next gen kepler to support dx 11.1, also take a year to rollout all cards

[MUYA]

How does that compare to 7970 theoretically? 80%? 90%? 110%?

[Vardant]

At 950MHz it should end up between 7950 and 7970.

[MUYA]

http://vr-zone.com/articles/rumour-n...ons/14823.html

Originally Posted by VR-Zone

Kepler introduces a radical new architecture but the overall structure may be somewhat similar to Fermi. GK104 features 4 overall Graphics Processing Clusters (GPC) - twice that of GF114. Each GPC has 4 SMs - same as GF114. However, each SM now consists of 96 SP or CUDA cores - again twice that of GF114. Each SM features 8 TMU. The end resut is 16 SMs, 1536 SP and 128 TMU. GK104 continues to feature 32 ROP and 256-bit memory interface like GF114.

The core clock is expected to be 950 MHz for the top bin, slightly higher than HD 7970. As previously rumoured, the hotclock for shaders has been dropped, so the shaders will operate at 950 MHz as well. Of course, the drop in shader clocks is compensated by massively increasing shader count - a similar strategy to AMD. The end result is 1536 SP clocked at 950 MHz for 2.9 SP TeraFlops and a 1/6 DP rate of 486 GFlops.

The 256-bit memory interface now uses 2GB GDDR5 and the memory controller is also improved over Fermi, now clocking at 5 GHz GDDR5 for a bandwidth of 160 GB/s. However, this is still massively short of HD 7970's 260 GB/s.

The die size for GK104 is "unofficially estimated" at 340mm2, which is slightly lower than Tahiti's 365mm2. The first GK104 product is due to be released in early Q2 2012.

Read more: http://vr-zone.com/articles/rumour-n...#ixzz1lyHgCoWA

[Redeemed]

Originally Posted by MUYA

http://vr-zone.com/articles/rumour-n...ons/14823.html

Interesting if true.

[shadow001]

Originally Posted by MUYA

How does that compare to 7970 theoretically? 80%? 90%? 110%?

Actually, it compares nicely to an HD6970 card, as it also has roughly the same floating point power(single precision), the same number of rops, the same amount of memory(2GB), though the HD6970 has 176GB/sec of memory bandwith, while those specs suggest 160GB/sec for the GK104 but it does have more texture units than the HD6970...It should be a pretty good card for 300$.


HD7950/7970 still have a lot more memory bandwith(up to 264GB/sec for the HD7970),3 GB of ram and more dual and single precision floating point math, and real world fillrate because of that memory bandwith advantage.

[Viral]

You can't really compare different architectures that easy, especially when one is not even out in the wild. It will probably perform around a 7950 in most situations where bandwidth limitations aren't too apparent. It could easily do better, I don't know, but unless they have some crazy efficient cache setup it's going to fall short whenever bandwidth is a concern.

It also sounds like they are following suit with one upping the model numbers (or two in this case, technically) like AMD did last gen.

[shadow001]

Originally Posted by Viral

You can't really compare different architectures that easy, especially when one is not even out in the wild. It will probably perform around a 7950 in most situations where bandwidth limitations aren't too apparent. It could easily do better, I don't know, but unless they have some crazy efficient cache setup it's going to fall short whenever bandwidth is a concern.

It also sounds like they are following suit with one upping the model numbers (or two in this case, technically) like AMD did last gen.

Both companies make cards that are extremely memory bandwith efficient these days, and keep in mind that a GTX580 has a 384 bit bus and 192 GB/sec of memory bandwith, and even an HD7950 still has 220 GB/sec of memory bandwith and 3 GB of ram to boot....The only time it wouldn't be bandwith limited is at lower resolutions and lower levels of AA use, where in the first case it's likely the CPU holding things back in many games, and the second case is as soon as AA settings and resolutions are cranked up there, i find it hard to believe it'll be a match for a cards like the GTX580 or even less the HD7950....


It'll still be a great card for 300$, but butting heads with cards costing 450$ is punching above it's weight class by quite a bit, but we'll see in 2 months when they're reviewed i guess.

[Xion X2]

Originally Posted by Vardant

At 950MHz it should end up between 7950 and 7970.

I'd rather have a 7950 or 7970 considering it has another 1GB of VRAM on the card.

I don't understand why Nvidia's doing this. One of the nicest ways to upgrade graphics on today's games are texture mods, and 2GB isn't going to get you a long way anymore.

I'll pass.

[Robs07M6S]

Originally Posted by Xion X2

I'd rather have a 7950 or 7970 considering it has another 1GB of VRAM on the card.

I don't understand why Nvidia's doing this. One of the nicest ways to upgrade graphics on today's games are texture mods, and 2GB isn't going to get you a long way anymore.

I'll pass.

Yeah my next card must be 3gb's in order for me to bother with an upgrade and if Nvidia doesnt have an answer to that soon enough then I know someone else who already does.

[mojoman0]

seems gpu memory is fairly vendor specified, so likely evga etc. will likely have a higher mem sku available eventually correct?

[Vardant]

It has 256-bit bus, it's either 2GB or 4GB. Considering, that except the name, that might still change, this is over-all still a mainstream card, then 2GB is just fine.

[Xion X2]

Unfortunately, since their mainstream card will have to compete with ATI's high-end until Q3-Q4, then 2GB isn't enough for those who like to mod their games.

You can get a 3GB 580 but they still sell for 580$ on newegg and are in very small quantities. On the other hand, you can get a 7950 3GB with equal or better performance for $460.

I guess it's possible they could release a 4GB version of Kepler mainstream, which is what they really need to do to compete.