ELSA hints GT206 and GT212

...they'll have to re-architect the MC and at the very least re-organize the ROPs into partitions of 8 units. (...) The other option is as you suggest - to maintain ROP organization of 4 per partition, but re-architect the ROPs themselves to create "fat" ROPs with higher fillrates than their current ROPs.

Both options have drawbacks. The first option would create an abundance of fillrate for entry-level GT2xx-derived SKUs because the base unit for ROP partitions would now be 8 instead of 4.
What I meant is that 8 ROPs or 4 fat ones are essentially the same, you just double everything. So the problem with lower-end parts will be there all the same. I do, however, see a solution. It's not necessary to use the new ROPs, they can use the old ones. I don't think they'll go putting GDDR5 on cheap cards, they'll probably stick with GDDR3.

But it's not a perfect solution either. In 2009, ordinary DDR3 will compete with GDDR3 and will probably be cheaper (because it will be made in large quantities for the new platforms from both Intel and AMD). Does nVidia's memory controller support DDR3? ;)
ShaidarHaran said:
It will be interesting to see what route they take. Maybe they'll choose an altogether different option. I've seen some suggest a 24 ROP/384-bit MC arrangement coupled with higher clocks to hopefully offset the lower fillrates. I somewhat doubt the 55nm process would enable the sort of clockspeeds necessary to achieve ~ parity with GT200's fillrates, though.
Are we talking about GT206? I can certainly imagine it with 24 ROPs and 384bit interface. Sort of, GT206 would be to GT200 what G92 is to G80. The problem in my opinion is that production of GDDR3 will gradually cease and we could see the same phenomenon as with old SDRAM and DDR SDRAM modules: when they became obsolete, they kept selling, but the prices were way higher than the then-mainstream parts. I'm not sure, but maybe their price even rose due to being EOL'ed.
 
I'm imagining GT206 with 16 ROP/256bit and 1GB low grade GDDR5, good core clock; GT212 with 24 ROP/384 bit and 1.5GB higher grade GDDR5 (maybe there won't be much mem clock difference).

Lower end GT206 card with fast GDDR3 (isn't 1.1GHz nice?), full 256bit bus, one cluster disabled, lower core clock. Likewise, a 320bit 1280MB GT212 with one cluster disabled.
 
What I meant is that 8 ROPs or 4 fat ones are essentially the same, you just double everything.

I'm still expecting to hear what there's to be "doubled" exactly in the current ROPs. If you bother to look at the two links from hardware.fr I posted before, their only problem is with 8xMSAA and that on G80/92/200 irrelevant of the amount of ROPs each architecture has. At least for GT200 neither the 1GB framebuffer nor it's 140+GB/s bandwidth are holding it back with 8xMSAA.

Neither Z-fillrates nor G/samples show anything lacklustering in NV's current ROPs (always up to 4xMSAA). Au contraire their by quite a bit higher than AMD has for each competing GPU.

I'm imagining GT206 with 16 ROP/256bit and 1GB low grade GDDR5, good core clock; GT212 with 24 ROP/384 bit and 1.5GB higher grade GDDR5 (maybe there won't be much mem clock difference).

Lower end GT206 card with fast GDDR3 (isn't 1.1GHz nice?), full 256bit bus, one cluster disabled, lower core clock. Likewise, a 320bit 1280MB GT212 with one cluster disabled.

Agreed; on a pure speculative basis if such a 206 would come with say 1.8GHz GDDR5@256bit it would end up at around 115GB/s and a 212 with say 1.7GHz GDDR5@384bit would end up at around 160GB/s. I'm still trying to figure why the amount of ROPs is so important in the end.

The only downside with 320/384bit wide busses is that the driver team would want to rip its hair out due to the different ram confirugations.
 
Last edited by a moderator:
Maybe. Maybe not.

:LOL: ok

Have you seen any confirmation of tape-out of either GT206 or GT212? I sure haven't. GT212 is on 40nm anyway, a process that is not yet ready for volume production so I doubt we'll be seeing anything which utilizes it next month ;)

The part that is launching next month is GT200b. It's just common sense.
 

Do you have information that indicates otherwise? Why cancel a project that's already taped-out for one that hasn't? There is a part due out next month, and it's not GT212. That leaves GT200b. Unless GT200b = GT206, in which case your attitude is wholly uncalled for. A change in internal naming convention isn't exactly information the public would be privy to, now is it?
 
Have you seen any confirmation of tape-out of either GT206 or GT212? I sure haven't.
That doesn't mean anything.

GT212 is on 40nm anyway, a process that is not yet ready for volume production so I doubt we'll be seeing anything which utilizes it next month ;)
GT212 is early 2009 part at best.

The part that is launching next month is GT200b. It's just common sense.
As i've said, maybe. Maybe not. We'll see.
 
That doesn't mean anything.


GT212 is early 2009 part at best.


As i've said, maybe. Maybe not. We'll see.

Like I said, the only way we'll see something other than GT200b next month is if NV internally re-named GT200b to GT206. There just hasn't been enough time since GT200 to have come up with something radically different and already have it "in production".

I don't believe NV had enough foreknowledge of GT200's yield rates to start design on a true successor early enough to push it out this quickly.
 
I'm still expecting to hear what there's to be "doubled" exactly in the current ROPs. If you bother to look at the two links from hardware.fr I posted before, their only problem is with 8xMSAA and that on G80/92/200 irrelevant of the amount of ROPs each architecture has. At least for GT200 neither the 1GB framebuffer nor it's 140+GB/s bandwidth are holding it back with 8xMSAA.
Yes, the huge drop is a bit strange - you'd expect it to achieve the same zsamples/s throughput at msaa 8x as at msaa 4x (like the rv770 or rv670 - heck the rv770 even needs at least msaa 4x to achieve its peak zsamples/s throughput), but there's a factor 3 drop. Maybe z buffer compression doesn't work with 8xMSAA (would be strange though)?
In any case, I can spot another deficiency of the ROPs, that would be (4 channel) FP32 blending. Performance drops to 1/8 that of FP16 blending and consequently is way below what a rv770 can do. Not sure if that would really matter somewhere though, and the rv770 sacrifices FP16 blend performance instead (which is now well below that of rv670).
 
I don't believe NV had enough foreknowledge of GT200's yield rates to start design on a true successor early enough to push it out this quickly.

Not that I disagree with your general statement, but other than conforming to the rules of English grammar, this particular sentence doesn't make, uhm, sense. Why would yield rates have anything do with the decision to design something earlier or later?

Pretty much all designs are started long before yield rates are known. You just assume that the mathematical models are more or less correct... and they usually are.
 
The only downside with 320/384bit wide busses is that the driver team would want to rip its hair out due to the different ram confirugations.

isn't that sorted out by now? they fixed trouble with the 8800GTS 320 (though performance limits remain, just aren't anormal as before), they have cards in both 512MB and 1GB models, and they already have 9600GSO and GTX 260 with differing buswidth.
 
Not that I disagree with your general statement, but other than conforming to the rules of English grammar, this particular sentence doesn't make, uhm, sense. Why would yield rates have anything do with the decision to design something earlier or later?

The low yield rates are one of the single biggest factors in NV's decision to transition away from GT200 ASAP, by my understanding.

Pretty much all designs are started long before yield rates are known.

Which is my point. NV had no way of knowing that GT200 would yield so low when they were at a point to begin a project on its successor, therefore any successor could not be a radical departure from GT200 but merely a die shrink with an optimized floor plan, perhaps.

You just assume that the mathematical models are more or less correct... and they usually are.

Usually, apparently not in this case, however. Unless rumors about GT200's yield rates from the likes of DigiTimes are incorrect.
 
Back
Top