Slashdot Mirror
Intel Discloses Its Forthcoming Discrete GPU Strategy and Design Efforts (hothardware.com) (hothardware.com)
MojoKid writes: Intel has been uncharacteristically vocal about its most recent plans to enter the discrete GPU market. Over the last year or so, the company has disclosed a few morsels of information and made some high-profile hires, in its bid to build-up and flesh-out its latest discrete GPU plans. This week, Intel decided to have a sit down with HotHardware, offering the opportunity to chat with Ari Rauch, Vice President of the Core And Visual Computing Group at Intel, to discuss what makes this most recent endeavor different from the company's previous and now discontinued attempts in the discrete GPU space. As a follow up, HotHardware also enlisted readership questions to engage with Intel about its upcoming GPU plans, compiling responses in a Q&A format.
In short, this isn't Larabee 2.0, not by a long shot. Intel is gearing up for a traditional GPU architecture design, coupled with some of the company's own strategic IP that it can bring to the table, to help differentiate its products. Further, Rauch noted Intel "will bring discrete GPUs to both client and data center segments aiming at delivering the best quality and experiences across the board including gaming, content creation, and enterprise. These products will see first availability over a period of time, beginning in 2020."
When questioned on their current silicon fabrication hiccups and delays and how it might affect Intel's ability to execute in this highly competitive space, Rauch noted, "we feel very confident about our product roadmap across software, architecture, and manufacturing." Based on some of the responses to product positioning questions, it also appears Intel is gearing up to address all performance envelopes as well, from entry-level to midrange and high-end graphics cards.
In short, this isn't Larabee 2.0, not by a long shot. Intel is gearing up for a traditional GPU architecture design, coupled with some of the company's own strategic IP that it can bring to the table, to help differentiate its products. Further, Rauch noted Intel "will bring discrete GPUs to both client and data center segments aiming at delivering the best quality and experiences across the board including gaming, content creation, and enterprise. These products will see first availability over a period of time, beginning in 2020."
When questioned on their current silicon fabrication hiccups and delays and how it might affect Intel's ability to execute in this highly competitive space, Rauch noted, "we feel very confident about our product roadmap across software, architecture, and manufacturing." Based on some of the responses to product positioning questions, it also appears Intel is gearing up to address all performance envelopes as well, from entry-level to midrange and high-end graphics cards.
From the interview:
Q: Will Intel’s new GPU architecture eventually migrate down onto the CPU or will the discrete and integrated solutions remain separate architectures?
A: Leveraging Intel’s broad portfolio of products is critical to building winning platforms: lots of performance, in compelling form factors, in compelling power envelopes. We’re excited by the opportunity to build technologies that will allow us to take experiences, features, and innovation to new and unique form factors, and to an install base of a billion screens around the world.
Are they going to improve the integrated graphics in their CPUs? (which currently is the weakest link in their offering, AMD Ryzen APUs have Vega GPU cores). According to the interview....... I don't know!
I think there is WAY more progress in the AMD and ARM front
They have developed no IP to do the things needed to compete with Nvidia or AMD in the next 2-3 years.
IP doesnt always have to be developed by a particular company. See Intels latest deal with AMD for integrated graphics.
I have been saying for a couple years now here that Intel is in very serious trouble. Especially after those layoffs and the PR announcement for a "cloud strategy." The first key point here is that its taken several years before it became obvious to most (even here) that Intel is in any trouble at all. The second key point is that Intel did know it years ago that they were in big trouble.
Intels biggest problem is that their vertical integration has really constrained them. Silicon (not just CPU's) doesnt leave an Intel plant without being branded Intel. They have older fabs that are idle because they wont sell time on them, and newer fabs that even at 100% capacity cant satisfy demand. The later wouldnt be a problem if Intel were the only source for a particular component by raising prices to decrease demand, but the reality is their competitors in total have far more capacity than they do.
It is because of all this that a company like AMD would trade off some IP to Intel. It doesnt fix Intels fundamental and now unfixable problem, which is that they will never be the market leader again, never steer the markets that they partake in. From here on out they can only react to what other market players are doing.
On the desktop process side, Intel was blindsided by the economy of AMDs chiplets, and they are still at least several years from an effective design. It isnt just about small dies on a single processor board, its about being wholly modular. The same chiplets that Threadripper uses are also used by AMD's low end Ryzen APUs.
Intel does have some "chiplet" experience but then too it was as a reaction to a blind-sided moment when their main competitor introduced multi-core to the consumer. It was a hack that they didnt explore but should have.
On the fabrication side, Intel is now dwarfed by the rent-a-fab market capacity on its entirety, and even individual rent-a-fabs are now overtaking them in capacity.
I've said it before and I'll say it again. Sell your Intel stock. Even if you arent manually in the stock market, check your 401Ks and Roth IRAs. They might be able to prevent becoming a Motorola, but even if they do its still bad. Very bad. Intel is fucked.
"His name was James Damore."
Its all fun and games when the scene data is static.
What has prevented RTRT for all this time has been dynamic scene data - things moving around - the acceleration structures like space partitioning trees are either too expensive to generate in realtime or sacrifice too much trying to deal with it.
Remember that GPUs have high memory bandwidth but absolutely terrible memory latency, not like CPU's where the opposite is true. Intel learned that the issue continued to remain insurmountable with their Larrabee failure. They could build those acceleration structures quickly on Larrabee, but then when it came time to render, the lack of memory bandwidth became the killer.
As far as I know there is still no acceptable solution. nVidia is claiming they have it, but in practice they are still mainly rasterizing.
"His name was James Damore."