What makes Intel’s new DG1 GPU unique and formidable | Warpcore

For years, Intel dominated CPUs and desktop chipsets, however, in the 2010s with the waning ascendency of Moore’s Law and increase in AI workloads, star of the Santa Clara based silicon valley icon has declined. It has been GPUs that have pushed the envelope in the past decade. It has been Intel’s weakness in the GPU space which has allowed traditional rivals like AMD and Nvidia to whittle away on Intel’s dominance in the traditional PC market. So to counter this dominance, Intel a couple of years ago hired Raja Koudri, the legendary GPU architect who was behind the revival of AMD’s Radeon GPUs. He also had a stint at Apple where he worked on the graphics stack of the iPhone which is known to be a core strength of the gadget. Since joining in 2017, Koduri, who is the senior VP and chief architect of Intel’s GPUs has been developing a new discrete GPU stack which finally showed up in the form of the DG1 at CES today. And Intel making a serious foray into the discrete GPU space alone is a huge deal, however, there are certain details that have come out which could make it very unique and formidable. Here’s why.

What is the pedigree of the Xe architecture

The first question for most people will be what’s this new Xe architecture that Intel has been talking up? Xe is Intel’s new graphics architecture which claims to give discrete graphics levels of performance. Intel claims the performance of this is 2x that of what its integrated graphics provided in 2019. This is also being provided as an integrated solution for the new Tiger Lake CPUs which are based on the new 10nm++ manufacturing node. Intel will also for the first time have a dedicated discrete GPU which will be based on this Xe architecture. So what are the key features of Xe architecture?

• Xe will have “One architecture” running across SKUs be it for a low power entry-level integrated solution like in the case of the Tiger Lake SoCs, a mid-range solution cutting across mid-range, enthusiast and datacenter workloads or exascale solutions which is the Xe HPC also called the Ponte Vecchio. In other words, Xe LP will scale from 5–50 watts which will for mainstream laptops, 75–250 watts will be Xe HP for desktop replacement laptops to workstations and beyond for Xe HPC.
• This architecture will be able to scale to 1000s of execution points with 40x better double-precision floating-point compute. There will also be a new scalable memory fabric called the XEMF along with a huge unified cache called the Rambo cache.
• The Xe HPC will be manufactured on Intel’s next-generation node — 7nm which it claims will provide 2x improvements on the 10nm node. This will be restricted to HPC while the other SKUs of Xe will be based on 10nm+. There will be other optimisations like the addition of Intel’s new Forveros technology for Rambo interconnect.
• Ponte Vecchio which is the first exascale GPU developed by Intel based on the Xe will be in the Aurora supercomputer. It will also have 16 compute chiplets. It is also said to be one of the largest GPU units of all time, in fact, it isn’t even a GPU in the traditional sense.

Where does that leave the DG1?

• Now the DG1 is based on this very pedigree. It can be assumed that what Intel showed off at CES 2020 was based on the Xe LP, or low power module which will power notebooks between 5–20 watt power envelop. Basically, most ultrabooks or something like the MacBook Air or the MacBook Pro with a discrete GPU.
• Intel showed a demo of Destiny 2 running at very playable frame rates with graphics being very impressive. They didn’t share any details for settings and configuration of the GPU.
• Since it is based on the Xe architecture many are assuming this is basically the same GPU as a Tiger Lake integrated module. The one architecture philosophy of this system should allow users to double up the integrated GPU with the discrete one for a massive boost in performance. This is a new way of thinking about GPUs as discrete GPUs from Nvidia and AMD can’t be paired with the Intel integrated graphics.

Why this time Intel could win? The team

• For the first time in perhaps a decade, Intel has an absolute work class team of leaders working on this GPU. They hired legendary GPU architect Raja Koduri from AMD in 2018. In 2019, Koduri brought in his former AMD colleague Jim Keller from Tesla. Some of the people who are working on Intel’s latest technology are the very best in the industry and have a legendary status. Both Kellar and Koduri were responsible for AMD’s revival and they also played a huge part in the silicon success that Apple is enjoying today.
• Intel’s inherent advantage of being the default CPU for workstations, notebooks and even data centres opens a window for an architecture like Xe which can be pair integrated solutions with discrete solutions.

At the end of the day, Intel has a very strong team working on this, with a fresh approach that plays to its strengths and well it leverages new technologies which could result in a disruptive GPU unit which could usurp the duopoly of Nvidia and AMD in the graphics space. Whatever be the case, the DG1 will be stepping stone towards this future and that’s why it is unique and important.

Words by Sahil Mohan Gupta

Originally published at https://warpcore.live on January 7, 2020.