Seeing the huge power consumption on the 8350 I thought why do AMD not copy the 32nm chip onto a 22nm die, just downscale the whole thing to go down to the newer sizes, maybe even 20 or 16 or 14nm, whichever is reasonably reliable at this point.
If they can produce the chips at a smaller process for a reasonable price, it would at least partially solve the huge power consumption of the chips. Would also help with Wendell's idea.
You're misled... The power consumption is not "huge", we're talking dollars per year compared to a haswell. Troll alert?
Shrinking the die would probably lower the yield on the chip.
when you go beyond a certain point with die size shrinks you dont see any power/performance gains under current architectures.
you have to develop completely new techno wonders (finfet etc) to make them work properly and that isnt cheap.
I dont really care about shrinking the current amd architecture down any further as it will never be viable for tablet/mobile.
As a desktop part it's fine... I dont see what all the fuss is about over power consumption
my last pc was a dual cpu xeon and BOTH cpu's drew 125w each and wasnt as powerful as the 8350
I don't understand, isn't the difference much greater under load, besides can't you fit more transistors on the chip by going to a smaller node, thereby increasing performance. I really would have liked to have seen a 4.7GHz Steamroller version of the 8350 which should have about 30% improved IOPs and give performance similar to the 3930/4930, they could have it phenom branded, and do a limited edition sale on the very best binned chips.
Then the 2015 Q1 chip could be a 22nm or smaller excavator chip with greatly increased transistor count, improved architecture, and using a different socket with DDR4 (with quad channel parts?) etc...etc... By then 22nm should be very, very mature enough to have good yields right?
Ok, yes, there is little difference as far as how much you are paying for one of the CPUs, but firstly the cooling is already a problem with single CPUs, let alone racks full of them. Also, the power delivery you need to feed these chips is substantially more than that of any other chip with comparable performance.
In short I'd think there is money to be gained, as moving to a smaller process should not be too hard, as the architecture doesn't need to be modified at all, and 22nm processes are already well established and seem to be quite reliable. The small price to produce the chips at a smaller process should not be more than the amount that would need to be spent on having the supporting infrastructure and electricity bills.
For a company as big as Google, every dollar they can save per CPU is a couple of millions they are saving over a month.
Even at this point I'd expect 22nm to be getting very respectable yields. The biggest differences I would think you get when moving down to a smaller process are firstly the power consumption. When you make the things smaller you will obviously need less current to flow through them, although this is partially reversed by the fact that leakage currents are slightly higher, but still you see steep improvements in power consumpion when you move to smaller processes. The second advantage is that the chips are smaller for the same transistor count. This would significantly improve the latencies experienced over the chips. There are actually several studies done on older 45nm chips where they compare the latencies between those and 100+nm chips, the latencies (somewhat obviously) scale pretty much linearly with the size of the chip.
Mobile viability??? lol
Yeah it should be, but wouldnt designing a new steamroller processor on a smaller node for the AM3+ platform be harder than doing it on the existing node size? Then a node shrink and architecture improvements to 22nm and then 16nm (which should be mature by then) could help them sustain a tick, tick, tick model for next few generations to close gap with intel, because hasn't intel been shrinking the die to get more processors out of each wafer for more money?
Wouldnt a steamroller version of the 8350 with their claimed 30% increase in IOPS at same frequencies produce greater energy savings relative to processing capacity?
APU's
dude.. die shrinks don't help cooling problems, they just make them worse. and intel ahd to make entirely new types of transistors for 22nm, and you can bet, sure as hell, that they patented it up the wazu.
You don't just copy and paste your existing designs into a smaller die. A different manufacturing process involves a lot more than changing the size of the transistors. It's a whole different process. Thus the term "manufacturing process". Your designs have to be modified to account for that, and changing any detail of a CPU design means it has to be tested and retested hundreds or thousands of times, then bugs need to be worked out, rinse and repeat.
Also, Intel is the only company who makes 22nm processors right now. They probably wouldn't allow AMD to use their fabs. The two companies who manufacture AMD's stuff, Global Foundries and TSMC, are working on 20nm processes of their own, but nobody is really sure when they will be available for mass production.
Also, AMD's CPUs do not have a "huge" power consumption. An 8350 uses more power than an i5, yes, but it is also an 8-core processor (and don't let anyone tell you otherwise). Some of that power difference is also just the higher clock speed. Temperatures and power usage increase exponentially with clock speed, so the 600mhz difference also makes up a portion of that.
An 8350 with 4 active cores and running at 3.4ghz probably doesn't use even 10% more power than an i5. Where AMD does have a problem, and you can argue all day about the cause of it, is that under those circumstances, the 8350 would only be about 75% as fast as the i5.
What I am getting at is they have been reducing die size whilst maintaing similar transistor counts to get more processors per wafer as node size has shrunk in order to make more money right? In otherwords, they are cutting their product up and making it smaller because their is not much competition in terms of performance.
And that if AMD started pushing agressively on the desktop segment, that intel would have to respond by not cutting down their processors so much?
I believe TSMC starts mass production @20nm 2014/1Q.
APU is not 8 core...
8350 is not APU
Not sure which direction that's going
Not the cooling problems as far as chip temperatures... Cooling difficulty due to the large amount of heat dissipated by the chip.
Obviously your power density is very similar on different processes as the transistor densities are higher although the power used per transistor is lower. Overall you will be feeding less power to a smaller chip, which will probably run just as hot, but be easier to cool due to the lower power.
Well my question at this point is how hard is it to move the 8350 to a 22nm process.
I would have expected it would be quite easy as you would not necessarily have to change any of the layout, instead you could just scale the whole chip.
Even if the per core power consumption is lower than intel CPUs the per core performance is lower for the same clock speed, meaning the performance per watt is lower. Either way there is space for improvement as far as the power consumption goes, and any cheap solution to that would be great for server systems based on the chips.
First, AMD has to determine that there is a market for it. That's probably the biggest hurdle.
Second, AMD has to basically recreate the Piledriver architecture to work with the 20nm process. You do have to change the layout because all of the parts don't scale the same. 20nm doesn't mean everything is at 20nm.
And last, some fab company outside of Intel has to be able to mass produce 20nm chips. That's not the case right now, but it will be shortly.
After all that, the power consumption wouldn't change much, if at all. The performance would not improve.
That stuff takes a new architecture, which is what Steamroller is for. Whether there will be an 8-core FX Steamroller is unknown right now. It's not on any public roadmap they've released.
I don't see a problem with the AMD processors now. If this is the best they can do to keep it the price that they are for the performance, then so be it.
I'd prefer the more power and more heat right now as opposed to less power and less heat, but less performance for the dollar. That's why AMD still exists.
People are saying that AMD wants to drop FX series and just have 6 and 8 core black-edition A10's instead. I have no comment on this nonsense.