NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
That's my thinking too. Only the possible 192bit bus wouldn't make sense in this regard. Unless it's a) either just not true, or b) it's there to enable cheap 1.5GB ddr3 cards which don't suck (but then still had too many rops).
I'd guess so, yes. It didn't impose any serious limitedness on earlier cards in this section of the market. It might not shine as bright in the quadro line-up, maybe.
Yeah. Still can't make much sense out of this. There either have to be more SMs, more GPCs or other changes so the pixel throughput is higher, or it is pointless - which could be after all those memory chips aren't populated192 Bit FTW.
.
Well, Fermi Mobile is finally available! that's cool to!.
http://gaming.do.co.za/topic/4824/r...-the-newly-launched-gtx460m-graphics-card.htm
Store availability is Friday (in Holland at least.)
http://gaming.do.co.za/topic/4824/r...-the-newly-launched-gtx460m-graphics-card.htm
Store availability is Friday (in Holland at least.)
Yeah. Still can't make much sense out of this. There either have to be more SMs, more GPCs or other changes so the pixel throughput is higher, or it is pointless - which could be after all those memory chips aren't populated
At the moment they will use the chips for the mobile parts. GTX460M with 192bit and 1,5GB.
imho it doesn't make more sense in the mobile space neither - that'll just draw more power without much performance improvement (even more so than on the desktop since I'd suspect memory clock will be similar but core clock lower - could be wrong though).
DavidGraham
Veteran
Man , Nvidia really needs to rethink her strategy , this big die philosophy is really getting old , uncompetitive and ridiculous.
This has nothing to do with NVIDIA following a 'bie die strategy' and everything to do with the fact their architecture is not as efficient as AMD's per-mm². GF100 is the consequence of a big die strategy; GF104/GF106 are not.
It could indirectly be argued in G80's case that the excessive focus on the largest family member might have made it harder for them to create very low-end derivatives ala RV610 because G8x had too much one-time functionality that didn't scale down. I'm not even sure that's really true, but you could certainly think so and then argue it's a consequence of the big die strategy. But GF104/GF106? Definitely not, no.
Also keep in mind that GF106 definitely has a 192-bit memory bus, and I *suspect* it has 240 SPs rather than 192. So while bigger than something comparable from AMD, it's not unbelievably massive either.
I think that's an oversimplification. Nvidia's architecture and transistor spend is also targeting a broader set of workloads. Any "perf/mm²" metric should be qualified. Also, AMD seems to be doing better on the density front and it would be interesting to know whether that's due to specific architectural traits or just them being better at overall semiconductor design.
That would mean there are five SMs in the GPC (in contrast to every other GF1xx GPU). Seems unlikely to me.
That's true, so feel free to replace that by "gaming at the time of release and in the following 6 months" or some such (you could argue it should be a longer period, but I'm thinking of the sale window here, not the usage window).
If you want to consider overall perf/mm² given a different focus though, you should consider that different chips may focus more or less on different things (i.e. CUDA with extra features for GF100, Quadro with lots of geometry for GF100/GF104, etc.) and that some of these focus changes may or may not be correctly reflected in the chip (e.g 1 GPC for GF106 and less L1 for CUDA on non-GF100 is good, but not everything is like that especially in features).
As for transistor density, keep in mind that custom design (even structured custom or similar) results in fewer transistors for a given functionality. Place-and-Route tools are very good at putting a lot of transistors/gates in a very small area while restricting process rules and DFM. The problem is they're not so good at making good use for them (especially but not only when optimising for speed). So this is one (but not the only) reason why comparing transistor density between NVIDIA and AMD can be misleading. It's easy to mistake less efficient design for less efficient implementation.
So yeah... it's complicated I suppose
DavidGraham
Veteran
I don't think it's the latter , semiconductor design techniques are not a secret or a dark art anymore , everyone can utilize and improve upon them , if your engineers are not good , you can always hire more skillful ones and offer them more money ,or you can carry out your own research and discover new methods .
If we look at the current situation in the industry , we can place most companies on equal footing in terms of design skills and innovations , where it really differs is the funds , production capabilities and design choices and philosophies .
Besides , ATi has to be way way better than Nvidia to be able to design such a relatively small chips in comparison to relatively large nvidia chips .. and I just don't see that happening , giving the history of both companies .
If anyone has a different insight , I would be more than happy to hear it.
Similar threads
- Deleted member 2197
- Architecture and Products
