Hydra is an early prototype. There's lots of positives mentioned, but of course little of the negatives. From a traditional perspective, I would wonder about the buffers and how the chip level communication happens. I think some tasks would be hard to divide with the hardware implementation they have. It would be easier in a DX9 situation moreso than a DX10 implementation, where things start getting much more complicated. Defered rendering, tone mapping, specular lighting, Shader AA & AA buffers, all of which I see as major issues.
The demos and info also make me wonder how the tasks are assigned, there's no clear split point so the division of labour of A renders the beams B renders the wall and floor would mean that those items need to be clearly defined. It sounds like the role of Lucids software and hardware is to try to act as a pre-GPU assembler/scheduler, however without shared resource pools it makes some taks very difficult and for GPU 1 and 2 to communicate would be very bandwidth intensive (edit: especially the add-in version). And it would require alot of tweaking to make the assembler efficient for new games, so once again you would need 'Lucid Optimized' titles like 'Xfire/SLi-ready' to get the full benefit.
Also they mention having different generations of cards doing the work with a GF6800 and a GF9800 doing the task together, however they do things like AF differently, let alone the DX generation differences. For the X1K -> HD series you have many more differences, and a few different similarities. Then doing AMD & nV, you could only barely do that in the last generation, this generation would be even trickier unless you change techniques where the two become GPGPUs IMO. they say DX10 and DX11 should be easier than DX9, but I think the exact opposite from a hardware standpoint, and even from an API standapoint, the features in DX10 let alone DX10.1 to me pose a much greater problem for such a method without so drastic change to what they are doing.
Now Raytracing however I can see it being much easiser, however if you simply turn the GPUs into raytracing co-processors in OpenGL/CL or DX11 then really you wouldn't need the LUCID solution anyways, and performance should scale very linearly. All you could need is a CPU (or CPU/GPU) and then salve GPUs acting as SPUs and then something to assemble and write the data to output buffer taking the role of the traditional ROP.
http://www.pcper.com/article.php?aid=607
http://www.pcper.com/article.php?aid=607&type=expert
Sounds great, but I'm very skeptical, especially since the person providing the details at IDF sounds more like a PR guy than a technical person, making difficult task sound like a simple division of labour, like the part where they say:
"Maybe 5 tasks to 1 or something like that; the results are then combined by the HYDRA chip and sent to a single GPU for output." very loosey goosey and alot lilke the promise of Supertiling before they actually tried to implement it in more complex games than the very closed environment of proffesional Flight SIMS.
Right now, I'm very skeptical, but it is interesting if they ever provide more details on how to do the complex stuff.
Oh Jebus, it's not only going to be offered as a MoBo but an Add-in card solutuon (thus would not be limited to just intel etc);
http://www.lucidlogix.com/technology/technologies.html
IMO this would add even more latency & bandwidth concerns, since it would have to use the chipset PCIe lanes 4 way + whatever CPU communication is required. That doesn't sound good at all.