Wednesday, January 25th 2012
NVIDIA Taunts AMD's GCN Architecture Performance
As AMD's Radeon HD 7900 series is finding ground in the market, and NVIDIA's competitive product line still without a concrete launch schedule, the mind games have begun. In an interview to NordicHardware, a senior NVIDIA official said that NVIDIA expected more from AMD's new GPU family. "Honestly, we expected more from our competitor's new architecture," the official said, indicating two interrelated things:
Source:
NordicHardware
- AMD's Southern Islands GPU family's performance levels are well within NVIDIA's expectations
- NVIDIA's new architecture will be a lot more powerful than Southern Islands, because it was prepared keeping in mind a faster architecture from AMD than what Southern Islands ended up being
87 Comments on NVIDIA Taunts AMD's GCN Architecture Performance
Chill out.:twitch:
Anyway, for you reading pleasure:
www.tomshardware.com/reviews/processor-architecture-benchmark,2974-15.html
en.wikipedia.org/wiki/Transistor_count
This really reminds me of the GTX 480 vs HD5870 match up. In which Nvidia rushed their chip into production just to gain back some of the product market. But what happened??? Well because Nvidia did this we got a GTX 480 (Which was really an unfinished GTX 570) that ran way to hot, showed up to the customers house DOA on a regular basis, and if not DOA it would often die shortly after little loose.(Of course there are a lot of exceptions I am just using blanket statements) Not to mention just being a huge power hog.
In the mean time.... AMD designed the HD6970 which while didn't beat the GTX 580 when finally released it did a really great job in the price vs. performance category.
The way I am speculating this will play out is Nvidia will release the GTX 680.... Which will be really a GTX 770/760. This will compete very well against the HD7970. But because it was rushed I believe it will have a great deal of many problems. Even if the GTX 680 doesn't end up having any issues it will still buy more time for AMD just to release another version (A better higher clocked version) of the HD7970.
While I have no doubt Nvidia will have no problem beating AMD's 25% to 30% performance lead.... you got to wonder if it was AMD's strategy.... release something conservative just to keep Nvidia on their toes and push them to release something before they are ready.
Of course I have nothing telling me this is how this will play out. It is all just speculation. But people say that history repeats it's self and I think it will again in this instance.
Sure most people don't care about transistors. Most people are stupid too.
GPUs are made of transistors. Whether people care about transistors or not, they are key and if a 3 billion transistor GPU (GF110) is so close to a 4.3 billion transistor GPU from the competitor, you better bet that a GPU with 50% more transistors and similar architecture will smack such competitor.
Maybe you can't warp your simple mind around the concept that GPUs are made of transistors and that more transistors means you can create more powerful GPUs if the architecture accompanies, but that's how it works.
GCN is the first time that AMD uses 1D shaders and that's one of the reasons their transistor count went up so badly. GCN is the fisrt AMD GPU to support half rate double precision, which requires a lot of transistors to implement. Transistors yes, not fairy dust. GCN also implements a lot of cache and memory management features, which also require a lot of transistors. And all of the above is why Tahiti is so big for so little performance gain.
Fermi already had all of this, or most of it, so Nvidia does not need to spend horrendous amounts of transistors again, on top of the ones they already spent on Fermi. Unlike GCN vs VLIW, Fermi vs Kepler won't be characterized by a large increase in transistor count per work unit. It's simples, simple math. Now get your head out your ass and start thinking, before calling others fanbois.
GK112?: Kepler 32 SMs -> 1024 CUDA Cores
GM110: Maxwell 128 SMs -> 4096 CUDA Cores
So, abouts from official sources that talk about Maxwell being ~10 TFLOPs
Fermi -> 1581.056 FMA GFlops (512 x 4 x .772) 512 Cores at 772MHz
Kepler -> 3481.6 FMA GFlops (1024 x 4 x .850) 1024 Cores at 850MHz
Maxwell -> 9945.088 FMA GFlops (4096 x 4 x .607) 4096 Cores at 607MHz
Estimations by Seronx^ not related to official sources above
Calculations based on this and official sources
Time will tell...
GK104 is a 16 SM part like Fermi
probably less actually in price since it is just Same Trannies + Clock Rate increase I would say GK104 will be
512 CUDA Cores with a clock near ~950 or so
512 x 4 x .950 = 1894.4 FMA GFlops
vs
512 x 4 x .772 = 1581.056 FMA GFlops
1894.4/1581.056 => ~1.20x
and if it has more SMs even better each SM you add though will probably decrease the clock about ~15-33MHz
G104 has two rumors about it is SM count which is either 16(512) or 24(768)
768 x 4 x .686 => 2107.392 FMA GFlops
GK104 -> 300-360mm²
Just saying
Most people are stupid as they don't care about transistor count? All hail the elitist tech master! :respect: I realize most people are a little nerdy on this site, but you my friend are as disconnected from reality as they come! :shadedshu
Fact is, AMD showed up first to the party with their new Lamborghini 7970. They got the party started, all the bitches took turns taken A ride. Come 2:01am Nvida is going Show up in their new Ferrari ... but its too late all the bitches are drunk and not going care. Oh whats this, AMD went and got their new 7975, 7980,7985 .. (cause we all know they are coming, Stocked OCed verson).
Late is Late.
512 CUDA Core GK104(100) will be around 950MHz with a TDP around 200Ws
Nvidia is disappointed because they were going to release a 2048 CUDA part with the 1024 CUDA part
But they can just stick with the 512 CUDA part and just clock it really high and keep TDP within 200W
---Remember this----
GeForce 8800 GTX - 90nm
PCIe 1.1 x16
484mm²
575MHz
128 CUDA Cores
GeForce 9800 GTX - 65nm
PCIe 2.0 x16
324mm²
675MHz
128 CUDA Cores
---------------------
Well they are doing this here:
GeForce 580 GTX - 40nm
PCIe 2.0 x16
520mm²
772MHz
512 CUDA Cores
GeForce 680 GTX - 28nm
PCIe 3.0 x16
~300mm²
~950MHz
512 CUDA Cores
The sad thing is gents.... I don't think these cards will release in February like some rumors have said. I have my reasons for saying this. If anything.... I think we will be seeing them around April.