If you were a big enough customer you could get a SKU for you, too. E.g. hyperscalers have Xeons which are not available for any other customers for any price.
They turned down Acorn about the 286, which led to Acorn creating the Arm, they have turned down various console makers, they turned down Apple on the iPhone, and so on. In all cases they thought the opportunities were beneath them.
Intel has always been too much about what they want to sell you, not what you need. That worked for them when the two aligned over backwards compat.
Clearly the threat of an Arm or RISC-V finding itself fused to a GPU running AI inference workloads has woken someone up, at last.
The problem is, console manufacturers know precisely how much of their product they anticipate to sell, and it's usually a lot. The PlayStation 5 is 80 million units so far.
And at that scale, the console manufacturers want to squeeze every vendor as hard as they can... and Intel didn't see the need to engage in a bidding war with AMD that would have given them a sizable revenue but very little profit margin compared to selling Xeon CPUs to hyperscalers where Intel has much more leverage to command higher prices and thus higher margins.
> they turned down Apple on the iPhone
Intel just was (and frankly, still is) unable to compete on the power envelope with ARM, that's why you never saw x86 take off on Android as well despite quite a few attempts at it.
Apple only chose to go for Intel with its MacBook line as PowerPC was practically dead and offered no way to extract more performance, and they dropped Intel as soon as their own CPUs were competitive. To get Intel CPUs to the same level of power efficiency that M-series CPUs have would require a full rework of the entire CPU infrastructure and external stack, that would require money that even Intel at its best frankly did not have. And getting x86 to be power effective enough for a phone? Just forget it.
> Clearly the threat of an Arm or RISC-V finding itself fused to a GPU running AI inference workloads has woken someone up, at last.
Actually, that is surprising for me as well. NVIDIA's Tegra should easily be powerful enough to run the OS for training or inference workload. If I were to guess, NVIDIA wants to avoid getting caught too hard on the "selling AI shovels" train.
They did push hard on their UMPC x86 SoCs (Paulsbo and derivatives) to Sony, Nokia, etc. These were never competitive on heat or battery life.
You probably meant Poulsbo (US15W) chipset
80 million in 5 years is a nothing burger as far as volume.
NVDA sold 153 million Tegra units to Nintendo in 8 years, so 1.5M units a month. That's just as comparable.
[1] https://www.servethehome.com/on-ice-lake-intel-xeon-volumes-...
And so that gave AMD an opening, and with that opening they got to experiment with designs, tailor a product, get experience and industrial marketshare, and they were able to continue to offer more and better products. Intel didn't just miss a mediocre business opportunity, they missed out on becoming a trusted partner for multiple generations, and they handed market to AMD that AMD used to be a better market competitor.
AMD isn't precisely a market competitor. The server and business compute market is still firmly Intel and there isn't much evidence of that changing unless Apple drops M series SoCs to the wide open market which Apple won't do. Intel could probably release a raging dumpster fire and still go strong, oh wait, that's what they've been doing the last few years.
AMD is only a competitor in the lower end of the market, a market Intel has zero issue handing to AMD outright - partially because a viable AMD keeps the antitrust enforcers from breathing down their neck, but more because it drags down per-unit profit margins to engage in consoles and the lower rungs and niches.
This is not true anymore, as it IS changing, and very rapidly. AMD has shot up to 27.3% of the server market share, which they haven't had since the Opteron days 20 years ago. Five years ago their server market share was very small single digits. They're half of desktops, too. https://www.pcguide.com/news/no-amd-and-intel-arent-50-50-in...
Intel, at one of its lowest low, still come up with lunar lake, which is not as efficiency as Apple M, but still, quite impressive.
I bet if they were focus on mobile when they are at their peak, they could come up with something similar to Apple M
It leads to mistakes like you mention, where a new market segment or new entrant is not a sure thing. And then it leads to mistakes like Larrabee and Optane where they talk themselves into overconfidence (“obviously this is a great product, we wouldn’t be doing it if it wasn’t guaranteed to make $1B in the first year”).
It is very hard to grow a business with zero risk appetite. You can’t take risky high return bets, and you can’t acknowledge the real risk in “safe” bets.
2010-2011 was also the time that AMD were starting to moan a bit about DX11 and the higher level APIs not being sufficient to get the most out of GPUs, which led to Mantle/Vulkan/DX12 a few years down the road. Intel did a bit regarding massively parallel software rendering, with the flexibility to run on anything x86 and implement features as you liked, or AMD's efforts for 'fusion' (APU+GPU, after recently acquiring ATi) or HSA which I seem to recall was about dispatching different types of computing to the best suited processor(s) in the system for it. However I got the impression a lot of development effort is more interested in progressing on what they already have instead of starting in a new direction, and game studios want to ship finished and stable/predictable product, which is where support from intel would have helped.
But certainly Intel wasn’t willing to wait for the market. Didn’t make $1 billion instantly; killed.
It's typical corporate venturing and reporting to a CFO. Google is not much better with them cutting their small(er) projects.
Sounds like they will someday soon.
There will always be giant, faraway GPU supercomputer clusters to train models. But the future of inference (where the model fits) is local to the CPU.
This relates to the Intel problem because they see the world the way you just described, and completely failed to grasp the importance of SoC development where you are suddenly free to consider the world without the preexisting buses and peripherals of the PC universe and to imagine something better. CPU cores are a means to an end, and represent an ever shrinking part of modern systems.