AMD: R7xx Speculation
- Thread starter Unknown Soldier
- Start date
- Status
Well, we could just boil this down to which is more "likely":
- 750 * 480 = 720GFLOPs
- 750 * 800 = 1.2TFLOPs
Jawed
5 FLOPs/ALU in both cases. How about a bit more creative math?
5 FLOPs/ALU in both cases. How about a bit more creative math?
Like 7 FLOPS/ALU ?
- 750 * 480 * 7 = 1.008 TFlops
Jawed
Legend
Definitely.
480:24 (96:24) is the same ALU:TEX as 320:16, 4:1
800:32 (160:32) is 5:1.
I have to say it is kinda entertaining that this still feels "up in the air".
Something that's worth noting: Vantage P scores should benefit much less than X scores regardless of the scale of changes in RV770. This is simply because P is with AF/AA off. So, I'm wondering if the P scores are a misdirection (being about 1.5x)...
Vantage X scores should be 2x, so the remaining question is, can RV770 achieve that solely through 4xZ per clock? Can RV770 do that with 24 TUs, or does it need 32 TUs?
In case it isn't obvious: I'm ignoring the die size (which is definitely in the region of 250mm2) and entertaining the idea that the damn thing is 800:32.
Somone mentioned 480:32 a while back, so I suppose I should throw it in there.
Jawed
Besides the point that the most important thing about it is that RV700 truly should yield a theoretical peak of ~1 TFLOP/s, that scenario above sounds a bit complicated to my layman's eyes. Assume you'd arrange those 480SPs in 5 clusters, you'd end up with 15 FLOPs per ALU. And no it doesn't have to be 4 or 5 clusters at any price, but that the whole number crunching stuff doesn't lead anywhere either.
By the way ATI never used to be that "fond" of MUL calls from what I recall from the past. If they'd add any single FLOP anywhere ADD would be the most likely candidate.
Question
So far r670 was seen as a 4 "processors" device nv92 as a 16 "processors".
Does Nvidia choice could hurt in this as GPU become bigger?
Could communication between the different "processors" start to hurt Nvidia more than ATI in near future (without major architectural change obviously)?
Could this add more complexity on software too (in regard to ATI choices)?
out of thread question.
So far r670 was seen as a 4 "processors" device nv92 as a 16 "processors".
Does Nvidia choice could hurt in this as GPU become bigger?
Could communication between the different "processors" start to hurt Nvidia more than ATI in near future (without major architectural change obviously)?
Could this add more complexity on software too (in regard to ATI choices)?
Last edited by a moderator:
I knew at some point that the whole processor thing would backslap eventually LOL. If I'd start as a layman I'd say that R6x0/RV6x0 has 4 very "phat" clusters and G8x/9x 8 quite "thin" clusters. So far we've somewhat verified that GT200 contains 10 clusters.
I'm still completely intrigued by the fact that we don't have any solid hardware numbers yet, besides clocks
And in its own way, if clocks are lower than previous generations, then ATI must be confident in the rest of the card's ability to make up for lower core clocks since that goes against their trend with previous generations
Anyways, the fact that more than a few people are claiming we'll be pleasantly surprised by the RV770's and the fact that nobody has leaked official hardware #'s on the card yet (when the Gigabyte placards were leaked, the GT200's specs were unveiled while ATi wasn't besides the obvious 512MB-256bit piece) seems to tell me that it's being closely hidden...
And in its own way, if clocks are lower than previous generations, then ATI must be confident in the rest of the card's ability to make up for lower core clocks since that goes against their trend with previous generations
Anyways, the fact that more than a few people are claiming we'll be pleasantly surprised by the RV770's and the fact that nobody has leaked official hardware #'s on the card yet (when the Gigabyte placards were leaked, the GT200's specs were unveiled while ATi wasn't besides the obvious 512MB-256bit piece) seems to tell me that it's being closely hidden...
Anarchist4000
Veteran
750MHz * 96 * ( 1.4*5 ) = ~1 TFLOP
If they're still doing the shader AA then having a little extra math power isn't a bad idea. Throw in the 1.4x shader clock domain with 750MHz and they get their teraflop.
If they're still doing the shader AA then having a little extra math power isn't a bad idea. Throw in the 1.4x shader clock domain with 750MHz and they get their teraflop.
8 TUs/32TMUs and 5 ALU-SIMDs would give you 800SPs.
That formula has one minor mistake
You have to have 2 FLOPs/SP and multiply with 0.625 to reach 1 TFLOP.- Status
Similar threads
