Slashdot Mirror
Intel Is in an Increasingly Bad Position in Part Because It Has Been Captive To Its Integrated Model (stratechery.com)
Once one of the Valley's most important companies, Intel is increasingly finding itself in a bad position, in part because of its major bet on integration model. Ben Thompson, writing for Stratechery: When Krzanich was appointed CEO in 2013 it was already clear that arguably the most important company in Silicon Valley's history was in trouble: PCs, long Intel's chief money-maker, were in decline, leaving the company ever more reliant on the sale of high-end chips to data centers; Intel had effectively zero presence in mobile, the industry's other major growth area. [...] [Analyst] Ben Bajarin wrote last week in Intel's Moment of Truth. As Bajarin notes, 7nm for TSMC (or Samsung or Global Foundries) isn't necessarily better than Intel's 10nm; chip-labeling isn't what it used to be. The problem is that Intel's 10nm process isn't close to shipping at volume, and the competition's 7nm processes are. Intel is behind, and its insistence on integration bears a large part of the blame.
The first major miss [for Intel] was mobile: instead of simply manufacturing ARM chips for the iPhone the company presumed it could win by leveraging its manufacturing to create a more-efficient x86 chip; it was a decision that evinced too much knowledge of Intel's margins and not nearly enough reflection on the importance of the integration between DOS/Windows and x86. Intel took the same mistaken approach to non general-purpose processors, particularly graphics: the company's Larrabee architecture was a graphics chip based on -- you guessed it -- x86; it was predicated on leveraging Intel's integration, instead of actually meeting a market need. Once the project predictably failed Intel limped along with graphics that were barely passable for general purpose displays, and worthless for all of the new use cases that were emerging. The latest crisis, though, is in design: AMD is genuinely innovating with its Ryzen processors (manufactured by both GlobalFoundries and TSMC), while Intel is still selling varations on Skylake, a three year-old design.
The first major miss [for Intel] was mobile: instead of simply manufacturing ARM chips for the iPhone the company presumed it could win by leveraging its manufacturing to create a more-efficient x86 chip; it was a decision that evinced too much knowledge of Intel's margins and not nearly enough reflection on the importance of the integration between DOS/Windows and x86. Intel took the same mistaken approach to non general-purpose processors, particularly graphics: the company's Larrabee architecture was a graphics chip based on -- you guessed it -- x86; it was predicated on leveraging Intel's integration, instead of actually meeting a market need. Once the project predictably failed Intel limped along with graphics that were barely passable for general purpose displays, and worthless for all of the new use cases that were emerging. The latest crisis, though, is in design: AMD is genuinely innovating with its Ryzen processors (manufactured by both GlobalFoundries and TSMC), while Intel is still selling varations on Skylake, a three year-old design.
"instead of simply manufacturing ARM chips for the iPhone"
What's simple about it? Intel's ARM was Xscale, which was based directly on DEC's StrongARM (which they purchased.) It was the fastest ARM core at the time, but while it [x]scaled up, it didn't [x]scale down. It had the highest power consumption at low clock rates of all the ARM cores.
Intel did not have an ARM-based product which would have been a viable core for the iPhone.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
I think this kind of analysis is quite premature. Presently, there is no mobile-worthy x86 option -- for lots of reasons. Until there is, I don't think you can judge Intel for their direction.
Presume, for a moment, that in a few years, Intel successfully produces an x86 proc for mobile specifications. It's distinctly possible, indeed even probable, that ARM becomes useless, and the entire mobile market moves to x86. What a boon for Intel to have not wasted time and effort during these middle-ground years.
We've lived through this before. I refer you to WAP. How many web developers spent how many hours fumbling through WAP-limited options, before the entire mobile market moved to full web technologies? What a wasted investment for any small company. And what a horrible experience in was for consumers.
We'll wait and see.
And that power-per-watt disadvantage vs ARM predates Intel's integration strategy and also their current process-size disadvantage. I don't see any evidence to the contrary in the linked story.
This just marks the end of an era. Moores Law is dead (and has been dead for quite some time). Intel will need some other way to innovate. All they have been doing is adding cores and trying to push up clock speeds.
Moore's law is about transistor counts. Adding cores adds transistors.
Even this is running into a dead end: because of physics.
They can still add cores for some time, if they can improve yields. First there is a process shrink and cores shrink, then the process is improved and cores grow again. Then we get a new process...
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
I still say it's Compaq, not MS, that deserves the credit here. Compaq created the PC compatible. Microsoft did have the foresight to make sure its software worked too, but at the time the money was in hardware and you more or less expected incompatibilities between generations of the machine. The idea of cross-vendor compatibility in the micro market was Compaq's, for sure.
This is why Microsoft is doing it. The realized they are not beholden to Intel. They made Windows RT (port of Win32 to ARM) so if the Intel x86-64 ship ever sank, it wouldn't take Windows down with it. They don't need it to sell like hotcakes; heck they don't need it to sell at all. They just need to to be there and ready if ARM overtakes Intel. It's insurance - a hedge against Intel imploding. If that should happen, they'll just transition to Windows for ARM, and all the software companies making Windows apps will (more or less) simply recompile their programs for ARM64, and Windows will carry on as if Intel never existed.
Moore's law is about transistors per unit area. Adding cores increases both. Only new manufacturing techniques to cram in more transistors will let the trend continue, and they are indeed pushing the limits of what's physically possible.
At 7nm we're talking features that are only about three dozen atoms wide. The current roadmap has 5nm production in a few years. This kind of thing is well outside my knowledge but I'm pretty confident you can't make devices smaller than a single atom, so they are rapidly approaching a wall one way or another!
=Smidge=
There was a really big turning point with MCA vs PCI as well. That's when it became plain that it was no longer "the IBM compatible".
When IBM couldn't force MCA as the standard it became plain that it had lost control over the direction of the design, and that we were now in a commodity hardware multivendor world.
Market Cap : Net Income
INTC: 235.5B : 4,450M
AMD: 14B : 81M
Yeah, those of us who climbed up from the bottom really do like meritocracy, because we know that people got there on skill, ability, and competence.
I am also one of those people. And I know why so many of us are so bitter about seeing handouts. We struggled, and we want to see other people struggle too. It's some kind of messed up desire for fairness, when really, we should be trying to make sure nobody else has to go through the bullshit we did to succeed. The part you missed in the above formula is actually the largest factor- luck. Get over yourself, asshole.
This is nonsense.
Any time someone throws out the word RISC in the context of modern superscalar processors, they invariably have no fucking idea what they're talking about.
The denotation between RISC and CISC existed because once upon a time, CISC processors had richer instruction sets at the cost of more cycles per instruction.
These days, all processors (relevant to this discussion) are essentially CISC, and run at more than 1 instruction per cycle. The terms RISC and CISC are dead terms.
All superscalar ARMs have instruction decoders that break them into smaller micro-operations, a la microcode.