The 32nm Fab process was canned at TSMC long time ago, and resulted in both AMD & NVIDIA having to squeeze 32nm designs onto existing 40nm PCBs. The TSMC 28nm process is up and running, albit not yet at full production capacity. AMD tapped out its Southerm Islands 28nm designs in Feb this year. The scene is set therefore for a shift from 40nm to 28nm cards.

The perceived effect at the Cruncher end will be massive. There is always an inherent speedup as the design gets a shrink, but this time its going down two steps. A bare minimum speedup of two times 40nm speeds will occur due solely to the shrink. There will be more, and is likely to be 3 or 4 times the 40nm speeds due to design improvements over and above a simple shrink.

The end result here will be to widen the Card capabilities even further, and create a need for another size of WU if the overall WU throughput on the server is to be held in check. A Dual 5870 will get through a Huge in around 14-16mins depending on o/c. That time will go down to 7-8 mins on an equivelent dual 7XXX card, highly likely to go below 5mins. A 7XXX equivelent to a twin 5970 will go through them in around 2-3 mins, maybe a little less.

How scalable is the Core app to a size increase to (say) double, maybe triple the current Huge size? If it is, maybe worth putting one on the drawing board to see how it rumbles, if its not already in the frame for the future. Full Production predictions are a little spaced out for 28nm as the respective Marketing departments get into full swing on 28nm, but a Q4 7XXX is possible (claims of earlier by AMD, but it will not be full production), with Q1 2012 likely, and full production Q2 2012. 5XXX is already EOL, and its now hard to get 5970s from outlets, classic sign of impending new products.

It can be viewed that this is all 9 months away, but its equally possible its only 3 or 4 months - end truth will lay in the middles somewhere. The effect here however will be marked, and a "son of Huge" will be needed, with the inherent need for a couple of months full debugging in production use (5XXX and 6XXX will do the latter easily enough). There maybe an architecture change, but its likel;y to be minimal as such, 6XXX had the key changes, so its vertually certain if a new WU runs on 5XXX and 6XXX, it will on 7XXX.

Regards
Zy

The 32nm Fab process was canned at TSMC long time ago, and resulted in both AMD & NVIDIA having to squeeze 32nm designs onto existing 40nm PCBs. The TSMC 28nm process is up and running, albit not yet at full production capacity. AMD tapped out its Southerm Islands 28nm designs in Feb this year. The scene is set therefore for a shift from 40nm to 28nm cards.

The perceived effect at the Cruncher end will be massive. There is always an inherent speedup as the design gets a shrink, but this time its going down two steps. A bare minimum speedup of two times 40nm speeds will occur due solely to the shrink. There will be more, and is likely to be 3 or 4 times the 40nm speeds due to design improvements over and above a simple shrink.

The end result here will be to widen the Card capabilities even further, and create a need for another size of WU if the overall WU throughput on the server is to be held in check. A Dual 5870 will get through a Huge in around 14-16mins depending on o/c. That time will go down to 7-8 mins on an equivelent dual 7XXX card, highly likely to go below 5mins. A 7XXX equivelent to a twin 5970 will go through them in around 2-3 mins, maybe a little less.

How scalable is the Core app to a size increase to (say) double, maybe triple the current Huge size? If it is, maybe worth putting one on the drawing board to see how it rumbles, if its not already in the frame for the future. Full Production predictions are a little spaced out for 28nm as the respective Marketing departments get into full swing on 28nm, but a Q4 7XXX is possible (claims of earlier by AMD, but it will not be full production), with Q1 2012 likely, and full production Q2 2012. 5XXX is already EOL, and its now hard to get 5970s from outlets, classic sign of impending new products.

It can be viewed that this is all 9 months away, but its equally possible its only 3 or 4 months - end truth will lay in the middles somewhere. The effect here however will be marked, and a "son of Huge" will be needed, with the inherent need for a couple of months full debugging in production use (5XXX and 6XXX will do the latter easily enough). There maybe an architecture change, but its likel;y to be minimal as such, 6XXX had the key changes, so its vertually certain if a new WU runs on 5XXX and 6XXX, it will on 7XXX.

Regards
Zy

At this rate our pc's will soon be able to do a 'full unit' not parts of one that some projects then put back together!

How scalable is the Core app to a size increase to (say) double, maybe triple the current Huge size?

Work unit size can be doubled or even tripled quite easily. I simply need to start producing larger units on the server and current clients will eat them up as is. There's actually two variables that I can tweak, size of one packet (1-256) and number of packets (1-x) in a work unit.

This built-in scalability has already brought us the the four different work unit sizes. At the moment Small work units are under utilized so when the need comes then I think we'll push Huge units even bigger and make the current Normal the size of old Huge. Then Small can handle all the slower cards and Tiny is still optimized for CPU versions.

On the other hand, if the new cards require tweaking or compiling the kernels then that's a job for our upstream (distributed.net). I don't know what their plans are regarding newer cards or run times.

-w