Intel announces GPU products and cross-processor APIs designed to compete directly with NVIDIA's data center GPU dominance.

Intel announced its Xe Graphics Architecture at the Intel HPC Developer Conference in Denver, Colorado, ahead of the Supercomputing '19 tradeshow. The company bills the Xe Graphics Architecture card as its first "exascale graphics card," based on a new 7nm architecture called "Ponte Vecchio." Intel is splitting the Xe Architecture into three designs for different segments: data center, consumer graphics cards, and AI use-cases; integrated graphics for processors; and the high-tier Xe HPC for high performance computing. This effort represents Intel's third major attempt at GPU prominence—it hopes to succeed where its Larrabee GPU and Xeon Phi manycore processors failed.

Rather than use the large die for graphics chips the way Nvidia and AMD do, Intel is adopting the Multi-Chip Module (MCM) design that breaks up one big chip into smaller "chiplets" connected via high-speed fabric. The modules employ other packaging technology advancements including Foveros 3D chip-packaging technology, which allows 3D stacking of dies and mixing of CPU, AI, and GPU processors; High Bandwidth Memory (HBM); and Embedded Multi-Die Bridge (EMIB) technology to tie HBM packages to the compute die. Xe Architecture cards will also feature a new scalable fabric called XE Memory Fabric (XEMF), which ties compute and memory together with a coherent memory interface that Intel claims will allow Xe to scale to thousands of nodes.

Kevin Krewell, principal analyst with Tirias Research, noted that Xe is not a brand-new graphics architecture but an evolution from Intel's integrated GPU technology in Core CPUs. "This is a design that is more like traditional graphics. [Intel is starting] with their traditional integrated graphics cores and building on top of that. Larrabee tried a GPU built on a CPU. In this case they are building a ground up GPU with GPU-like features and not trying to do anything too weird. And now they've got a real GPU guy running the group," he said. That "real GPU guy" is Raja Koduri, senior vice president of the company's Core and Visual Computing Group, who previously led AMD's integration of CPU and GPU cores on one die and later pioneered graphics at Apple.

Intel is targeting Nvidia's dominant CUDA programming language with its new OneAPI programming model, designed to simplify programming across GPU, CPU, FPGA, and AI accelerators. Applications can move seamlessly between Intel's different types of compute architectures depending on where they will be most efficiently processed. Intel has also developed the Data Parallel C++ Conversion Tool to convert CUDA code to OneAPI. The company stated that OneAPI will be open-source and will work with other vendors' hardware, though it did not specify which vendors.

Despite its ambitions, OneAPI faces a significant maturity gap. As Krewell observed, "It's more than a shot at CUDA because they want to replace CUDA. OneAPI is a very ambitious program, trying to combine all of the different processor elements under one umbrella API. So it's a very aggressive program and they are building it out piece by piece. But right now it's at version 0.5. CUDA is at version ten. So they've got a ways to catch up."