tkrow21 Posted November 11, 2009 Posted November 11, 2009 If this doesn't crumple under 3xSLI or 3xCF like the original technologies do, this is going to be amazing. Share this post Link to post Share on other sites More sharing options...
IVIYTH0S Posted November 11, 2009 Posted November 11, 2009 (edited) Did you look at the game benchmarks? Most of them are in the 80% range. I know I know, I hope it doesn't sound like I'm wishing the Hydra a failure....just venting my anger from my economics hw on this silly benchmark If this doesn't crumple under 3xSLI or 3xCF like the original technologies do, this is going to be amazing. Or 4x fire , I'm more interested if they can first perfect dual GPU scaling regardless of condition. Then my 4870x2 will get a little nudge more power, and then comes the 8800GT being upgraded from the occassional physx card, to 3rd GPU Edited November 11, 2009 by IVIYTH0S Share this post Link to post Share on other sites More sharing options...
Waco Posted November 11, 2009 Posted November 11, 2009 I'm more interested if they can first perfect dual GPU scaling regardless of condition. Eh, that'd have to be a completely GPU-bound scenario and even then it's not possible to get exactly 100%. Hell, 85% is about as close to perfect as we're likely to see. Getting another 45% or so with a 3rd GPU would be AWESOME though. Share this post Link to post Share on other sites More sharing options...
Zertz Posted November 11, 2009 Posted November 11, 2009 (edited) Getting another 45% or so with a 3rd GPU would be AWESOME though. Indeed, but I don't think it's going to happen. Keeping one or two GPU's fed with data isn't too bad, but keeping three of them busy? That's much harder. We shall see! Edit: More than two video cards isn't going to happen, the Hydra 200 chip only has 16 PCI-E lanes. Edited November 11, 2009 by Zertz Share this post Link to post Share on other sites More sharing options...
jammin Posted November 11, 2009 Posted November 11, 2009 There are three versions of the Hydra 200: The LT22114 and LT22102 support 2 GPUs, while the LT24102 (which is what they are using here) supports 2/3 or 4. - Lucidlogix Share this post Link to post Share on other sites More sharing options...
Zertz Posted November 11, 2009 Posted November 11, 2009 The LT22114 and LT22102 support 2 GPUs, while the LT24102 (which is what they are using here) supports 2/3 or 4 cards. - Lucidlogix Okay, I worded that wrong. There's only 16 lanes going into the chip. They can have more lanes out (just like NF200), but it creates an obvious bottleneck (just like NF200 doesn't perform quite as well as native PCI-E lanes). Share this post Link to post Share on other sites More sharing options...
jammin Posted November 12, 2009 Posted November 12, 2009 OK, I get ya. Do you think only having 16 upstream lanes will really hamper things once you start adding more GPUs? Share this post Link to post Share on other sites More sharing options...
IVIYTH0S Posted November 12, 2009 Posted November 12, 2009 Okay, I worded that wrong. There's only 16 lanes going into the chip. They can have more lanes out (just like NF200), but it creates an obvious bottleneck (just like NF200 doesn't perform quite as well as native PCI-E lanes). Ouch that does raise a problem doesn't it . Especially with future cards. Share this post Link to post Share on other sites More sharing options...
jammin Posted November 12, 2009 Posted November 12, 2009 Here is the block diagram: Share this post Link to post Share on other sites More sharing options...
Zertz Posted November 12, 2009 Posted November 12, 2009 (edited) OK, I get ya. Do you think only having 16 upstream lanes will really hamper things once you start adding more GPUs? It definitely won't help scaling. Let's say you have 4 cards that can take 16 lanes, but you can only throw 16 lanes worth of data at them. With two cards, it's (currently) not a problem, but I can't see how they can scale well past that with their current silicon. There you go, it's easier to visualize the issue from that diagram. Just add 1-2 more blocks at the bottom while keeping the same 16 upstream lanes and the problem becomes obvious. Edited November 12, 2009 by Zertz Share this post Link to post Share on other sites More sharing options...
The Smith Posted November 12, 2009 Posted November 12, 2009 Okay, I worded that wrong. There's only 16 lanes going into the chip. They can have more lanes out (just like NF200), but it creates an obvious bottleneck (just like NF200 doesn't perform quite as well as native PCI-E lanes). I don't think it works quite the same way. With the NForce 200, I think the CPU talks to every CPU at the same time. It sends the information to all of them. Also the inter-GPU traffic does not go through the natives lanes; it is just routed by the nForce 200. More here. With Hydra, I believe the CPU would be like if it was talking to only one GPU. Then, the Hydra splits the orders and sends them to each GPU separately, as we have seen here. I am not sure if this is equivalent though. I may very well be wrong though. If I am, please correct me. Also there are more benchs here. Share this post Link to post Share on other sites More sharing options...
IVIYTH0S Posted November 12, 2009 Posted November 12, 2009 (edited) Also there are more benchs here. definitely some more impressive results there...though this one made me laugh 2x 260s were worse than the 260 and 4770, I wonder why they didn't do two 4890s?? Edited November 12, 2009 by IVIYTH0S Share this post Link to post Share on other sites More sharing options...
Recommended Posts
Please sign in to comment
You will be able to leave a comment after signing in
Sign In Now