The New and Improved "G80 Rumours Thread" *DailyTech specs at #802*

Man I'm just chomping at the bit to see more info on this beast.

 

G80 least can be sealed as more than 1.5X times 7950GX2 performance .

 

One thing to be careful of is that if the ALU's do indeed run at twice the clockspeed of the rest of the core, we might hear some pipeline counts that are actually "effective" pipeline counts: that is, there are actually half as many, but running at twice the speed.

Just food for thought.

 

This explains the bus size differences from the gtx to the gts, possibly each mini gpu will have access to its own bus. Hmm a ring arrangment for the memory control sounds like a possibility too

 

no mini GPU. only one die for sure.

 

Only in theory, lets see it in real world, driver work is going to be very important specially if they can't get the performance up to that 1.5- 2 times its going to be hard to compete with r600

 

Oh yeah saying what geeforcer said like the cell for the gpu

 

38.4 billion texels per second:

• 8 non-filtering pipelines, running at 1350MHz - used for vertex fetch, texture buffer fetch, constant buffer fetch = 10.8G "texels" per second
• 48 filtering pipelines, running at 575MHz - yer "bog-standard" bilinear, trilinear, AF texture pipelines = 27.6G texels per second

Total 38.4G texels per second.

Jawed

 

OK, naive question here, doesn't double-pumping the shader ALUs make the problem of memory latency (effectively) twice as serious?

 

Frankly speaking, this turn I think it will be hard for R600 to compete with G80

 

I see what you mean but perhaps there's no direct connection, you need a narrower bus cause you have less ALUs (and TMUs and ROPs probably..).

 

We've confirmed something?

 

Didn't Intel learn not to do this earlier w/ netburst.

 

If your design is already fundamentally able to hide all normal cases of latency (e.g. bilinear filtering) then clocking is just a number to play with.

As far as I can tell it's a matter of register file organisation, versus SIMD-per-clock versus pipeline length, versus thread-control complexity (thread despatch, queues).

Having said all that, G80 sounds hairy. A whole new kind of hairy. Default-scalar ALUs is, to me, a bigger thing (on this scale) than relative clocks.

Jawed

 

well you can perceive" ATI 80nm " form RV560/RV570 . both of them are unoverclockable.

 

I see no difference between double pumping the ALU's and doubling the number of pipelines, performance-wise. Remember that the ALU's themselves won't need to have any direct access to external memory. There is, however, the potential benefit of reduced die size for the same performance.

 

The memory channels are mostly likely not reserved for specific stuff. Each memory channel instead supplies it own 'processor' block that includes everything like ROP, TMU, ALU etc. I now believe in a Cell-like architecture where each of these groups maybe can share some bandwidth on a grand ring memory bus.

Call it multi-core if you like.

G80 GTX = 6 cores
G80 GTS = 5 cores
etc

 

with the way drivers are right now, the r600 shouldn't have any issue with pumbling the g80