BlueFalcon said:
How can HD8970 be 25-50% faster than HD7970 GE on 28nm? AMD has not made a 550mm2 die (ever?) and given their power consumption for HD7970GE was already near 230-240W. Even if with a more mature 28nm node, even if they shaved off the power consumption of a 1050mhz Tahiti XT part to just 200W, I can't see how they can get 25-50% more performance in just 50W, unless AMD wants to go over 250W of power usage. Also, HD7970GE was released on a 6 months more mature 28nm node which allowed them to bump the clocks from 925mhz to 1050mhz. That means it already took advantage of some 28nm node maturity. I am most hesitant to believe in the 25-50% increase from HD7970GE because AMD went from 40nm 389mm2 die size of HD6970 to a 28nm HD7970 365mm2 die and that full node move down allowed them to bump performance about 45-50% but power consumption stayed roughly the same (http://www.techspot.com/review/603-best-graphics-cards/page11.html). Based on that I don't see how they can net 25-50% more performance on the same 28nm node. Last time AMD was stuck on the same node, it was when they went from HD5870 to HD6970 and performance barely went up 15% on average. Most of the increases came at higher resolutions where HD5870 ran out of 1GB VRAM and in games that use tessellation because the geometry engines were upgraded in Cayman. The only ways I can see how a 25-50% boost can happen is if AMD would increase die size way beyond 365mm2 to 500+ (which would be unheard of for the firm), or do a complete redesign of the Tahiti XT core by instead starting with Pitcairn as a leaner gaming chip by dropping most of the double precision compute functionality of Tahiti to reduce transistor waste. Then the starting base would be a 212mm2 chip not a 365mm2 one and they could essentially double everything inside Pitcairn to make a 420mm2 monster lean gaming chip. But given AMD's focus on HSA/HPC, I can't see them ditching double precision of HD7970 on HD8970. I think that's part of their strategy moving forward. With NV things are different because they have a history of making very large chips (8800GTX = 484mm2, GTX280 = 576mm2, GTX480 = 526mm2, GTX580 = 520mm2) and because GK104 is under 300mm2, they have a lot more room to increase performance. GK110 already sells in K20, K20X Tesla parts. NV can easily build a 520-550mm2 Titan card due to their experience and existing proven manufacturing of K20/X parts. Additionally, unlike AMD that already went to a 384-bit bus and 288GB/sec memory bandwidth, GK104 is starved by just 192 GB/sec of memory bandwidth and yet it's not much slower than HD7970GE. If NV widens the bus, that alone could net a 40% performance increase for them. Instead of being an HD4870 vs. GTX280 situation, this could end up being HD2900XT/3870 vs. 8800GTX all over again. Not sure if this is legit but if true, Titan could be a monster: http://www.techpowerup.com/179605/First-NVIDIA-GeForce-Titan-780-Performance-Numbers-Revealed.html |
Actually you bring up a lot of good points! However I must fix a few things you said:
-The 680 uses a 256 bit bus so it isn't half what they are going to.
-The 7970 GHz is 10-15% stronger so it IS a noticable difference.
-The 8970 will use 25% more stream processors so there you go on that clear cut buff.
-The 8970 has 50% more ROP's, and 30% more TMU's.
-The overall pixel and texture fill rate are 40-50% faster than the 7970.
-The memory clock is 6500 MHz effective. (Crazy higher than the 7970)
-I have measured my power use durring furmark (While fully overclocked), and it was at 223w (Barely more than my old 6950 at 220w). Rememeber a pcie lane + 6-pin + 8-pin =300w of power. The 8970 still has plenty of power to play with since it is only barely using more than 2 x 6-pins of power. Though I will say they will have to watch out from becoming a 480 ;).
http://videocardz.com/35050/amd-radeon-hd-8970-rumored-specification-analysis
I think the things I mentioned are enough to make a 50% increase in performance theoretically possible. Remember I said 25-50%. I don't think it will be more or less than that, but it could be on either end of the spectrum ;).
