Originally Posted by Abundant Cores
Originally Posted by Particle
I feel like that whiteboard image needs explained. The three columns show memory layouts using traditional hardware design nomenclature. That is, the numbers are: (Mega) Pages x Page Size (Bits). In DRAM, page width is the same as interface width, so that 128x32 chip from the first column is 128M pages x 32 bits wide. That column shows at that bottom what type of memory it is (GDDR5) and how many pieces there would be (8). Consequently, you can figure out what you want to know by multiplying whichever values interest you. For capacity, you would take 128M x 32 x 8. You end up with 32 gigabits (4 GB) of total space and an interface width of 32 bits x 8 chips (256 bit). The columns themselves represent different memory layouts and the values within each column show different capacities for that layout by utilizing different densities but the same number and type of ICs as the other values in that same column. To my eye, I would expect column 1 to represent a high end tier of products, column 2 a mid tier, and column 3 a value tier.
Column 1: GDDR5 x 8 pieces
128x32(x8) = 4GB, 256-bit, GDDR5
64x32(x8) = 2GB, 256-bit, GDDR5
32x32(x8) = 1GB, 256-bit, GDDR5
Column 2: GDDR5 x 4 pieces
128x32(x4) = 2GB, 128-bit, GDDR5
64x32(x4) = 1GB, 128-bit, GDDR5
Column 3: DDR3 x 8 pieces
256x16(x8) = 4GB, 128-bit, DDR3
128x16(x8) = 2GB, 128-bit, DDR3
64x16(x8) = 1GB, 128-bit, DDR3
Thats interesting, if what your saying is right then then the R9-285 with a 256Bit bus (down from 384Bit) would the same performance or slower than the 7970 as it only has 15% more SP's and it dropping 30% memory bandwidth.
I can't see that unless they have completely lost the plot.
Tell me, how many memory IC's are on the 7970?
Why? Nvidia released GTX 580 with around 192GB/s and then released GTX 680 with around 192GB/s. And 7870, while it has lower theoretical, seems to do better in high bandwidth situations.
Note how 7870 has 153GB/s of bandwidth (theoretical) yet it handily outperforms the GTX 670. AMD could probably get away with a 7870 style configuration for memory, save a bunch of die space, or replace that die space with something that's actually needed, like more ROPs and shaders.
7870 doesn't fall away from 7950 at all, and 7950 supposedly has 240GB/s of theoretical bandwidth. 7970 is near 280GB/s depending on models. And 153GB/s of AMD theoretical bandwidth is not enough to slow it down at 1440p 8xMSAA.
I guess what I'm saying is that 256-bit 6ghz GDDR5 for AMD would be more than enough for an AMD GPU to work just fine, even at 1440p 8xMSAA.
It does highly that something is seriously broken with Nvidia GPU's memory controllers, but no one wants to touch the subject. Something with 192GB/s maximum theoretical should not lose to something at 153GB/s theoretical.
153GB/s is around 80% of 192GB/s.