years ago, this was impossible as AMD fought for survival. With competitive new silicon now in hand, the time is right for AMD to invest in software development and challenge Nvidia as effectively as it is challenging Intel.