Friday, August 14th 2020
Xe HPG is Real, Intel's Gaming GPU Releases in 2021, without HBM
Intel on Thursday at its 2020 Architecture Day event announced a high performance gaming variant of its Xe graphics architecture, which it calls Xe HPG. The Xe architecture is designed to scale between tiny iGPUs and mobile discrete GPUs as Xe LP, up to scalar compute processors under Xe HP, and beyond to HPCs and supercomputers, under Xe HPC. The combination of the client graphics feature-set of Xe LP, with the scale of Xe HP, results in Xe HPG. Intel is designing Xe HPG for a third-party semiconductor foundry, and hopes to debut it in 2021.
In our older graphics detailing the Xe LP, we tried to explain just how easy it is for Intel scale up the iGPU to a discrete GPU SoC. This is done by simply dialing up the Xe slices, and dropping in dGPU ancillaries such as a PCI-Express host interface and memory controllers for the prevalent client-segment discrete graphics, namely GDDR6. There will be additional components, such as ray-tracing hardware. Intel is gunning for DirectX 12 Ultimate logo compliance, and ray-tracing forms a big part of that.Even before the Xe HPG, the first discrete GPU based on Xe is expected to be derivatives of Xe LP built on Intel's 10 nm SuperFin node, which are essentially the Xe LP iGPU of "Tiger Lake" spun out on dies of their own, with memory and PCIe interfaces. These should make up mobile discrete graphics cards that can augment performance of the Xe LP iGPU of notebooks with "Tiger Lake" processors.
In our older graphics detailing the Xe LP, we tried to explain just how easy it is for Intel scale up the iGPU to a discrete GPU SoC. This is done by simply dialing up the Xe slices, and dropping in dGPU ancillaries such as a PCI-Express host interface and memory controllers for the prevalent client-segment discrete graphics, namely GDDR6. There will be additional components, such as ray-tracing hardware. Intel is gunning for DirectX 12 Ultimate logo compliance, and ray-tracing forms a big part of that.Even before the Xe HPG, the first discrete GPU based on Xe is expected to be derivatives of Xe LP built on Intel's 10 nm SuperFin node, which are essentially the Xe LP iGPU of "Tiger Lake" spun out on dies of their own, with memory and PCIe interfaces. These should make up mobile discrete graphics cards that can augment performance of the Xe LP iGPU of notebooks with "Tiger Lake" processors.
10 Comments on Xe HPG is Real, Intel's Gaming GPU Releases in 2021, without HBM
A bit like Intel trying to be a dGPU company but only presenting slides. Its amazing how they've already managed full confusion on naming even before anything is out :D
"10nm Superfin"
:roll::roll::roll:
While I dont care for the premium range at all (spending twiceas much for + 30% is just stupid) Having three players in the 3060/3070 midrange would really put some pressure on prices in a segment that really matters to the majority of gamers...
But now 2021 sounds more like Q3/4 2021 so nothing substancial and they miss out on all that hype that comes with nextgen consoles this year...
The Xe architecture is designed to scale between the low end, both integrated and discrete; this is known as Xe LP. It is also able to scale up to scalar computing (definition of scalar) processors, this is known as Xe HP, and even further to high performance computing and supercomputers under the name Xe HPC.
If you take the client graphics-oriented features of LP, with the wide range of features under HP, you get HPG, which is their gaming-oriented lineup of Xe GPUs.
Or to simplify:
Xe HPC: High performance computing, supercomputers
Xe HP: Scalar computing
Xe HPG: Gaming-oriented GPUs, mixing featuresets from HP and LP
Xe LP: Low powered client GPUs, integrated and low-end discrete
Each year i use to wait that when Intel will bring hardware AI instructions to get in the game with Nvidia (Tesla Motors and Volvo AI contract is with Nvidia)
Intel wants to play it realtime in graphics world.
Which reminds me watching a video in which linus was explaining that Intel Discrete and igp had an internal presentation where Igp was proposed to move forward while discrete was left behind because of higher tdp etc etc and AI was noware a threat for Intel.