Intel this week announced that it had taped out and powered on its first graphics processor based on the Xe-HPG architecture. The company also reaffirmed that it is working on a stack of discrete Xe-HPG GPUs that will be used for mid-range and enthusiast-class gaming PCs sometimes next year. In the meantime, Intel started to ship its DG1 discrete GPU based on the Xe-LP architecture for entry-level gaming PCs. (via SeekingAlpha / Intel Earnings Presentation).
“We powered on our next-generation GPU for client DG2,” said Bob Swan, CEO of Intel, during the company’s earnings call with analysts and investors. “Based on our Xe high-performance gaming architecture, this product will take our discrete graphics capability up the stack into the enthusiast segment.”
Intel’s first discrete GPU in two decades — the DG1 — relies on the same Xe-LP architecture that is used for the company’s latest built-in GPUs found in codenamed Tiger Lake processors. Intel is currently shipping its DG1 GPUs for revenue and expects the first PCs with its discrete graphics inside to hit the market later this quarter.
“Our first discrete GPU DG1 is shipping now and will be in systems from multiple OEMs later in Q4,” said Swan.
Intel’s Xe-LP GPUs have a significantly higher performance than Intel’s previous-generation graphics solutions, but since Xe-LP GPUs have to be integrated into CPUs, they are tailored primarily for low power consumption and efficiency in terms of transistor count. This is not the case for other Intel Xe architectures, namely Xe-HP for datacenters, Xe-HPC for supercomputers, and Xe-HPG for gaming PCs. In fact, Xe-HPG combines numerous peculiarities of other three architectures.
“We have been working since 2018 on another optimization of Xe-HP targeted gaming,” said Raja Koduri, chief architect of Intel, at the company’s Architecture Day in August. “That microarchitecture variant is called Xe-HPG. […] We had to leverage the best aspects of the three designs we’ve had in progress to build a gaming optimized GPU. We had a good perf-per-watt building block to start with, Xe-LP. We leveraged the scale from Xe-HP to get a much bigger config and we leveraged the compute frequency optimizations from Xe-HPC.”
Intel’s Xe-HPG GPUs will support hardware-accelerated raytracing along with other features, which will have an impact on the architecture of its execution units and/or sub-slices. Furthermore, since Xe-HPG GPUs will be made at an external foundry (i.e., at TSMC), Intel used a lot of third-party IP (e.g., memory controller and interface, display interfaces, etc.) to optimize design costs.
Intel got its first Xe-HPG silicon from its foundry partner back in mid-August and has tested it internally since then. So far, the company has only confirmed that the GPU could be powered on, but this is a good sign in general.
Intel’s family of Xe-HPG graphics processors will consist of multiple GPUs targeting market segments that will span from mid-range all the way to enthusiast level. So far, Intel has not disclosed how many discrete Xe-HPG graphics chips it plans to launch next year. Intel also did not reveal at this time whether it had taped out and powered on its big flagship Xe-HPG GPU, or something smaller and cheaper. Typically, GPU companies like AMD and Nvidia tend to introduce their big GPUs first and then follow up with smaller processors. However, every rule has an exception. For example, Nvidia started to roll-out its highly-successful Maxwell architecture from mainstream offerings.
In any case, right now Intel is bringing up its Xe-HP graphics processors for datacenters that are made using its 10nm Enhanced SuperFin process technology and demonstrate performance of around 40 FP32 TFLOPS as well as an unknown Xe-HPG graphics processor.