Physics Processing Unit

[DAPUNISHER]

Originally posted by: slash196
Sounds damn cool. Physics is becoming a bigger and bigger part of games, and a dedicated card could increase the quality dramatically. Of course, it would have to be a PCI card, so the actual bandwidth is limited, but I think its a great idea. I just hope is gets substantial development support.

Why would it have to be PCI? PCIe is are already starting to supplant it, and this tech may not require much bandwidth anyways.

 

[slash196]

It'll have to be PCI because the PCI-e slot is taken by the GPU. Kind of a no-brainer.

 

[Duvie]

Actually makes more sense!!!

Why not have hardware specific components...CAD users by things that wil help CAD users....AI and PPU for gamers on their uber expensive gaming cards.....sound specific processors for people who do professional sound mixing...

Why would INtel or AMD what to include everything liek this in a cpu when the fact is most users are surfing the net and recreationally playing games.....

Maybe a whole new line of Xeons can be created to be application specific , but the cost to develop and develop the chipsets probably prohibits it, or prohibits it based on the size of each individual market...I smell major premium on something like this...

 

[DAPUNISHER]

Originally posted by: slash196
It'll have to be PCI because the PCI-e slot is taken by the GPU. Kind of a no-brainer.

Oh, so there are no boards with multiple PCIe slots available with a variety of speeds from 1-16x right now :roll: No brainer, more like no brain? J/King of course but you are definitely having a brain fart to think the vid card is the only PCIe slot already available to be filled on some boards already out.

 

[Duvie]

HOw about PCI-x slots liek one puts a high bandwidth SATA Raid card in....100-133mhz slot etc.....

 

[DAPUNISHER]

Originally posted by: Duvie
HOw about PCI-x slots liek one puts a high bandwidth SATA Raid card in....100-133mhz slot etc.....

That is almost exclusively workstation&server board territory so far. PCIe is a pretty versatile interface that should accomodate a card with this chip on it, and is already becoming the de facto standard for a new gen high-end gamer's board. By the time cards based on this are out it will be even further entrenched.

 

[Duvie]

Originally posted by: DAPUNISHER

Originally posted by: Duvie
HOw about PCI-x slots liek one puts a high bandwidth SATA Raid card in....100-133mhz slot etc.....

That is almost exclusively workstation&server board territory so far. PCIe is a pretty versatile interface that should accomodate a card with this chip on it, and is already becoming the de facto standard for a new gen high-end gamer's board. By the time cards based on this are out it will be even further entrenched.

Yeah I agree but I was think for the professional users like CAD guys....

The abit mobo for dual opterons which looks more geared towards enthusiasts I believe has one...I think with games gearing up for SMP due to dual core I think professional and enthusiast can get blurred together...Can you imagine dual core dual opterons????


Probably the pci-e is easier to push as more manistream desktop...definitely...

 

[Murd0ck]

Yes I agree it's a good idea to read with caution. That specific quote was from the Ageia site and should definately be considered a forward looking statement.

Does anyone know who is using their physics engine? I read something about U3? and Ubisoft among some others.

I also noticed this - again from ageia's site "AGEIA's NovodeX SDK provides interfaces and plug-ins for all of the major tools vendors, including Alias, Discreet and Softimage"

"PhysX Processor Architecture has been designed to enable radical acceleration of:

Rigid body dynamics
Universal collision detection
Finite element analysis
Soft body dynamics
Fluid dynamics
Hair simulation
Clothing simulation "

another
"NovodeX Physics SDK

The NovodeX Physics SDK is a stable, high-performance solution for game developers to enable physics-based gameplay and effects in PC and console titles. A powerful API for the PhysX PPU, NovodeX enables game developers to inject both software-only and hardware-accelerated features into their games. The NovodeX Physics SDK is also the first and only asynchronous (multithreaded) physics API capable of unleashing the power of multiprocessor gaming systems.

Key features:

Stable, high-performance solver
First and only multiprocessing physics API
PC and console support
Works alongside other game engines
Supports vehicles, rag dolls, and character controllers
Integrates with any renderer
Full complement of code samples and tutorials
World-class developer support and custom solutions "

 

[Lonyo]

STOP PRESS:
BREAKING NEWS:
READ THIS NOW.

ARTICLE ON THIS TECHNOLOGY
It will answer many points raised. PLEASE READ BEFORE FURTHER COMMENTS. KTHX.

"Yes we're supporting it and its going to be cool."

Mark Rein
Epic Games Inc.

"To be honest I don?t think gamers will spend money for dedicated physics hardware (unless it?s included in something they already buy like their video card) so I?m not sure if this will ever catch on?
However, if dedicated physics hardware could be guaranteed the improvement to the physics realism in games would be incredible."

Kevin Stephens
Monolith Studios

"Hardware Acceleration always beats software ? therefore, time sliced CPU-based physics will impact performance and have to be dumbed down to satisfy bandwidth issues?.

GD: What are the current bus implementations you have on the road-map?
Curtis: We presently have both a 4X PCI Express and PCI supported on our chip as the interfaces.

 

[DAPUNISHER]

No arguments here, if there is a potential market, you can rest assured they will exploit it Hypertransport is well suited to whatever they want to connect, so any of the potential combos is possible with it.

Thanks Lonyo :beer:

 

[otispunkmeyer]

Originally posted by: DanDaMan315
I agree with MrControversial, this sounds like a flop. With these multicore CPU and GPU's coming out I see know need for a dedicated "Physics" chip. Atleast in the near future.

cpus can do a few hundred, it would take so many cpus to do what a single chip can do on its own.

dual core still wont come near it, and even then, the games will still have to be coded to make use of that second cpu

 

[DAPUNISHER]

LMAO! I talk about the borg computing model and they got the idea while p[ounding brews and watching star trek! That rules!

Looks like as always, it all depends on dev support, and a way to package it to sell. I predict they never make it to that IPO, but get bought out by a big player, unless their visonaries who will license but refuse to sell.

 

[otispunkmeyer]

Originally posted by: Cat

Originally posted by: MrControversial

Originally posted by: Cat
I've been using Novodex for about a year now. A single CPU cannot come close to what this chip is promising, so I don't think it's useless. It's very exciting, actually. I am concerned about reading back thousands of 4x4 or 3x4 single or double precision float matrices 60 times a second.

Single core processors will be dinosaurs until some revolutionary nanotech comes along. Industry and academia saw multiprocessing coming a long time ago. Five years ago, I would have been excited about this, because unloading hefty FP processing off of a single-threaded CPU would have been a godsend. However, with CPU's that process six, eight or more threads on the horizon it seems like it's a bit too late. With consoles like Xbox already ditching components (APU) to be handled by a powerful mult-core CPU and the PS3 using the multi-core Cell as a broadband/media processor as well what's stopping the next wave of desktop processors to implement something like the PPU on chip?

When I made my "useless" comment, I did not have a Pentium 4 and an Athlon64 in mind.

When do you see a general-purpose processor being capable of performing like this one supposedly does? I don't see it happening soon.

yeah, you simply cant beat something that was made specifically to do the job. my car can easily go race round a track, but if i wanna do it fast and proper then ill need a specilised track car....something like a lotus exige would be nice....ahhh.....mmm...

 

[slash196]

Originally posted by: DAPUNISHER

Originally posted by: slash196
It'll have to be PCI because the PCI-e slot is taken by the GPU. Kind of a no-brainer.

Oh, so there are no boards with multiple PCIe slots available with a variety of speeds from 1-16x right now :roll: No brainer, more like no brain? J/King of course but you are definitely having a brain fart to think the vid card is the only PCIe slot already available to be filled on some boards already out.

Yeah, I had a brain fart. Although with the huge-ass video cards around nowdays, some of those PCI-e slots are blocked.

 

[slash196]

This developer seems convinced...

 

[otispunkmeyer]

if u dont like it dont buy it...how hard is that?

 

[DAPUNISHER]

Originally posted by: otispunkmeyer
if u dont like it dont buy it...how hard is that?

But it is so much more fun to shake their fists at the sky and bitch about everything under the :sun: than subscribe to logic and reason It is the age old human condition of fearing change despite the inevitablity of it.

 

[Duvie]

"Physics and graphics are both areas where dedicated hardware can exploit the problem domain's underlying parallelism to deliver far more performance than a sequential CPU."

pretty much sums it up why this needs to be taken out of the cpus hands.....


Still seems if I read it right there is a need for dual gpus to have taht one gpu to process the physics data....Did others read that as well....Will this really need to be in conjunction with next generation graphics cards???

 

[MrControversial]

Originally posted by: DAPUNISHER

Originally posted by: otispunkmeyer
if u dont like it dont buy it...how hard is that?

But it is so much more fun to shake their fists at the sky and bitch about everything under the :sun: than subscribe to logic and reason It is the age old human condition of fearing change despite the inevitablity of it.

Change is inevitable and expected, but this seems more backwards than forewards. I welcome change, but I was hoping to see it go in a different direction like what the next gen consoles are doing. It looks like consoles and PC's are diverging again.

 

[ribbon13]

Bleh, let them develop it. As long as it's part of the graphics card I don't care. I just don't need YACTB. (Yet another component to buy)

 

[otispunkmeyer]

i jus think its nice to have the option (very soon) boost processing power for hopefully not alot of money. if its got a nice price and theres plenty of support for it, and benchmarks show theres a definate gain then id consider buying it......i wouldnt be able to afford these massively mulitcore cpus of the future untill i get a good job with my degree. if i could have some of that power now and for a decent price then why not.

 

[DAPUNISHER]

They are just referring to how concurrency eliminates some of the previous constraints. Here is the way it'll work succinctly put

The adoption of 3D GPUs reduced the CPU rendering load in most games, but interestingly, as graphics content expanded in scale and scope, the requirements on the CPU for preparation, housekeeping and other tasks increased. A similar phenomenon can be anticipated with the adoption of PPUs. The increased depth and quality of physically interactive environments will expand the requirements for AI, game logic and even rendering. In short, the CPU ?thinks and orchestrates?, the GPU ?renders and displays?, and the PPU ?moves and interacts?, and all complement each other as a platform for spectacular gameplay.

The CPU will be freed to handle a tougher AI load this way, perhaps making my speculation about a dedicated AI engine unecessary.

 

[Mark R]

This is actually quite an exciting product.

Physics has been starting to become a limiting issue - look at some recent games: sure the water looks nice and bump-mapped, with some sprites to represent a splash - but think how much better it could look if the water was truely fluid. How about trees that waved in the wind, or had branches that could break off? What about a car that actually took damage realistically - instead of just having a door fall off when a critical level is reached?

It's actually interesting that developers are now starting to turn to custome silicon to perform such tasks - I'd always considered them the domain of the general purpose CPU, although there has been a lot of work recently on accelerating some physics operations using GPUs (e.g. collision detection). However, while some applications can, with some trickery, be offloaded onto the GPU, not all can.

The important thing about specific-purpose processors is that they can be made much more powerful and much more cheaply than general purpose CPUs. There are GPUs on the market now which are essentially 16 core processors. Their cost is a fraction of that of a general-purpose CPU, their power consumption greatly smaller, yet their number-crunching strength is enormous by comparison. There has been some research in using the enormous power of modern GPUs to help in computer analysis of medical X-rays: A single GPU can considerably out-perfoum multi CPU xeon workstations.

Just to give another example - you can now buy encryption processing units for high-end servers. Busy secure web-sites need huge amounts of CPU power for the encryption - what do you do if you've already got 8 Xeons? Specific purpose encryption cards can encrypt data as fast as a SCSI RAID system can supply it - at a price comparable to a single CPU.

The development of such products is unlikely to be significantly affected by the emergence of multi-core CPUs. Multi-core CPUs are one way of increasing computational power, cheaper and more effective than using seperate processors. However, if your task is repetitive and parallelisable, which 3D rendering, geometry and physics all are - then a single-purpose solution will give performance far in excess of that of a general purpose solution, and often at significantly reduced cost.

Of course, there's no reason why such specific purpose processing units couldn't be integrated into a general purpose unit. However, I suspect that this will be less good than an add-on processor board - too many different product levels (CPU speed, co-processor cores, co-processor speeds, etc), different types of circuit design may have different optimal balances for manufacturing, general purpose computing platform has historically had low-speed interconnects due to upgradability - a single purpose board can have much higher CPU-RAM interconnect bandwidth.

 

[Duvie]

Good stuff Mark R!!!

 

[MrControversial]

Originally posted by: ribbon13
Bleh, let them develop it. As long as it's part of the graphics card I don't care. I just don't need YACTB. (Yet another component to buy)

But you won't get to experience Doom4's awesome physics engine that's so good I can't think of a cool marketing acronym for it.