LucidLogix Fires up Multi-GPU Computing With Faster, More Flexible HYDRA 200 Chip

[PP Mguire]

It shouldnt be any less exciting than x58. And besides, if this chip does what its supposed to do, the p55 boards will rape.

 

[pr0n Inspector]

Why would it be adding any latency. It potentially does the same thing as nvidia's chip.

Hydra receives orders from the CPU, splits them into workloads suitable for different GPU's, then send them to GPU's. When the GPU's finished their jobs they sent the results back to Hydra. Hydra combines them then sent it to the primary GPU for display. Why wouldn't this add latency?
Also, NF200 does not do the same thing.

 

[human_error]

Hydra receives orders from the CPU, splits them into workloads suitable for different GPU's, then send them to GPU's. When the GPU's finished their jobs they sent the results back to Hydra. Hydra combines them then sent it to the primary GPU for display. Why wouldn't this add latency?
Also, NF200 does not do the same thing.

Current xfire/SLI:
1. Get draw commands
2. Split commands to both cards (if not doing AFR then dissect the image for the gpus to render their part).
3. Send data to both cards
4. Both cards render their part
5. Card 2 sends its data to card one for combining before sending to VDU.

Hydra:
1. Get draw commands
2. Detect different load capabilities on cards
3. Split commands to both cards (this will use a similar part to AFR i believe, so each card does a whole frame instead of tiled/split frames like super AA can do).
4. Send data to both cards
5. Both cards render their part
6. Card 2 sends its data to hydra which rediects it to card 1 (simple connection, not any latency) and the frame is interjected between frames generated by card 1 to the VDU (again very simple, no real latency).

With both of these setups the overall latency is going to be so close i'd say it will be indistinguishable. The hydra chip is meant to split the directX commands between the cards which are allowed to render the image using their own methods (if different). The image is then simply sent to be interjected between frames by card 1 - due to the hydra splitting the workload properly then you don't have to worry about rejoining parts of the same frame from different cards, or having to assume the cards run at a similar/same speed and having to sync the frames between the two properly (which is the cause of a lot of the overhead in current setups).

To split the data the hydra chip doesn't have to do too much work - once it knows the relative capabilities of each card it can simply direct the directx commands between them with no extra work needed - i.e. if card 1 is twice as fast as card 2 you just do:

Draw 1 -> card 1
Draw 2 -> card 1
Draw 3 -> card 2
Draw 4 -> card 1
Draw 5 -> card 1
Draw 6 -> card 2

Which isn't too expensive as you just direct the command down the apropriate pci-e connection.

 

[Deleted member 3]

I wonder when Nvidia will buy them out......

I doubt they will, Intel is a big investor, I see them buying it before nv gets a chance.

 

[Mussels]

what i'm interested in is how its actually done - if they dont use crossfire/SLI but use their own method, then we could start seeing ram being additive in multi GPU setups (EG, if they take half the screen each, but actually get to use all of the ram per card - 512MB card + 1024MB card stops dropping to 512MB, and gives us 1.5GB of usable ram)

 

[Bjorn_Of_Iceland]

So will this still use "profiles" for efficient performance?

 

[human_error]

what i'm interested in is how its actually done - if they dont use crossfire/SLI but use their own method, then we could start seeing ram being additive in multi GPU setups (EG, if they take half the screen each, but actually get to use all of the ram per card - 512MB card + 1024MB card stops dropping to 512MB, and gives us 1.5GB of usable ram)

From what i've read about the system it is basically an AFR rendering style that they adopt - this is how they overcome compatability issues between different card vendors and is why you can add any combination of cards without worrying about overhead. Looking at how they described the system i doubt the gram will be additive in this implementation, but i do think that the cards would be able to use all of their own ram - even if you had a 512mb and a 2gb card the 2gb can still use of of its ram (the system basically makes the cards act as if they were in a single card config).

I may be understanding their implementation wrong of course but from what i've read this is how it will be working.

 

[Deleted member 3]

what i'm interested in is how its actually done - if they dont use crossfire/SLI but use their own method, then we could start seeing ram being additive in multi GPU setups (EG, if they take half the screen each, but actually get to use all of the ram per card - 512MB card + 1024MB card stops dropping to 512MB, and gives us 1.5GB of usable ram)

This is a second generation chip. There are plenty of demos out there for the first, google it. They use their own method, that's the whole point.

 

[theubersmurf]

Forgot about that, but P55 multiple GPU's just doesn't get me all excited

What I'd like to see, not so much because of the Lucid Logix chips, but more over pcie 3.0, usb 3.0, and sata 3.0, is a refresh of the x58 motherboards. If manufacturers could get rid of the stocks they have and make new ones with those updated interfaces, and hopefully lucid logix chips (if they work as advertised) That would be great.

 

[RallyRoach]

So, if I'm understanding this correctly, say we have 30 frames to render in a second, and a GTX 295 and a 4850 installed on a system with this chip. The GTX is roughly twice as fast as the 4850 on a given game/setting. So for every 1 frame sent to the 4850, the GTX gets 2? The 4850 renders 10 frames, while the GTX 295 renders 20 frames, all in the same second and the Hydra chip just outputs this as 30 frames per second?

 

[btarunr]

Watch! http://vimeo.com/6700209

In action.

 

[Disparia]

Showing output of one GPU on the right, completed frame on the left.

Instead of a brute method of splitting work, the scene is intelligently broken down. Different amounts of RAM per card will work because textures don't need to be in memory on all cards.

 

[PP Mguire]

THe problem though is still within drivers of multi vendors. Putting ATI and Nvidia together on the same platform and making them work together at a driver level is almost impossible.

 

[Mussels]

THe problem though is still within drivers of multi vendors. Putting ATI and Nvidia together on the same platform and making them work together at a driver level is almost impossible.

true, as well as directX levels.

you cant mix a DX11 and a DX10 card and expect to run DX11 games, for example.


as for the rest of it, it gives performance boosts (each card can use all its ram, hence its 'additive' - poor word, but you get the meaning) and it lets you say, crossfire a 3870 with a 4870