Wow, nervana, and now movidius. Intel are eating a lot of non nvidia machine learning hardware. Theyre also really into opencl, half the talks at this years international workshop on opencl were from intel http://www.iwocl.org/attend/sessions/ Interesting times :)
That's definitely one way of saying Intel is all over the place lately because it has no idea what's "the future."
Many-core CPUs (Phi), FPGA's (Altera), ASICs (Nervana), and now the GPU-like Movidius for machine learning.
You could argue that "they're just experimenting" to see what will take off. But I don't know if paying $17 billion for something (Altera) still fits in the "experiment" category. Was acquiring McAfee antivirus for $8 billion an experiment, too? What if FPGAs aren't the future of machine learning, and they fade out? Does Intel write-off the $17 billion, like it did the $10+ billion it spent on the mobile market.
And it's now trying to make x86 chips, while at the same time making ARM chips better for the competition. I don't think Intel (its current CEO) has any clue about what the company should focus on. Sooner or later, its PC market profits will be squeezed/wasted on these acquisitions, and Intel will need to show something for it.
No, it would not need to write off the $17 billion, it would simply spin Altera back out again. It was a successful stand-alone company and FPGAs have plenty of applications that have nothing to do with machine learning to make it a viable segment of the market for a long long time to come.
McAfee is a different matter and the mobile market could have been won but not in the way that Intel hoped.
They could have won the mobile market, at the cost of severely hurting their desktop market. In a way, Intel was scared of their atom being "too good". They had to suck, because the major premium they earn is selling i3/i5/i7/xeon, a low end x86 chip that performs well would have earned them good parts of the mobile market at low margin, and lost sales in the desktop market at high margin.
Already the cheap end x86 notebooks were ditching i3 for atom, and being "good enough".
The instruction set is irrelevant. In modern CPU, decoding the X86_64 instruction set takes about 80K transistors. Compare this with 1,175 millions transistor packed into a Skylake - it does not even factor.
What is relevant is the ability to compute within a power envelope. Intel didn't try to scale down its new designs for mobile. It tried to scale up old cores with new technology - something that they got good at during the tick-tock development flow. And guess what, new core designs beat process improvement. Intel eventually reached same power envelopes as ARM with more computing power, and better overall performance due to better memory bus access. But it reaches them after 8-12 months after Qualcomm does, so it's a full generation behind in terms of mobile development.
Specific to this acquisition, Movidius packs 150 Gflops under 4W of power. Skylake packs 100Gflops under 120W of power. The difference here is more than 2 degrees of magnitude. Moving the designs to the new Intel processes should cover another significant improvement. So that's what's going on.
> Specific to this acquisition, Movidius packs 150 Gflops under 4W of power. Skylake packs 100Gflops under 120W of power. The difference here is more than 2 degrees of magnitude. Moving the designs to the new Intel processes should cover another significant improvement. So that's what's going on.
Where did you get the measurements? Link please.
M2 to Skylake is like apples to oranges. The RISC cores are based on Leon, and then you have 12 SHAVEs/VLIW cores, which are very good at number crunching and GPU/DSP-like tasks, but they are not really general purpose CPUs as such. I think you should exclude them from your comparison.
Anyway, good for Movidius, I wish 'em all the best.
Added: Intel created a huge moat and wall around their x86 architecture. It worked very well for them in keeping out the rest of the world. The problem is, they inadvertently bricked up their gate and burned their drawbridge and now they are imprisoned in their own castle.
Intel has the luxury to bet on multiple things. And I think thats a smart strategy becasue Intel is fighting Nvidia, IBM, Mellanox, ARM & many other competitors on multiple fronts. Betting on "the future" is not a single thing, its exploring multiple technologies that have the potential to become their own version of "the future".
> That's definitely one way of saying Intel is all over the place lately because it has no idea what's "the future."
As has no one else. Everyone who said that he knows the future has been proven wrong so far, at most people can hope to try to see a possible future, nothing more.
My first reaction was to ask: 'Why would not these technologies (acquired from Intel-Phi, Altera, Nervana, Movidius) not be successful in their own niches?'
Now I understand that it is easier to coalesce around a champion technology and too hard to navigate around optimising multiple methods. But I'd expect that the differences between them are so large that they would have their own niche implementations.
Trying to project the future is hard. That is why we have index funds. Invest in a bunch of different assets to reduce risk. The main issue is there is some correlation between the companies which screws up the math.
They know they won't beat nvidia on CUDA and playing on their turf won't work, so they make sure to win on opencl. They might not know which solution will win on that field but if they have all of them they're sure to hold a "winner".
