More Details About Lucid's Hydra Revealed - Object Based Rendering, RISC, 4w TDP

[AuDioFreaK39]

More details about Lucidlogix Hydra revealed



Uses object based rendering

Last night we had an interesting chat with Nir Cohen who is the senior product marketing manager at Lucidlogix, as he was vistiing Taipei to talk to some potential customers and he gave us a few more details about the Hydra chip and what it can do.

Although he wouldn't go into too much detail, we did find out some new and interesting aspects on what Hydra can do. First of all, the implementation done by ELSA Japan is not the same setup as that which will be found on consumer motherboards, as the Hydra chip can function in two different modes. To make a replacement for SLI or CrossFireX, the Hydra chip has to be configured to run in graphics performance mode, but ELSA's implementation doesn't work like this, as ELSA is using the chip as a PCI Express controller with some additional functionality.

The Hydra chips have a built in RISC processor which makes it a much more flexible solution than your average PCI Express controller. ELSA's implementation is said to be cheaper than using a Tesla server with four cards from Nvidia, but we'll have to wait and see what ELSA's official pricing will be until we know this for sure. In non graphics performance mode up to four cards can be used, but in graphics performance mode you're limited to two cards at the moment, but Lucidlogix is working on allowing for up to four cards.

We also found out that the interface on the ELSA solution that connects to a workstation or server is using an external PCI Express x1 connection, although as this is only used for transferring data to be computed and already computed data between the CPU and the Tesla cards, this is meant to be more than enough for the intended usage scenarios. We're not sure how this works with other solutions, such as using Quadro cards.

Another interesting aspect is that its possible to cascade multiple Hydra chips, although at the moment this is limited to two chips, but this would allow for more cards to be used. This could at least in theory allow for four GPU graphics cards, but Nir suggested that in the future up to 8 GPU's per Hydra chip could be possible, in theory at least.

For those that are worried that this will be another nForce 200 chip, there's no need to worry as Nir assured us that the Hydra has a TDP of a mere 4W and as such it shouldn't even need a heatsink, although we expect motherboard makers to add one just in case.

The reason why Hydra scales so well is because its the Hydra chip that controls the rendering and Lucidlogix is doing it quite different than the way AMD and Nvidia are doing things. Using multiple graphics cards have so far meant alternate frame rendering, split screen rendering or tile based rendering. Lucidlogix on the other hand is going for object based rendering which is an entirely new approach which means that each GPU get allocated parts of an object of the scene that's being rendered and then the Hydra chip pieces together the small parts into the final frame which is then displayed on the screen.

The proof is still in the pudding so to say, as Lucidlogix has yet to show a running demo of its Hydra chip, but with a bit of luck this should happen during CeBIT which kicks off early next month. So far we're only hearing whispers about potential consumer partners, but again, there's a chance that more details will be revealed during CeBIT.

Source

[XSAlliN]

If CUDA is the reason, then it will work same as FASTRA.

[Mr Roboto]

I want to believe the Lucid is real and can perform similar to what they claim. However, this is something that's revolutionary as far as the gamer\enthusiast goes yet we barely hear anything about it not to mention a working demo. Why aren't they shouting from the rooftops showing off this technology even if they have yet to make any deal with motherboard manufacturers? If it can perform half as good as what they claim they won't have to worry about selling it, motherboard manufacturers will be lining up to be the first to feature the Lucid Hydra. They need to be hyping this up and showing some benchmarks to back up their claims.

If they did this any motherboard that's released with this chip would be back ordered for months! Plus they have Intel capital behind them. It just doesn't make a whole lot of sense to me. I patiently await some benchmarks, until then label me a skeptic.

[dirtypop]

Can anyone explain in basic terms why object-based rendering should be faster than every-other frame rendering for multiple cards? It doesn't make sense to me.

[clonez]

i dont think its faster most of the time (>2 gpus should be managed better, however)

its simply a very good way to split workloads on gpus with different perfomance and vendors (and microstuttering shouldnt be a big issue, if its done right and hydra splits the workloads fast enough, no matter how many objects are displayed)

[Aberration]

Can anyone explain in basic terms why object-based rendering should be faster than every-other frame rendering for multiple cards? It doesn't make sense to me.

Because this allows asymmetric cards with regard to ability.

If you had a 280 and a 9600 and did AFR you can obviously see there would be issues.

But with object based the 9600 could a few objects in the frame while the 280 took the bigger objects.

Then you take into account differing RAM per card and what each one would be able to handle.

[trinibwoy]

Object based rendering huh? I'd love to see that work with rendering something even as simple as a cubemap or with Carmack's proposed voxel based terrain renderer. This also renders a Z-only passes a moot point since that requires running all objects through the transform and rasterization stages. Hogwash!

If it was easy to just parcel out individual objects to be rendered to different GPUs AMD and Nvidia would have done it ages ago. How can Lucid do this so easily when they have a lot less information than the graphics guys do about their own hardware?

In terms of RAM, there's no win there. How would Hydra know which objects are more expensive (like which ones are going to be tessellated, or which have the most demanding shaders) or which textures need to go to which card? Some algorithms also rely on maintaining the drawing order of primitives. They're making it sound like you can just split up the objects that belong to the scene and delegate them to different cards without worrying about any dependencies. Sounds like a pile of bull until they offer up some proof that it works.

[geo]

^^ maybe thats why there isn't public demo yet?.. one would have thought, ATI with their X2 cards woul have come up with this tech

[RaZz!]

i've read about their plans to use object based rendering earlier, but i still can't imagine how that works. especially if different cards are used where one card is considerably faster than the other. another thing is: what if one card is heavily overclocked and would produce artifacts on a regular frame based rendering - what would be the effect on object based rendering?

it's definately a nice approach though, but i think if lucid really succeeds with this technique, ati and nvidia will do the same. it's just that ati and nvidia don't feel the need for something like that as the customers are paying for their crap sli and crossfire anyways. so why bother doing so when it's not "necessary" for them? but that's just my pov.

[EniGmA1987]

Why would anyone run a card for gaming or whatever that produces artifacts? That is when you turn your OC down...

[trinibwoy]

^^ maybe thats why there isn't public demo yet?.. one would have thought, ATI with their X2 cards woul have come up with this tech

Yep, I'm still trying to understand why people are falling for Lucid's so-far-empty promises and a few slides. SLI and CF aren't perfect but there are very good reasons for that. And those reasons don't go away when somebody puts some pretty pictures up in powerpoint.

[RaZz!]

Why would anyone run a card for gaming or whatever that produces artifacts? That is when you turn your OC down...

it's not about whether i do game like that or not (however, i never mentioned gaming anyways ), it was more of food for thought on what WOULD happen.

[BababooeyHTJ]

Object based rendering, sounds like a can of worms driverwise. I really cant see a company without a working product being able to write solid drivers when Nvidia and ATI which are obviously well established and very weathy have driver issues of their own.

Pics/benchmarks or it didn't happen.

[ToTTenTranz]

Object based rendering huh? I'd love to see that work with rendering something even as simple as a cubemap or with Carmack's proposed voxel based terrain renderer. This also renders a Z-only passes a moot point since that requires running all objects through the transform and rasterization stages. Hogwash!

Indeed, there was a time in the videogame industry to which we could call the "Carmack Age". That age is long gone, now. Carmack is more interested in rockets now, so let him be.
Between something Carmack proposed and something else that would allow for near-perfect multi-GPU scaling, guess which one I'd choose?

If it was easy to just parcel out individual objects to be rendered to different GPUs AMD and Nvidia would have done it ages ago. How can Lucid do this so easily when they have a lot less information than the graphics guys do about their own hardware?

And who said it was easy? Or are you assuming it must be easy because a new company came up with the concept? Assuming that all the great minds must be already in Intel, nVidia and AMD? I find that hard to believe.

In terms of RAM, there's no win there. How would Hydra know which objects are more expensive (like which ones are going to be tessellated, or which have the most demanding shaders) or which textures need to go to which card?
Some algorithms also rely on maintaining the drawing order of primitives. They're making it sound like you can just split up the objects that belong to the scene and delegate them to different cards without worrying about any dependencies.

Wait and see. Maybe they can do it, maybe not.

Sounds like a pile of bull until they offer up some proof that it works.

Do you think Intel would invest in something that "sounds like a pile of bull"?



EDIT:
To all those "superskeptics" around, here's an image from a 6 month old article:



It is object based. There are even videos showing it.

[Xenogias]

I imagine this will have a lot of caveats and limitations when it finally emerges. Like every other revolutionary, "why didn't I think of that before?", tech.

[trinibwoy]

Between something Carmack proposed and something else that would allow for near-perfect multi-GPU scaling, guess which one I'd choose?

Carmack's proposal is rooted in existing technologies that work and are based on reality (The terrain design process in Crysis is voxel based along with many medical imaging applications). You would take vaporware over the musings of one of the 3D graphics world's most prominent pioneers?

And who said it was easy? Or are you assuming it must be easy because a new company came up with the concept? Assuming that all the great minds must be already in Intel, nVidia and AMD? I find that hard to believe.

It doesn't have anything to do with being smarter than Intel, nVidia and AMD. It's about understanding how graphics rendering works and why Lucid's claims are ludicrous on the surface. All of the fancy algorithms they claim to have will be running on the CPU. The Hydra chip is nothing but a fancy PCIe bridge like N200.

Do you think Intel would invest in something that "sounds like a pile of bull"?

Who knows what Intel's motivation is here. Hydra supports good old SFR and AFR as well. They could simply want a solution that isn't tied to any IHV's chipset or technology and free themselves of a dependency there.

To all those "superskeptics" around, here's an image from a 6 month old article

Yes, we've all seen the PCPer article. So a missing piece of the floor in UT is proof now? They need to do a lot better than that to validate their claims. The reason we're skeptical is because we have some basic understanding of how this stuff works.

[Solus Corvus]

So a missing piece of the floor in UT is proof now? They need to do a lot better than that to validate their claims.

It's a lot more then just a missing piece of the floor. Most of the objects missing in one scene are present in the other. Some of the objects are rendered in both scenes (the gun, for example), probably as a way to avoid the dependency conflicts you mentioned earlier.

[Xel'Naga]

I can't believe people have doubts about Lucid Hydra.
Do you realize who is behind this? George T. Haber who built the UltraSPARC FPU, Moshe Steiner the leader of the team that built the first MMX unit for Intel. And so on...
The people behind Lucid Hydra are the dream team of digital circuits, people who could get a seven figure salary at any company they wanted to and they all put there money and next years of their life in Lucid.

[jmke]

seeing is believing as with most things in this industry

[NapalmV5]

how about a hydra for raid cards: areca arc1231 sli/tri/quad

[russian boy]

IMO too much hype.

I can't believe people have doubts about Lucid Hydra.
Do you realize who is behind this? George T. Haber who built the UltraSPARC FPU, Moshe Steiner the leader of the team that built the first MMX unit for Intel. And so on...
The people behind Lucid Hydra are the dream team of digital circuits, people who could get a seven figure salary at any company they wanted to and they all put there money and next years of their life in Lucid.

Are we seeing birth of first ever Lucid fanboy?

[trinibwoy]

It's a lot more then just a missing piece of the floor. Most of the objects missing in one scene are present in the other. Some of the objects are rendered in both scenes (the gun, for example), probably as a way to avoid the dependency conflicts you mentioned earlier.

Yeah there are a lot of fallbacks including synchronization points, reverting to SFR or AFR etc. But that's exactly the point, all of these contingencies lower performance far below the holy grail of linear scaling.

AFR is the simplest approach and the easiest to get close to linear speedups. However, we've seen how hard it's been for Nvidia and AMD to consistently get near that with most applications. For AFR to work perfectly there has to be no inter-frame dependencies or synchronization points. Both will go up dramatically with a so-called object based approach.

Given that, we are now to believe that an intra-frame approach with vastly greater and more complex dependencies and synchronization points will achieve even better scaling than AFR? Even if you have no knowledge of graphics rendering your logical conclusion should be obvious.

The point is that Lucid will only scale linearly in applications with very simple rendering engines. But then AFR will too so what's the point? People seem to think Lucid is offering some magic bullet that removes the problems plaguing existing methods. But all it does is make all the problems even harder to solve.

[grimREEFER]

breakthrough for supercomputing, but i predict nothing for the gamers will come out of this unfortunately

[Manicdan]

this is just the next step in the design of SLI/Xfire. to obtain every bit of performance out of the cards, more complex designs are needed. i however am ashamed that neither AMD or NVidia came up with this first. an X2 card would be perfect to squeeze in a high bandwidth chip thats built to run off one set of drivers and do it right for their own brand. i do remember rumors of the 4870x2 using shared memory, but even they backed down for the same old simple solution.

im still using just one card in my pc, cause i know that when i plan to upgrade to a new one, the old one will not be compatible. i still feel very strongly that both companies really need to look into a universal solution (for their own chips) to keep people upgrading.

also isnt it possible that Lucid can offer better than linear performance? if im bottlenecked when it comes to running 2560x1600 4xAA, then wouldnt i see a huge increase in fps (like more than double, but probably less than triple) if only 60% of the RAM was needed per card? i know my example may not make much sense, but if a card is trying to pack too much into its ram, then having to work with a little more than half the textures may really help it more than AFR ever could?

[LordEC911]

this is just the next step in the design of SLI/Xfire. to obtain every bit of performance out of the cards, more complex designs are needed. i however am ashamed that neither AMD or NVidia came up with this first. an X2 card would be perfect to squeeze in a high bandwidth chip thats built to run off one set of drivers and do it right for their own brand. i do remember rumors of the 4870x2 using shared memory, but even they backed down for the same old simple solution.

Who is to say this is the right step that was needed to take in the industry... As mentioned many times already, this has yet to be proven/shown.

also isnt it possible that Lucid can offer better than linear performance? if im bottlenecked when it comes to running 2560x1600 4xAA, then wouldnt i see a huge increase in fps (like more than double, but probably less than triple) if only 60% of the RAM was needed per card? i know my example may not make much sense, but if a card is trying to pack too much into its ram, then having to work with a little more than half the textures may really help it more than AFR ever could?

Better than linear is possible, even taking into account margin of error, though only in very rare situations.

breakthrough for supercomputing, but i predict nothing for the gamers will come out of this unfortunately

Why?
Supercomputing doesn't really need this, it already scales linearly in most situations.

It doesn't have anything to do with being smarter than Intel, nVidia and AMD. It's about understanding how graphics rendering works and why Lucid's claims are ludicrous on the surface. All of the fancy algorithms they claim to have will be running on the CPU. The Hydra chip is nothing but a fancy PCIe bridge like N200.

Hmmm... I thought there was a basic processor, RISC, built into their chips that did all the load balancing?