In an unprovoked tweet, in the style of a certain golf playing president who now likes kimchi, senior vice president of the Core and Visual Computing Group, general manager of edge computing solutions and chief architect at Intel Raja M. Koduri set off a meme by tweeting from Intel News, “Intel's first discrete GPU coming in 2020:”
Every pundit, blogger, and web poster mildly associated with GPUs immediately posted a story saying, “I told you so.” At last count there were over 270 re-tweets.
It’s hard to get lathered up about these almost off hand comments, what they lack in specifics would fill volumes. And, for the record, Intel has been designing and building GPUs since the Haswell came out in 2012. Intel had a team in Folsom CA, and another in Hillsboro OR for almost two decades designing graphics engines — anyone rememerb the i740, or a Larabee?Actually, it was at a recent analyst meeting where Intel CEO Brian Krzanich informed attendees that Intel will release its first discrete graphics chips in 2020, followed by an entire series of GPUs targeted at gaming, Artificial Intelligence and data center industries. After Krzanich’s comment, the company didn’t provide any details or confirmation. Instead, it linked back to an earlier announcement about the hiring of Koduri.
Give me back my silicon, or at least pay for it
Intel’s current integrated GPUs occupy 60% of the die of processors like Kaby Lake — 60%. And you know how much Intel gets paid for that expensive silicon? Zero, nadda, no dollars $0.00.
And although Intel has four or five AI engine solutions, depending upon how you count, a k-core GPU isn’t among them. Meanwhile Nvidia and AMD are getting all the AI/ML training jobs and pushing Xeons into the background in the data center, the one growth spot in all of PC land. Obviously, that won’t do. So, Kranich, like Pope Francis has thrown dogma to the winds and not only thought the unthinkable, but actually done it — acknowledging that GPUs exist, and that they are important. The first sign of the change in religious belief came with the revolutionary Kaby Lake-G.
Clearly, Intel isn’t going to go on buying chips from AMD for very long, although AMD can probably sell them to Intel cheaper than Intel can build one. Nonetheless, Intel is a vertical manufacturer with massive silicon factories that costs tens of billions of dollars, and even if they can’t get the bargain basement price for a part, they need to keep those ovens warm and running.
2020 is just 18 months away. Koduri has been at Intel for 6 months. Do you really think Intel’s discrete GPU project started after he arrived? Some of the 270 re-tweeters, pundits, and bloggers must know it takes four to five years to architect, design, and manufacture a GPU. And to that point, the famous and revered Jim Keller isn’t so much a great architect as he is a really great project manager. Because he knows the intimate details of a processor design, he also knows how to get a team to focus and stay on track — that’s the real secret sauce.
What are the odds?
Those of you who like betting, should start setting odds for Intel being able to really show a dGPU at CES 2020. You can also set odds for when they will actually ship it. Remember, the two most experienced, most powerful and most invested GPUs companies in the world aren’t taking a nap or on vacation. As successful as Nvidia is, Jensen Huang runs that company with the credo that you make one mistake, just one, and bam, you’re out of business, AMD with its limited resources spread over two processor families makes every dime and every second count. AMD is probably one of the most efficient companies in the industry right now. Intel is large, bureaucratic, and top-down driven. Intel can never match the efficiency or drive of AMD and Nvidia. That’s not a slam against Intel, you just can’t get that big, and cover all the corners Intel does and not have semi-rigid structure.
So, my crystal ball, dusty, scratched, and cracked as it is, says we won’t be able to buy a mass-produced dGPU from intel until 2021 or later.