PDA

View Full Version : Mental Images - NEWS announcement


pete016
03-08-2006, 01:51 PM
www.mentalimages.com (http://www.mentalimages.com/)

www.mentalimages.com/1_1_news/news_texte/060306.html (http://www.mentalimages.com/1_1_news/news_texte/060306.html)

also
www.artvps.com (http://www.artvps.com/)

Lorecanth
03-08-2006, 11:39 PM
This could be big... hardware accelerated mental ray...

CupOWonton
03-08-2006, 11:56 PM
This isnt realy big news. These cards become outdated within a year. Litteraly. with the CPU prower increasing steadily every year, paying $13000 for a render box would be a waste of money when you could buy 13+ $1000 computers that only use base components for net rendering but high end processors and ram.

From The original stats that was given out with the PURE render cards and the Render Drives, they seemed to be lying about their test results, especialy when involving Vray. We attempted a similar benchmark test on the ChaosGroup forums, and found that the only way it would take Vray to render SO LONG is if they had no clue how to actualy use it and possibly use LightTracer rather than actualy using Vray's actual GI system.

Render cards are not the wave of the future if theyre going to be propriotary to Mental Images software. That would greatly limit what people would be using. If they had card that would accept high end processors stritly for rendering purposes, that would be much better. Takes all the other processing off the motherboard.

I believe the next step in calculating alternate information using a card is a physics card for calculating object interactions on a large scale and updating that with the CPU and GPU. We've already gone back to linking cards like we use to with 3d accelerators. So its possible to see a set of linked cards each doing a different task before we can compress it all back onto 1 card again. Assuming we would even go back to doing it all on 1 card. It may close the space the information has to travel, But you only need so much information at a time.

Lorecanth
03-09-2006, 12:43 AM
I disagree...that figure of 13,000 is completely ignoring the cost of licensing mental ray for those 13 machines. Which if dual processor are going to be licensed at $2,000 a pop. So your cost for those machines just went to $39,000. Vray is a great rendering app, but a lot of studios have a signifigant investment in mental ray in both knowledge and licenses. So at the least this is rather interesting.

CupOWonton
03-09-2006, 03:03 AM
I disagree...that figure of 13,000 is completely ignoring the cost of licensing mental ray for those 13 machines. Which if dual processor are going to be licensed at $2,000 a pop. So your cost for those machines just went to $39,000. Vray is a great rendering app, but a lot of studios have a signifigant investment in mental ray in both knowledge and licenses. So at the least this is rather interesting.

Didnt MentalImages change the way they lisence MentalRay reciently? Maybe with just how they work it with say Max or Maya *unlimited render nodes* and you dont have to buy a lisence for every computer, just 1, and install it on the rest legaly wile only using 1 lisence for unlimited render nodes. And if its on a card, its on a card, I doubt they'd ask for more money selling you a card with no software.

$13000 is what a render drive costed. + 1 computer + 1 lisence of Maya/Max/MentalRay etc *est $6000?. All that for 1 render drive.
You can outfit computers with 3dsmax using 1 registered lisence, and vray or mental ray because both of them have an unlimited render node lisence *asside from the 10 limit on Distributed renders for Vray* So you only have to pay for 1 version of max in that instance. I doubt someone is going to go the other route and attempt to buy an un nessisary work lisence for every computer.

BillSpradlin
03-09-2006, 05:02 AM
CPU's become outdated within a year yet you don't see everyone upgrading their 500 proc renderfarms to the latest and greatest processors every year do you? Obviously not.

This stuff is really going to lean towards being useful to those in the architectural and automotive design fields more so than it will be for feature film visual effects. I can't see this being a bad thing overall though and mental images seems to be on the right track in what they are developing. Especially since the other big renderer out there for film visual effects probably won't be around publicly a whole heckuva lot longer. Expansion to other fields is a great way of catering to a wider range of customers and clients which means more money and more development.

CupOWonton
03-09-2006, 08:07 AM
CPU's get outdated fast, sure, but its easier to just add a new computer to the current farm than replace a $13000 piece of equipment that can only be used one way. Render Farms/blades are better.

I'm in the architectural 3d rendering field, and all I ever hear are complaints about ArtVPS's products. Usualy the comment goes something like this - " We used it, then we got vray, we never touched the stupid machine again after that."

I also noticed their benchmarks lack any comparason to that of MentalRay or Vray as they had before. Unless theyve been placed far far away from the.. 2 benchmarks they have now, which only compare their products to themselves, not to competition. I wonder why?

Lorecanth
03-09-2006, 08:09 AM
Didnt MentalImages change the way they lisence MentalRay reciently? Maybe with just how they work it with say Max or Maya *unlimited render nodes* and you dont have to buy a lisence for every computer, just 1, and install it on the rest legaly wile only using 1 lisence for unlimited render nodes. And if its on a card, its on a card, I doubt they'd ask for more money selling you a card with no software.

$13000 is what a render drive costed. + 1 computer + 1 lisence of Maya/Max/MentalRay etc *est $6000?. All that for 1 render drive.
You can outfit computers with 3dsmax using 1 registered lisence, and vray or mental ray because both of them have an unlimited render node lisence *asside from the 10 limit on Distributed renders for Vray* So you only have to pay for 1 version of max in that instance. I doubt someone is going to go the other route and attempt to buy an un nessisary work lisence for every computer.

I was refering to mental ray standalone... and sadly no they havn't changed their rendering policies on that.

CupOWonton
03-09-2006, 08:21 AM
I was refering to mental ray standalone... and sadly no they havn't changed their rendering policies on that.
Thats a rather stupid way of marketing a product when they allow for unlimited nodes through Max. :shrug:

mustique
03-09-2006, 02:02 PM
Nvidia's Gelato renderer coupled with quadros, might have looked like a threat for ARTVPS. Maybe also for Mental images. So they get together. Now I'm really curious.

Will we see "realtime raytracing" sooner than imagined? Will even gamers buy Quadro's some day? Will we think of buying a umm- RTR card before we even think of an PPU?

Or will it be that we'll embrace our good old CPU ?..

rasamaya
03-09-2006, 04:21 PM
I cant wait!!This will be like waiting for Zbrush 2.5, TG2 or Modo....

enygma
03-09-2006, 04:28 PM
Real-Time raytracing has been achieved, but not in a high resolution form. Some guys over at the Saarland university in Germany presented a paper at Siggraph last year called RPU: Ray Processing Unit, where they used a standard run of the mill Xilinx Virtex II FPGA. The chip used a variance of OpenRT rewritten in VHDL to work on the chip and managed to provide real-time raytracing capabilities at lower resolutions (like 640x480).

Mercury Computers also demonstrated OpenRT running on a 10 CPU cluster of Opterons in real time. They also demonstrated it running on Cell as well.

I know for a fact that there are efforts to try and make higher resolution and huge datasets be raytraced in real-time that don't involve going with faster CPUs or GPUs.

albedo4800hp
03-09-2006, 04:53 PM
Just some random thoughts why CPU might not be the right way to go in the first place or why even waiting for more CPU power did not really fullfill our dreams of fast or realtime raytracing or at least a rather interesting statement, well it is just a guess so take it with a grain of salt. Quoted from the last Raytracing news
( see http://www.acm.org/tog/resources/RTNews/html/rtnv18n1.html)

[starting quote]

Gordon Stoll replied:

Ah, this is something I've thought about a lot. Fifteen years ago, 1990, would be (I think) a 33MHz 80486. That had onboard floating-point, which is obviously important for RT. The actual semiconductor scaling trend (including both the number of transistors and their switching speed) predicts an improvement of 3,125 times since then. Do you think that the core CPU performance of RT has actually gone up that much?

My guess is that it didn't, and that instead RT fell off of the performance curve right around then because RT breaks pretty much everything that we've put into CPUs since then (deep pipelines with branch prediction, superscalar, out-of-order, SIMD/vector operations, etc, etc, etc). I've been thinking about trying an experiment or doing a literature survey to try to support this, but for now it's just a crackpot theory. Maybe I'll try to convince Eric to figure it out. :)"

[ending quote]

I don't see why he thinks SIMD/vector operations fall out of this, that IMHO is not true

shehbahn
03-09-2006, 05:52 PM
>I don't see why he thinks SIMD/vector operations fall out of this,

because by definition rays cannot guarantee homogeneous processing of data (and why shader code is processed "vertically") - unlike REYES that deals with homogeneous grids of micropolygons (where shader code is vectorised and processed "horizontally").

in other words, most of the predictive branching and other caching Intel has put into their CPUs are constantly defeated by the randomness of both code and data accesses incurred in a tracer, while processing interpreted shader code one instruction at a time over large grids of upolys amounts to running small loops of code that makes very homogeneous data access.

albedo4800hp
03-09-2006, 06:44 PM
>I don't see why he thinks SIMD/vector operations fall out of this,

because by definition rays cannot guarantee homogeneous processing of data (and why shader code is processed "vertically") - unlike REYES that deals with homogeneous grids of



That is true for the general case but surprisingly it is not so bad as it seems at least depending on how you organize your algorithm. For primary rays that is not true at all and even for secondary rays especially shadow rays neither. It becomes worse for reflection and refraction rays but only for highly curved objects. And besides most instructions you would probably need for shading a ray are vector instructions as well there is no huge difference to REYES or standard scanline rendering. Access pattern to memory are probably worse I agree. So from all features put in the CPU I think that this is the one were raytracing has benefited most from.

shehbahn
03-09-2006, 07:39 PM
>And besides most instructions you would probably need for shading a ray are vector instructions as well there is no huge difference to REYES or standard scanline rendering.

you are confusing a SIMD vector with a math vector i think - and yes there is a huge difference : most production shaders range in the 100 000 instructions. if you run this vertically, every ray hit sample will run all these instructions, then proceed to the next sample. a REYES shader will execute each instruction of the shader on an entire grid of contiguous upolys, then proceed to the next instruction. there are many benefits to this :
- a single shading instruction can fit in L1 cache and runs in a very tight loop (benefitting from almost every advantage of the superscalar pipelines)
- no predictive branching problems
- texture access guaranteed to be contiguous for the grid
- easy derivatives, which means shader space filtered patterns (ie. far less sampling for anti-aliasing)
of course this comes at the (huge) price of global illumination...

>Access pattern to memory are probably worse I agree. So from all features put in the CPU I think that this is the one were raytracing has benefited most from.

the very argument that he is making is because traced shaders have lots of branching code and not very homegeneous processes, they do not benefit from these features nearly as much, to the point where the nature of the code is in fact running against the branch predicting and L1 caching.

pete016
03-13-2006, 07:52 PM
ARTvps have up dated their website

jbradley
03-14-2006, 03:37 PM
Didnt MentalImages change the way they lisence MentalRay reciently? Maybe with just how they work it with say Max or Maya *unlimited render nodes* and you dont have to buy a lisence for every computer, just 1, and install it on the rest legaly wile only using 1 lisence for unlimited render nodes. And if its on a card, its on a card, I doubt they'd ask for more money selling you a card with no software.

Not for Maya. Only Sattelite rendering is supported, and only up to 8 CPUs, not per machine. Max is the ony renderer that supports unlimited render nodes with mental ray, and only through the Max renderer, not stand alone MR.

This is one of my major beefs now that Autodesk acquired Alias.

CupOWonton
03-14-2006, 07:46 PM
Not for Maya. Only Sattelite rendering is supported, and only up to 8 CPUs, not per machine. Max is the ony renderer that supports unlimited render nodes with mental ray, and only through the Max renderer, not stand alone MR.

This is one of my major beefs now that Autodesk acquired Alias.

But that deal was struck before the aquesiton of Alias wasnt it? Like, back on max 7. Maybe theyll work another deal with Mental Images to get unlimited render nodes for maya now that they have it too.

pixelmonk
03-15-2006, 01:55 PM
But that deal was struck before the aquesiton of Alias wasnt it? Like, back on max 7. Maybe theyll work another deal with Mental Images to get unlimited render nodes for maya now that they have it too.

With the onslaught of renderers for Max, it probably wasn't a big deal if MI and Autodesk did a deal like that. It's not like Max people were buying lots of extra MR licenses. Hell.. if I used Max, I'd probably try Vray instead. Also, Audodesk probably paid them some amount of money for unlimited licensing to happen, oh and of course to help further development for MR for Max. I could see it as a win-win for both sides. I do hope Autodesk does this for Maya too. Unlimited satellites would be killer.

CGTalk Moderation
03-15-2006, 01:55 PM
This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.