THE Hydra Demo

[VirtualLarry]

Originally posted by: Janooo

Hydra appears to the OS as a single GPU.

The distribution engine as it is called is responsible for reading the information passed from the game or application to DirectX before it gets to the NVIDIA or AMD drivers. There the engine breaks up the various blocks of information into "tasks" - a task is a specific job that HYDRA defines that can be passed to any of the 2-4 GPUs in the system.

It appears Hydra will be breaking DirectX calls into smaller pieces and managing their multitasking.
It remains to be seen if the scaling is going to be better.
This is no AFR so there shouldn't be any micro stutter.

The question I have is - if this is indeed better quality or scalability than existing SLI or CF implementations, then why? If it's better algorithms, then it would seem cheaper to me (no added hardware needed), if those algorithms were implemented in the drivers.

I think it remains to be seen what the added "magic" hardware does. Is it simply yet another PCI-E bridge chip, like the NV200, or does it have extra bells and whistles?

 

[VirtualLarry]

Originally posted by: nitromullet
Also, Hydra claims to split up the workload 50/50 between two gpus, whereas with AFR each gpu renders every second frame and because of this each gpu duplicates a lot of the same work as the other.

I'm also not sure if there will be a 'timing requirement' per se... Since all the gpus are working on all the frames, you don't have a situation where one frame will be ready 'too early' the way as you would with AFR in SLI/CF.

If all the GPUs are working on all of the frames (certain GPUs responsible for certain polys?), then what happens to AA implementations? Will they look "off", if part of the scene is composited from, say, an NV card, and part from an ATI card, and if they use slightly different AA implementations?

Then again, because of the Vista single-instance WDDM driver restriction, perhaps cross-vendor multi-GPU won't even be possible, and thus not even an issue.

 

[VirtualLarry]

Originally posted by: dunno99
But if you've ever did any GPGPU programming, you'd know that reading data off of the graphics card moves more along the lines of x1 speeds. The reason being that the communication of these chips are optimized for receiving data, not sending data. So that means the two cards are communicating at much lower speeds, which isn't even close to enough bandwidth for a single buffer synchronization in a 60fps game (check the wikipedia entry on SLI if you don't believe me). Then again, who knows...maybe they're playing in 640x480.

I disagree. Are you claiming that all current video card PCI-E implementations are severely broken? This isn't AGP, a one-way street. PCI-E has 16 lanes of bi-directional bandwidth. There shouldn't be any problems sharing frame buffers across cards, over the PCI-E bus.

 

[SSChevy2001]

Originally posted by: VirtualLarry
The question I have is - if this is indeed better quality or scalability than existing SLI or CF implementations, then why? If it's better algorithms, then it would seem cheaper to me (no added hardware needed), if those algorithms were implemented in the drivers.

I think it remains to be seen what the added "magic" hardware does. Is it simply yet another PCI-E bridge chip, like the NV200, or does it have extra bells and whistles?

Cause with current SLI and CF setups the memory bandwidth and framebuffer are not shared. If you have 2x 4870 512MB in CF your only going to have 512MB framebuffer and 115GB/sec bandwidth. With the Hydra engine your 2x 4780 with have 1GB framebuffer and 230GB/sec bandwidth.

CF and SLI
DX code -> both GPUs via AFR ( which suffers micro-stutter and requires custom profiles to work. Also the memory is duplicated, wasting the extra memory )

Hydra Engine
DX code -> DX 50% GPU 1 and 50% GPU2 ( Each GPU has it's own memory to work on the task given to it. )

Software emulation of what hydra would more than likely not work. It would be like running physx game on your without a PPU or GPU Physx.

----

Biggest question is what happens to current multi-gpu single cards like the 4870x2 or 9800GX2? Also how wil Hyrda handle power consumption when not gaming?

 

[R3LIC]

Just got around to reading the reports and posts about this thing.
It seems quite promising if they can deliver. I have one question though that I haven't seen asked anywhere.

Would this chip be capable of being put on and add-in card by itself ?

 

[jkresh]

r3lic, so far it has been suggested as either on the motherboard or on a dual gpu card (4870x2 or something like it), from what I have seen of its specs I think it could be on its own pci express card (probably would want an x16) or better yet on something like a crossfire "mastercard" (ie have a gtx280 or 4870 or something with this chip on it, that allows for multigpu even if you don't have it on the motherboard) so you don't loose a pcie express slot.

 

[SSChevy2001]

Originally posted by: R3LIC
Just got around to reading the reports and posts about this thing.
It seems quite promising if they can deliver. I have one question though that I haven't seen asked anywhere.

Would this chip be capable of being put on and add-in card by itself ?

Here's what they said, looks like the answer is no.

There are two likely scenarios for potential designs: on a motherboard or on a graphics board.

Also current multigpu cards don't look to work with it.

 

[dunno99]

Originally posted by: VirtualLarry

Originally posted by: dunno99
But if you've ever did any GPGPU programming, you'd know that reading data off of the graphics card moves more along the lines of x1 speeds. The reason being that the communication of these chips are optimized for receiving data, not sending data. So that means the two cards are communicating at much lower speeds, which isn't even close to enough bandwidth for a single buffer synchronization in a 60fps game (check the wikipedia entry on SLI if you don't believe me). Then again, who knows...maybe they're playing in 640x480.

I disagree. Are you claiming that all current video card PCI-E implementations are severely broken? This isn't AGP, a one-way street. PCI-E has 16 lanes of bi-directional bandwidth. There shouldn't be any problems sharing frame buffers across cards, over the PCI-E bus.

Yeah, you're right. I should've looked further into the new stats on CUDA and some such. The bandwidth currently stands at about 2GB/s one can get off of the cards. Which means, for a 60fps game, that's about 33MB/frame. Which I still sincerely doubt is enough to synchronize between two cards running at 1600x1200 or so (which is slightly less than 5 full frame buffer syncs).

PCI-E broken? No. PCI-E is doing very well in what it was meant to do. What's broken is using PCI-E to synchronize between two cards. The point being that if communication isn't minimized between two cards, there's no chance that this will fly. The fact that Crytek uses lots of full screen passes (AO, color correction, etc...) means that the framebuffers need to be synchronized most of the time or one will notice visible artifacts along the seams where the memory is divided between the two cards.

 

[Nemesis 1]

Ya ask good questions , Seems to me part of the ans. are in the linked article

 

[BFG10K]

Intercepting API calls like that sounds iffy. I want to see this thing working properly in a range of games without application profiles before I'll consider it a legitimate solution.

 

[Nemesis 1]

Don't we all?? Heres to crossed fingers.:sun:

 

[taltamir]

i have seen another suggested hardware implementation, where the hydra chip replaces the plx chip on X2 cards, it will allow 2 to 4 GPUs per hydra chip to be used in a single slot, and it will not appear as SLI / CF to the drivers. Instead working as a single card, scaling 100%, with none of the multi GPU issues...

that looks pretty cool.

Despite intel funding, I don't think it really targets larabee, since larabee will inherently be an array of shrunk, very highly oced in order p1 variants. Rumored to be seperate 10 cores in the first model. Which will then use drivers to translate all the directX commands to highly parallel x86 code. This means there is no need for any sort of hydra like chip...
But if they could make it necessary, and then jack up prices... lots of money to be made.

 

[Nemesis 1]

Yes thats why I am so excited. This isn't about intel its about performance enhancement.

If this works it opens up a lot of possabilities. For other GPU companies to suddenly have good products. The thing I am concerned about is no roadmap for DX10. DX10.1 is suppopse to be ready befor 09. So thats of some concern . ATI is ok but Dx10. for now no support.

 

[Nemesis 1]

When this was announced . I did this thread . I thought it was a good thread for its Time.

I can tell you this for other forums that had same subject . This was by far the most civil. Something you men should take pride in .

Now I believe we can expand upon the recent Msi news with a good base from which to resume. The Tech chat!

 

[exar333]

I think it's pretty cool. Some actually new tech would breathe some life into the video forums! (not just a new GPU; same old story...)

It would be very cool to just buy the best card around for price and/or performance and just know you could throw it in with your old card, regardless of brand.

SLI sucked the big one (it was a PITS and a hassle) and maybe this will make it better.

Here's to hope!

 

[gigahertz20]

AMD or Nvidia will gobble up this company if their technology really works, imagine if Nvidia GPU's had linear performance scaling, while AMD was still stuck with their crossfire technology....everybody would buy Nvidia video cards then.

My other major concern is that this technology could end up like AGEIA's PhysX - great potential but gobbled up by one of the mega-players rather than turning into a product on its own. Honestly after hearing the entire presentation I was curious why NVIDIA or AMD hadn't already thought of this - the potential for being bought up is extremely high here.

 

[Nemesis 1]

Originally posted by: gigahertz20
AMD or Nvidia will gobble up this company if their technology really works, imagine if Nvidia GPU's had linear performance scaling, while AMD was still stuck with their crossfire technology....everybody would buy Nvidia video cards then.

My other major concern is that this technology could end up like AGEIA's PhysX - great potential but gobbled up by one of the mega-players rather than turning into a product on its own. Honestly after hearing the entire presentation I was curious why NVIDIA or AMD hadn't already thought of this - the potential for being bought up is extremely high here.

In the above qoute the presenter stated a fact that stiks out like a sore thumb.

If Hydra works as claimed it is a big winner as all hear would agree if it works as claimed.

What the presenter failed to see as did you is the other fact that sticks out like sore thump . If the real potencial is their its Worth Billions of dollars. I don't see AMD NV making billion dollar offers. After they read present licencies. And Provisions made.

 

[gigahertz20]

If Hydra works as claimed it is a big winner as all hear would agree if it works as claimed.

The Lucid Hydra engine must work decently if a big motherboard maker like MSI is already putting the tech on a motherboard. They wouldn't be installing it on a motherboard if the tech was a complete failure or had alot of issues. The link below did state it was a prototype motherboard, so who knows when we'll see benchmarks or even if MSI decides to keep it on there. I just hope to see benchmarks sometime within the next few months if it stays.

http://www.iopanel.net/forum/thread31404.html

http://www.techpowerup.com/101...ments_Lucid_Hydra.html

 

[aka1nas]

Originally posted by: gigahertz20

If Hydra works as claimed it is a big winner as all hear would agree if it works as claimed.

The Lucid Hydra engine must work decently if a big motherboard maker like MSI is already putting the tech on a motherboard. They wouldn't be installing it on a motherboard if the tech was a complete failure or had alot of issues. The link below did state it was a prototype motherboard, so who knows when we'll see benchmarks or even if MSI decides to keep it on there. I just hope to see benchmarks sometime within the next few months if it stays.

http://www.iopanel.net/forum/thread31404.html

http://www.techpowerup.com/101...ments_Lucid_Hydra.html

From what I read on XS, it's basically being used as a replacement for the NF200 chip and may only be doing PCI-E switch duties.

 

[Nemesis 1]

Yep . They make many claims. The biggest one bits my ass is no RT for at least 10 more years . Many claims being made . We at least are seeing a Card with it on . for upcoming unreleased processor. Which many claims are made about also . Most are young men shouting off mouths. The good guys don't say much there any more unless ya go deeper inside extreme with pay as you go . LOL. Easy hack.

 

[gigahertz20]

Originally posted by: Nemesis 1
Yep . They make many claims. The biggest one bits my ass is no RT for at least 10 more years . Many claims being made . We at least are seeing a Card with it on . for upcoming unreleased processor. Which many claims are made about also . Most are young men shouting off mouths. The good guys don't say much there any more unless ya go deeper inside extreme with pay as you go . LOL. Easy hack.

WTF? English is not your first language is it?