Anandtech: Lucid Hydra 100 - Enabling SLI/CrossFire on Any Platform

[Epsilon84]

http://anandtech.com/tradeshows/showdoc.aspx?i=3379

Wow, a new type of multi-GPU interface to make CF/SLI obsolete?

Basically it uses a more efficient way of load balancing between GPUs which enables it to achieve near linear scaling with additional GPUs, or so they claim.

Lucid, with their Hydra Engine and the Hydra 100 chip, are going in a different direction. With a background in large data set analysis, these guy are capable of intercepting the DirectX or OpenGL command stream before it hits the GPU, analyzing the data, and dividing up scene at an object level. Rather than rendering alternating frames, or screens split on the horizontal, this part is capable of load balancing things like groups of triangles that are associated with a single group of textures and sending these tasks to whatever GPU it makes the most sense to render on. The scene is composited after all GPUs finish rendering their parts and send the data back to the Lucid chip.

Thoughts? On paper, at least, it looks very promising!

EDIT - New article @ Anandtech, a more detailed look at Hydra 100: http://anandtech.com/video/showdoc.aspx?i=3385

[turtletrax]

Bring it on! I am sick of NV holding SLI ransom so they can peddel thier crappy wares. I wont even go SLI on X58 if there is a NV brigde chip on it. If the PCI-E is going through a NV chip I wont buy it at all.

[SimBy]

http://www.xtremesystems.org/Forums/...d.php?t=196389

[naokaji]

1: I'll believe when its avalaible in retail.
2: I'll believe when its avalaible in retail.


seriously, its one of those things that sounds too good to be true.

[systemviper]

I love it, it would make the motherboards so much better, (fingers crossed)

[Epsilon84]

http://www.xtremesystems.org/Forums/...d.php?t=196389

Damn, I must have missed that thread.

Well, more discussion is welcome in this thread, if the mods allow.

[Sampsa]

I'll go see it live today, lets see..

[shiznit93]

will this thing microstutter lol?

[skindogg]

As sweet as this sounds, we're never going to see it, due to two possible scenarios:
A: performance is much less than the respective SLI/CF chipset
B: it is the same or better and NVIDIA/ATI sues them into the ground.

[Blauhung]

I believe this may be Intel's way of enabling dual GPU solutions for Larabee, but this is my own personal opinion and not coming from anything I've seen internally.

[perkam]

I don't think people understand the magnitude of this chip. If Intel is able to harness this technology, it will not need to buy the Nvidia chip to allow SLI on its motherboards.

Ideally, the universality of the project means Larrabee, Rx00 and GTx00 will be supported.

Perkam

[Eldonko]

Whoever thought of this is going to be very rich if it works.

[JohnZS]

Whoever thought of this is going to be very rich if it works.

Or get sued...and then purchased by...nVidia!
or Intel

[rozzyroz]

i would imagine that it will be blocked through drivers much in the same way that nvidia blocks sli... if there is no nvidia sli chip, then disable vidiocard 1...

i hope there is no way to block this new chip, and we can use what ever we want on any motherboard... wonder if you could multi gpu different manufacturers?..... one nvidia and one larabee.. and maybe a third ati card to "mix" things up a bit..

[vitaminc]

Seen it live at IDF exhibit last night.

Basically the current SLi method is like a 2-4 stage pipeline with very frequent stalling and flushing overhead, thus the scalability is far from linear. Hydra does it by distributing tasks per frame to different GPUs instead of different frames to different GPUs. How they divide and arbitrate the tasks are, of course, proprietary.

The chip itself sits between the NB and GPU, taking one PCI-express 16 lane.

One of my biggest concerns, technically, is choking of the PCI express bandwidth for those ultra powerful GPUs. The chip takes in 16 lanes and divide up the task to either 2 GPUs x 8 lanes each or 4 GPUs x 4 lanes each if I remembered correctly.

The other disappointment is that the chip is designed to distribute graphical tasks using better algorithms so its application to GPU data crunching (CUDA, Larabee, etc) is limited.

The chip runs at 5W only, so that's a big positive compare to those incompetent Nvidia dumb overclocking phucks. Wasn't hard to see Nvidia imploding when they went down the path of P4...

All and all, they sampling to bunch of mobo venders and reviewers right now, product expected to be on market in 1H 2009.

[Tony]

I believe this may be Intel's way of enabling dual GPU solutions for Larabee, but this is my own personal opinion and not coming from anything I've seen internally.

wouldn't that just add latency when you don't need it if they opened up the drivers ?

Intel won't need a "bridging chip" for its own GPU solution ...will they?

[Helmore]

I believe this may be Intel's way of enabling dual GPU solutions for Larabee, but this is my own personal opinion and not coming from anything I've seen internally.

Larrabee is already a lot better suited to multi-GPU than ATI's or NVIDIA's cards are and that's mainly because of the way it renders. It still does need a interconnect between the cards, but nothing too outlandish and a board that has an X58 chipset on it should to the trick. That's just my guess from what I've read though and I'm no expert on these things.

[ToTTenTranz]

One of my biggest concerns, technically, is choking of the PCI express bandwidth for those ultra powerful GPUs. The chip takes in 16 lanes and divide up the task to either 2 GPUs x 8 lanes each or 4 GPUs x 4 lanes each if I remembered correctly.

PCI-Express in crossfire/sli is also used for data exchange between GPUs. With this architecture you don't need that because apparently the work between GPUs is distributed beforehand.
Since the chip is coming embedded on motherboards (and high-end ones of course), it will only use PCI-express 2.0+, and that's 16GB/s bandwidth. There's no way a 3d game would use more than 16GB/s bandwidth for the graphics card, not in the next few years. That would be like transferring the whole game to the graphics cards every second... if your hard drive / optical media ever allowed it.

The other disappointment is that the chip is designed to distribute graphical tasks using better algorithms so its application to GPU data crunching (CUDA, Larabee, etc) is limited.

Well they do market the chip for gaming and professional rendering.. I don't really see the point in using it for data crunching.




And how about merging the two threads about Lucid?

[makaka]

1: I'll believe when its avalaible in retail.
2: I'll believe when its avalaible in retail.


seriously, its one of those things that sounds too good to be true.

1: +1
2: +1
Completely right , too good to be true

[Xoulz]

Yeah, some of us were talking about this a month ago. I'm with Perkam in my thinking. There was some speculation whether or not Intel was going to let Nvidia make chipsets for the new platform. I think Intel did this on purpose so later when Nvidia loses it's SLI chipsets to open source SLI that this Lucid chip will provide. They can't be seen as exclusionists or being monopolistic.

This is exactly what will eliminate micro stutter or any multi GPU inherent problem.

[Epsilon84]

Or get sued...and then purchased by...nVidia!
or Intel

Get sued for what?

Besides Lucid is apparently largely funded by Intel, so if they get sued they're practically suing Intel.

[Xoulz]

Or get sued...and then purchased by...nVidia!
or Intel

Please understand what your talking about before you make such *lacking in knowledge* statements. Lucid hold 50 patents on this Hydra chip....!



.

[BlueAqua]

Sounds amazing and would definitely be worthy of an upgrade. I wonder with Larabee and this you could really put a lot of gpu's in one machine.

[Sr7]

I give this about a -.5% chance of actually working, at least with ATI/NV GPUs. If anyone knows their products, it's the GPU vendors. If they can't get more than the scaling they get now, there's no way someone who is abstracted from the hardware/drivers by one level will be able to do any better. Not to mention there are always particulars in each vendor's implementation which can't just seamlessly be made to work together. Then what happens in games where there are elements from one frame read to the next? The same thing that happens now.. no scaling.

The only hacky method they could do would add a ton of latency to the process, and no one wants to sacrifice responsiveness for a higher framerate. This also sounds like a recipe for microstuttering.. mismatched cards with different amounts of processing power, spanning multiple vendors and generations? 1 frame rendered take 5ms while the next takes 25ms on an older card?

[fireice2]

Hope this one really works.