Yes, but Intel problems are not the same of the other foundries.
14nm was supposed to enable two things, first a cost advantage enough to tilt the balance on mobile market towards Intel product, and a cost advantage enough to give Intel foundry business a good value proposition so it could begin to be noticed on the market. It failed miserably on these two counts, and it seems that it is also being detrimental to Intel other businesses as well.
And with 10nm being a straight shrink of this mess, that basically kills whatever chances Intel had of being a dominant mobile player, and that postpones the foundry movement towards the end of the decade, when TSMC will be even bigger than today.
I may be well be wrong and 10nm turns out to be the node that will deliver these things, but I'm not optimistic here, really. 10nm is already delayed and we are hearing about 14nm issues by the time we were supposed to be discussing its phase out. 10nm is high risk as well.
Yea, I agree. I am far from an expert on this, but IMO, intel needed 14nm to be a compelling product in mobile, cheap and efficient enough to overcome their lack of integration. Instead they got delay after delay, poor yields, and mediocre performance. Their is a lot of criticism of Sklylake for lack of performance gains, but to me, nothing is more disappointing than 14nm atom.
I find it worrying that they are still talking about 14nm ramp up difficulties. Broadwell has been on the market for a year now.
The entire premise behind the contra-revenue program is that they would subsidize 22nm mobile products because once 14nm products arrived these costs issues would not exist anymore, and 10nm would be even better. Today Intel isn't confident enough on what (and when) 10nm will deliver to the point they are scaling back their mobile efforts, and not a single word about their foundry efforts. It's a real mess.
So they're giving up?
That density normalization argument really backfires. It essentially shows that despite having inherently higher density at process level, that the actual product having less density due to sub-optimal architecture. I bet these slides are provided by process technology to prove "Hey, it is not our fault".
No, it actually doesn't. Do you know what designing a CPU architecture to run at twice the frequency does to density?
Yes i do know. First of all move A9X to low vt high performance process, up Vcc and there you are. A frequency difference of about factor 2 can be achieved by the measures i mentioned above without changing uarch. And A9 seems to have enough thermal headroom to do so.
So they're giving up?
I think part of Intel's problem is that they are trying to do mobile on a process that spans to 4 Ghz+. So for Cannonlake SoCs to really work they are going to need to get aggressive on density and voltage. Perhaps they are unwilling?
Intel needs to lift the foot off the pedal a bit. And go on a 3 year product cycle.
I.e. 2016,17,18 skylake 14nm and those 3 years work on getting 10nm right. Then 2019 starts the year of the new CPU.
Releasing new CPUs that only offer 5 percent gains is just going to lead to product glut,customer confusion and shortages as the process changes end up being harder than expected.
People don't upgrade on a yearly cycle anyways.
And the people buying new computers hold off because of lack of new CPUs but won't buy last years CPU because they see the new one just came out.
The entire premise behind the contra-revenue program is that they would subsidize 22nm mobile products because once 14nm products arrived these costs issues would not exist anymore, and 10nm would be even better. Today Intel isn't confident enough on what (and when) 10nm will deliver to the point they are scaling back their mobile efforts, and not a single word about their foundry efforts. It's a real mess.
Yeah, I think you're right. But remember that they still have prioritized mobile over desktop on 14 nm.
Having two different process techs on the same node, but optimized for low power vs high performance would probably be good. The foundries have that. And I actually though Intel used to have it too, but I'm not certain.
The notions that one comes across on this forum certainly are amusing at times, aren't they?"Just up Vcc and there you are." If it only was that easy.
So here's an interesting point that most don't realize - per-core power allowance is actually quite comparable between mobile and many-core server SKUs for Intel. It's actually desktop which is outlying at ~20W+ per core versus 5-8W for mobile and server.Yeah, I think you're right. But remember that they still have prioritized mobile over desktop on 14 nm.
Having two different process techs on the same node, but optimized for low power vs high performance would probably be good. The foundries have that. And I actually though Intel used to have it too, but I'm not certain.
"Just up Vcc and there you are." If it only was that easy.
Doing both are not going to get you the gains you dream.Could you please stop quoting me out of context? I was stating that you up Vcc in addition to moving to low Vt process. I was never claiming you are getting +100% performance with Vcc alone.
Doing both are not going to get you the gains you dream.
I am not telling you what we are going to get, i am just stating that we are getting precisely the gains i claimed. And that is not even considering binning from the fast process corner. So stop telling me that's not possible.