Industry sources say discrete Intel Xe cards will arrive mid-2020

[DanceswithUnix]

Intel definitely wants a slice of the GPU pie so i expect for the first couple of generations they'll sell at a loss to get foothold into the market.

Probably. I gather they gave away racks of the compute boards to try and seed that market.

[excalibur1814]

If he have not, then Intel have definitely hired the wrong guy.
But at this level ( minus when politics are involved ) you are only allowed to make the same mistake once.

Thats not true ... look at the windows 10 team !

Or Apple. Or Google. Or any big company.

[Xlucine]

Yes they have. They bought a graphics company, marketed the technology as Intel's, created lackluster discrete graphics cards which didn't sell all that well until they killed it off. Google "i740".

Then with Larrabee they decided they could do graphics with really wide SSE instructions on lots of Atom class cores because someone high up said that everything should run x86 instructions from embedded to graphics cards. Intel have always had a problem with drivers, so making a software centric GPU seemed a dumb move, they eventually gave up and released the product as a compute only card with no graphics output.

Intel have no problem making boards and shipping products, but the one thing Koduri could show them is how to make graphics compute scale beyond small integrated units as even their big integrated graphics hit a wall before they got to being big enough to need a freestanding card.

I'm expecting this to be expensive, hot, slow with buggy drivers and reliable hardware. If they are prepared to lose money on it it might not be expensive.

Given recent price hikes, I suspect modern GPUs are at much higher profit margins than historically (at least, non-ray tracing ones with sensible die sizes). With that kind of competition there's a lot of room for cheap cards to still turn a profit (followed by AMD cutting 5700 and 5500 prices to match, and nvidia sweating nervously). Selling at a loss will probably make current 570 prices look steep

[stoatwblr]

Intel have no problem making boards and shipping products

Let's be really clear. Intel is a midddling memory foundry with a CPU which got lucky thanks to IBM, Microsoft AND the existence of AMD (IBM required there be a second source cpu supplier)

There were better CPUs, there were better architectures and there were definitely more power efficient systems available - the x86 was one of the slowest (per clock) and least power efficient out there so it should be no surprise that Intel was able to make them faster and sippier over the years - The business competition got swept away by the power of the the 900 pound IBM marketing gorilla and then the rest of the world got wiped out by the avalanche of the Microsoft yeti.

Intel hasn't had to innovate much over the years, except to keep AMD from nipping at its heels - and the amount of unlawful activity it's shown to be involved in to keep AMD out of retail space shows how scared it's _really_ been of the competition.

If Intel is seen to be leveraging its near-monopoly in the CPU space to enter and become a major player in Graphics space, the competition regulators in a bunch of jurisdictions are going to start looking _very_ closely at their activities again. As such, you're unlikely to see "cheap Intel cards" for the simple reason that the moment sales of them start impinging significantly on sales of AMD/Nividia discrete parts those competitors will be down at the regulator offices thumping on the counters.

[stoatwblr]

Adding to my last comment:

One of the things that _really_ irritates the guys I work with (large planetary imaging datasets) is how Nvidia keeps dumping driver support for older devices.

Cuda might be "cool" and the "only way" for now, but everyone's frantically learning to use OpenCL for portability and trying to port as much as they can to not be Cuda-dependant.

They've been looking longingly at Xeon Phi units for a while. We even have a couple of eval cards kicking around - which although long in the tooth now, are better supported than Nvidia cards of the same age.

In the end the things that power this kind of thing may not even _be_ GPGPUs, but being able to treat your "gpu" as a cluster of x86s has its own programming simplicity for number crunching work when you're not processing vectors and triangles.