-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
ATI's next generation plans outed - Before Nvidia gets GF100 finished
- Thread starter bhavv
- Start date
Problem is - on 40nm the GF100 even with a very tweaked respin isn't going to yeild massive increases in performance or reduced power/heat on the high end parts... the design really really needs to be on the right process not shoe horned onto something it was never designed for.
The only thing they are likely to achieve from tweaked respins is more economical low and mid range parts, with somewhat reduced power needs/heat output. (I kinda hold out hope for a "GTX260" style part that has massive OCing potential
)ATI has its own problems tho, doubling up the numbers on evergreen to go to next gen would give a much smaller performance increase than the 4870->5870 jump.
Last edited:
Problem is - on 40nm the GF100 even with a very tweaked respin isn't going to yeild massive increases in performance or reduced power/heat on the high end parts... the design really really needs to be on the right process not shoe horned onto something it was never designed for.
The only thing they are likely to achieve from tweaked respins is more economical low and mid range parts, with somewhat reduced power needs/heat output. (I kinda hold out hope for a "GTX260" style part that has massive OCing potential
ATI has its own problems tho, doubling up the numbers on evergreen to go to next gen would give a much smaller performance increase than the 4870->5870 jump.
Though you did say the same thing quite a lot when it came to 4800s > 5800s, which turned out to be incorrect.
We simply don't know how they'll perform and speculating in such a way, as you should now be aware of, doesn't mean much.
I think your mis-remembering what I said, I said that there are diminishing returns in the scaling of ATI's current architecture and that 4800-5800 wouldn't be a consistant doubling of performance - and theres not its more like 1.6-1.7x or so on average, and moving from 5800->6800 (or whatever) again wouldn't double the performance if you doubled up the numbers it would be more like somewhere in the region of 1.5x the performance and if you did it again probably more like 1.3x.
There is some speculation as to the exact numbers sure - but the diminishing returns are a real issue and somewhere in that ballpark, which is why ATI were going for a next generation design with NI.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
You seem to be ignoring a couple MAJOR points, it will be bigger with MORE power. IE if they bump shader power by 50%, it would be 50% bigger, 35% faster, and still smaller than a Fermi with still higher yields, uh oh.
Do remember that for Fermi to be released in a more powerful second version or refresh with more shaders, it too would have to get bigger. Except when you're sub 20 cores per wafer at their current size, they clearly can't go bigger till a die shrink.
A very tweaked respin, call it a B1, if they do it(probably will) will actually get around 10% bigger for the SAME shader count because the two basic problems with 40nm, inconsistant quality via's, and inconsistant transistor size are pretty simple to fix. Via's requires simply putting 2 or more in instead of one, they are very small, it still increases die size. As for the other fix, its simply spacing them out a little further so if some transistors overlap due to inconsistant sizes its less of a problem. Both fixes require more die space, that will reduce possible dies per wafer, but should increase yields a decent amount. They still won't come close to Cypress yields, because its 40% smaller(which would become about 50%).
Thats Nvidia's biggest problem, AMD has room on 40nm to go 30-40% bigger in die size and stay around 225W, Nvidia has NO space on 40nm for any added core logic.
A tweaked Fermi will have better yields, NOT better performance, the via's and spacing out fixes for 40nm increase yields, it doesn't reduce power, it doesn't increase possible speeds, it fixes yields and thats it.
A B1 won't beat a 5870, let alone a 5870 with +30% power.
For die size, prices and complexity SI is unlikely to be a massive leap, there just isn't the power or size available on 40nm for a double shader part. With some more lessons learnt and a few more power tricks and increasing size marginally they can bring a probably 20-30% performance boost for little cost to the consumer, but if you're waiting for a 3200 shader part with 80% increase in performance over the 5870, that won't be happening till 28nm.
I'm not sure why people expect massive performance increase from an architecture viewpoint.
For all the architectural advancements in Fermi, it really doesn't perform as anything except a 285gtx at same clocks and shader counts would. Architecture in terms of GPU generally means what fancy new DX features it can do, increases in efficiency to save power and continue allowing shader/pipelines counts to increase, every generation the shader/pipeline count has dictated performance and will continue to do so.
Associate
- Joined
- 10 Nov 2009
- Posts
- 1,099
- Location
- London
But surely improving the design to work around the transistor issues reduces the need to overvolt the chip to fix slightly broken transitors, thus reducing the power requirements. Likewise for clockspeed, a better quality chip will allow a higher frequency for the same voltage/power/heat production.
And if they can increase yield significantly then they also have the option of either producing a better binned part or lowering the voltage/TDP and throwing away/lower binning GPUs that are marginal (but that they are now forced to make work due to the incredibly low yields).
I'm not saying this will happen, overcoming the large die-size issue is a major problem, but there are performance improvements that can be made.
Its one thing to use the extra space to improve yields, reduce heat/power while still maintaining a smaller core size for low and mid-range parts, but scale that back up to a performance part and you end up with a bigger core again.
On this I agree with DM any respin will be about making more economical parts, rather than pushing the performance envelope or massively reducing heat/power on high end parts.
On this I agree with DM any respin will be about making more economical parts, rather than pushing the performance envelope or massively reducing heat/power on high end parts.
Take note that D.P. is implying that a Cypress refresh will be big, hot and power thirsty like GF100s which is something he doesn't know for sure.
I don't think anyone is claiming it won't be larger, hotter and more power hungry than a Cypress GPU, but to claim it'll be just like GF100 in those regards is again grasping at anything to make out like GF100s aren't an utter failure.
Exactly right.
The mere fact of improving yields will allow for higher clocked parts, or parts with 512 SP, or both.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
The problem being as I said, it would end up significantly faster than a Fermi at the same size and this is the problem.
If Fermi was 60% faster with 60% more power used and 60% more expensive, I wouldn't mind, I might get one. Its when its 60% more power, 60% more expensive and 10% faster that it becomes useless.
If AMD make a core 60% bigger than their current one, it would likely use 40-50% more power but also be 60% faster, thats fine.
The problem is Fermi would also need a similar increase in size to increase its shader count, which would simply level out the power/size/performance with the "new" AMD chip, it would still be massively bigger more powerful and more expensive.
D.P, no, none working parts, don't really decrease power, they are cut off, the problem with 40nm is that its leaky as hell on the GOOD transistors, AMD suffer from this on all their working parts as do Nvidia, its a leaky process, even if every transistor was working ITS A LEAKY PROCESS. Fixing yields does not alter the quality of the working transistors.
There would be a rather small decrease in excess power with a few more parts able to come in with lower voltage and lower power, but not many.
As I said in another thread, Fermi isn't hugely power hungry for what it is. It uses about 60% more power for about 60% more size and 50% more transistors. A fix won't reduce the clock speeds, power usage or really anything else, it will increase the amount of cores with working shader clusters, it will do it at the cost of die size aswell. Frankly AMD seem to be suggesting a 10-15% increase in the die size they wanted to implement "40nm" tweaks, add that onto Nvidia's core and you're up to a core around 75% larger.
Even if you can marginally increase clocks, that gives AMD SO much room to manouvre, Nvidia just can't make a part with more shaders, they are at the exact limit of what can even be made and obviously WAY beyond the limit of what can be made at a profit. AMD however could go 50% bigger, with 50% more shaders, and still come in smaller, cheaper and with a 50% performance improvement at 40nm.
Fermi 2, as in a Fermi with more shaders and a significant increase in performace, IE a proper next gen, will not be till 28nm, they can't do it before then. Its looking increasingly like AMD can increase shader count and power before then, and can likely do it further and cheaper if they move to GloFo for 28nm.
Nvidia can't compete on price/performance/yields, till they utterly change their basic design to a small core design. The question is will they, or maybe more accurately, will they do it soon enough.
If you would assume that optimistically a B1 respin would give Nvidia the 512sp's(6% increase) and maybe a 10% bump in clock speeds, well, you're talking about a 15% increase in performance if you're lucky. AMD have the headroom to cram another 50% shaders, maybe more if their new uncore increases efficiency by doing the same work in less space. a 15% bump from a B1 respin just isn't going to compete with a 30-40% performance bump from AMD, and AMD will still be making cheaper chips to do so.
If Fermi was 60% faster with 60% more power used and 60% more expensive, I wouldn't mind, I might get one. Its when its 60% more power, 60% more expensive and 10% faster that it becomes useless.
If AMD make a core 60% bigger than their current one, it would likely use 40-50% more power but also be 60% faster, thats fine.
The problem is Fermi would also need a similar increase in size to increase its shader count, which would simply level out the power/size/performance with the "new" AMD chip, it would still be massively bigger more powerful and more expensive.
D.P, no, none working parts, don't really decrease power, they are cut off, the problem with 40nm is that its leaky as hell on the GOOD transistors, AMD suffer from this on all their working parts as do Nvidia, its a leaky process, even if every transistor was working ITS A LEAKY PROCESS. Fixing yields does not alter the quality of the working transistors.
There would be a rather small decrease in excess power with a few more parts able to come in with lower voltage and lower power, but not many.
As I said in another thread, Fermi isn't hugely power hungry for what it is. It uses about 60% more power for about 60% more size and 50% more transistors. A fix won't reduce the clock speeds, power usage or really anything else, it will increase the amount of cores with working shader clusters, it will do it at the cost of die size aswell. Frankly AMD seem to be suggesting a 10-15% increase in the die size they wanted to implement "40nm" tweaks, add that onto Nvidia's core and you're up to a core around 75% larger.
Even if you can marginally increase clocks, that gives AMD SO much room to manouvre, Nvidia just can't make a part with more shaders, they are at the exact limit of what can even be made and obviously WAY beyond the limit of what can be made at a profit. AMD however could go 50% bigger, with 50% more shaders, and still come in smaller, cheaper and with a 50% performance improvement at 40nm.
Fermi 2, as in a Fermi with more shaders and a significant increase in performace, IE a proper next gen, will not be till 28nm, they can't do it before then. Its looking increasingly like AMD can increase shader count and power before then, and can likely do it further and cheaper if they move to GloFo for 28nm.
Nvidia can't compete on price/performance/yields, till they utterly change their basic design to a small core design. The question is will they, or maybe more accurately, will they do it soon enough.
If you would assume that optimistically a B1 respin would give Nvidia the 512sp's(6% increase) and maybe a 10% bump in clock speeds, well, you're talking about a 15% increase in performance if you're lucky. AMD have the headroom to cram another 50% shaders, maybe more if their new uncore increases efficiency by doing the same work in less space. a 15% bump from a B1 respin just isn't going to compete with a 30-40% performance bump from AMD, and AMD will still be making cheaper chips to do so.