Analyzing Intel's Discrete Xe-HPC Graphics Disclosure: Ponte Vecchio, Rambo Cache, and Gelato

[PontiacGTX]

Exascale for Everyone

 

 

Intel says that it is hard not to notice the ‘insatiable’ demand for faster, more power efficient compute. Not only that, but certain people want that compute at scale, specifically at ‘exascale’. (It was disclosed at a high-performance supercomputing event, after all). For 2020 and beyond, Intel has designated this the ‘Exascale’ era in computing, where no amount of compute is good enough for leading edge research.

 

 

On top of this, Intel points to the number of connected devices in the market. A few years ago analysts were predicting 50 B IoT devices by 2020-2023, and in this presentation Intel is saying that by mid-2020 and beyond, there will be 100 billion devices that require some form of intelligent compute. The move to implementing AI, both in terms of training and inference, means that performance and computational ability have to be ubiquitous: beyond the network, beyond the mobile device, beyond the cloud. This is Intel’s vision of where the market is going to go.

 

 

Intel splits this up into four specific categories of compute: Scalar, Vector, Matrix, and Spatial. This is certainly one blub part of the presentation I can say I agree with, having done high-performance programming in a previous career. Scalar compute, is the standard day-to-day compute that most systems run on. Vector compute is moving to parallel instructions, while Matrix compute is the talking point of the moment, with things like tensor cores and AI chips all working to optimize matrix throughput. The other part of the equation is spatial compute, which is derived from the FPGA market: for sparse compute that is complex and can be optimized with its own non-standard compute engine, then an FPGA solves it. Obviously Intel’s goal here is to cover each of these four corners with dedicated hardware: CPU for Scalar, GPU for Vector, AI for Matrix, and FPGA for Spatial.

 

 

One of the issues with hardware, as you move from CPU to FPGA, is that it becomes more and more specialized. A CPU for example can do Scalar, Vector, Matrix, and Spatial, in a pinch. It’s not going to be much good at some of those, and the power efficiency might be poor, but it can at least do them, as a launching point onto other things. With GPU, AI, and FPGA, these hardware specializations come with different amounts of complexity and a higher barrier to entry, but for those that can harness the hardware, large speed-ups are possible. In an effort to make compute more ubiquitous, Intel is pushing its oneAPI plan with a singular focal resource for all four types of hardware. More on this later.

 

 

Intel’s X^e architecture will be the underpinning for all of its GPU hardware. It represents a new fundamental redesign from its current graphics architecture, called ‘Gen’, and pulls in what the company has learned from products such as Larrabee/Xeon Phi, Atom, Core, Gen, and even Itanium (!). Intel officially disclosed that it has its first X^e silicon back from the fabs, and has performed power cycling and basic functionality testing with it, keen to promote that it is an actual thing.

 

 

So far the latest ‘Gen’ graphics we have seen is the Gen11 graphics solution, which is on the newest Ice Lake consumer notebook processors. These are out in the market, ready to buy today, and feature performance 2x over the previous Gen9/Gen9.5 designs. (I should point out that Gen10 shipped in Cannon Lake but was disabled: this is the only graph ever where I’ve seen Intel officially acknowledge the existence of Gen10 graphics.) We have seen diagrams, either potentially from Intel or elsewhere, showing ‘Gen12’. It would appear that ‘Gen12’ was just a holding name for X^e, and doesn’t actually exist as an iteration of Gen. When we asked Raja Koduri about the future of Gen, he said that all the Gen developers are now working on X^e. There are still graphics updates to Gen, but the software developers that can be transferred to X^e have been already.

 

Since I didnt find the thread related to this here is the link of the analysis

Edited January 14, 2020 by Developer

[ENTERPRISE]

Interesting read and Intel is not wrong. Compute tasks are growing exponentially in all facets of computing. That being said it would seem Intel with their new GPU venture is focused on the compute model and not so much your typical GPU for 3D/Gaming applications. That is at least what I am getting from the article.

 

I would like to see Intel also focus on the standard gpu consumer needs as well in order to create more competition in this area. We need someone to keep AMD and Nvidia in touch. Granted Intel have some learning to do in this area as Nvidia and AMD have been in the GPU business for a while now.