AMD Accelerator?

[LungExpansion]

I heard an interesting rumor that AMD may allow an interconnect between its CPU and the AEIGA PPU on the motherboard much like the old Co-Processor days. When Physics are passed to the Opteron CPU there is a dedicated interconnect which will offload the physics to the AEIGA PPU for processing and return it through the Opteron. Since we know Physics is very poor on a CPU this would be interesting.

Needless to say I think its not true but interesting enough to post.

Also not sure if this would go in the CPU or video area.

 

[erwos]

Originally posted by: LungExpansionWhen Physics are passed to the Opteron CPU there is a dedicated interconnect which will offload the physics to the AEIGA PPU for processing and return it through the Opteron.

This is totally nonsensical. You don't "pass physics" to the CPU. You pass instructions. The CPU has absolutely no idea what said instructions are for. (As an aside, 486SX's co-processor actually functioned as the only CPU - it was a full 486DX, more or less. ) I also don't know why you think regular CPUs are bad at physics calculations - the Opteron is smoking fast at FP math, which, when you get down to it, is all physics really is.

Now, your physics _library_ might include some sort of switch that essentially checks to see if an AEGIA PPU is available - if it is, then it'll use that, and if not, it'll use your processor. But a co-processor socket is tremendously overkill for this kind of thing.

-Erwos

 

[Jeff7181]

I think it's more likely that AMD will put the PCI-Express controller on die, which would essentially create an "interconnect" between the CPU and PPU or GPU.

 

[themusgrat]

Completey OT, but why boycott EA????? BF2 is a good reason not to. And having PCIe, RAM, and HD controllers all implemented on die would be the ultimate. While I'm at it, every one of them, but at least we would see huge benefits from PCIe.

 

[LungExpansion]

Originally posted by: erwos

Originally posted by: LungExpansionWhen Physics are passed to the Opteron CPU there is a dedicated interconnect which will offload the physics to the AEIGA PPU for processing and return it through the Opteron.

This is totally nonsensical. You don't "pass physics" to the CPU. You pass instructions. The CPU has absolutely no idea what said instructions are for. (As an aside, 486SX's co-processor actually functioned as the only CPU - it was a full 486DX, more or less. ) I also don't know why you think regular CPUs are bad at physics calculations - the Opteron is smoking fast at FP math, which, when you get down to it, is all physics really is.

Now, your physics _library_ might include some sort of switch that essentially checks to see if an AEGIA PPU is available - if it is, then it'll use that, and if not, it'll use your processor. But a co-processor socket is tremendously overkill for this kind of thing.

-Erwos

Thats not true. While the CPU might be good at math its a dog when it comes to Physics. I highly recommend downloading the AEIGA vs P4 demo where they put a pair of dual core P4 cpu's doing physics against a single AEIGA PPU. The CPU's are crushed and peg out at 100% cpu and could only render about 10FPS but a single AEIGA PPU had no problem doing what the 4 cores could not and was rendering well above 30FPS.

 

[Munky]

Originally posted by: LungExpansion

Originally posted by: erwos

Originally posted by: LungExpansionWhen Physics are passed to the Opteron CPU there is a dedicated interconnect which will offload the physics to the AEIGA PPU for processing and return it through the Opteron.

This is totally nonsensical. You don't "pass physics" to the CPU. You pass instructions. The CPU has absolutely no idea what said instructions are for. (As an aside, 486SX's co-processor actually functioned as the only CPU - it was a full 486DX, more or less. ) I also don't know why you think regular CPUs are bad at physics calculations - the Opteron is smoking fast at FP math, which, when you get down to it, is all physics really is.

Now, your physics _library_ might include some sort of switch that essentially checks to see if an AEGIA PPU is available - if it is, then it'll use that, and if not, it'll use your processor. But a co-processor socket is tremendously overkill for this kind of thing.

-Erwos

Thats not true. While the CPU might be good at math its a dog when it comes to Physics. I highly recommend downloading the AEIGA vs P4 demo where they put a pair of dual core P4 cpu's doing physics against a single AEIGA PPU. The CPU's are crushed and peg out at 100% cpu and could only render about 10FPS but a single AEIGA PPU had no problem doing what the 4 cores could not and was rendering well above 30FPS.

But whose demo was it? If Aeiga coded the demo, I'm sure they'd want to make it look like the cpu's couldnt handle it to create a demand for their product, but what it all boils down to is the algorithm. For example, when ripping and encoding a dvd, there's heavy use of iDCT math functions. There are several software algorithms for the same calculations, and the difference between the fastest one and the slowest one can be as much as 10x, all while running on the same cpu. With game physics, it gets even more competitive because a lot of things in games are only approximation models of the real thing. As long as the physics look convincingly real, it doesnt matter whether it's being modeled by an optimized cpu algrithm or processed exactly to the last drop by dedicated hardware. In this case, i'd actually prefer the optimised cpu model because it doesnt require additional hardware.

 

[Zap]

Originally posted by: themusgrat
Completey OT, but why boycott EA????? BF2 is a good reason not to.

EA has put out any number of games that are really fun, but they are notorious for putting out buggy code - essentially you the consumer is paying to beta test it. Answer me this: How soon after BF2 came out was there a patch for the game? We're not talking an Unreal Tournament "patch" that basically adds features and new maps, but a patch that fixes stuff they shipped out broken.

 

[linkinpark342]

*waits for fix of red/blue/green name bug as i find this the most annoying bug i've ever found in a game. Then a fix for the god damn server browser which, at this point in time, is worse than making a script to ping every ip address in the world (from some list that i have no clue where to find) and seeing which have the BF2 ports open*
BF2 is a good game at heart but EA is loosing a lot of people by not fixing their crappy coding.

back on topic: I agree with munky. I would rather see true benchmarks by a third party of the PPU before i rejoice on anything that AMD does related to this. However, an integrated pci-e link would be nice

 

[biostud]

read this:

http://theinquirer.net/?article=29155

 

[erwos]

Originally posted by: LungExpansion
Thats not true. While the CPU might be good at math its a dog when it comes to Physics. I highly recommend downloading the AEIGA vs P4 demo where they put a pair of dual core P4 cpu's doing physics against a single AEIGA PPU. The CPU's are crushed and peg out at 100% cpu and could only render about 10FPS but a single AEIGA PPU had no problem doing what the 4 cores could not and was rendering well above 30FPS.

As someone else has already noted, AGEIA isn't going to put out a demo that makes them look bad. The very fact that they used a P4, with its relatively slow FP math, isn't inspiring much trust in me. Indeed, it would be pretty easy to rig such a test - just use some relatively inefficient algorithms on the CPU side, and toss in a few corner cases where the "PPU" shines its best. They did the same sort of demos with onboard T&L, and it turned out to be not much faster in most benchmarks - if it was at all. (Anyone got the original 3DMark?)

And, really, you still make no sense. PHYSICS IS MATH! If you can do math fast, you can do physics fast, and vica versa.

I'm not saying I think their product is trash. First, I haven't seen it. Second, the idea of a physics API backed by a dedicated accelerator is appealing. Even if it's slower than the CPU, it still frees up the CPU to do other things, like graphics. I've been following this closely, because we could really use some quaternion calculation speedups on some of the older machines in our lab.

-Erwos

 

[erwos]

Originally posted by: biostud
read this:

http://theinquirer.net/?article=29155

Some idle speculation that you could use an HT link as a peripheral bus is not really all that impressive. Adding more PCIe lanes to the chipset and/or adding the PCIe controller to the CPU would seem to be a cheaper and simpler way of doing what then author suggests.

-Erwos

 

[Jeff7181]

Originally posted by: Zap

Originally posted by: themusgrat
Completey OT, but why boycott EA????? BF2 is a good reason not to.

EA has put out any number of games that are really fun, but they are notorious for putting out buggy code - essentially you the consumer is paying to beta test it. Answer me this: How soon after BF2 came out was there a patch for the game? We're not talking an Unreal Tournament "patch" that basically adds features and new maps, but a patch that fixes stuff they shipped out broken.

But BF2 is teh pwnzors, biatch! Woot! Madden roxxors 2!

 

[carlosd]

Munky and erwos are right. You woulnd't show a demo to make your product look bad or yes?. I think actual CPUs are good enough for game physics. And as said there should be alternative methods for the CPU to look better processing phisycs because in the end the phisycs are nothing more than math calculations made by a processor. And I really doubt than the AGEIA PPU have 3 times more FPU power than a dual core PD o A64 CPU.

 

[Griswold]

Originally posted by: LungExpansion
I heard an interesting rumor that AMD may allow an interconnect between its CPU and the AEIGA PPU on the motherboard much like the old Co-Processor days. When Physics are passed to the Opteron CPU there is a dedicated interconnect which will offload the physics to the AEIGA PPU for processing and return it through the Opteron. Since we know Physics is very poor on a CPU this would be interesting.

Needless to say I think its not true but interesting enough to post.

Also not sure if this would go in the CPU or video area.

I think that was mentioned in an interview with AMD's CTO (might have seen the news bit at DailyTech, not sure), so its not really a rumor. But I think AMD has professional applications in mind, like servers and supercomputer possibilities.

The question is rather, why bother with AEIGA? Upcoming game engines will simply offload these calculations to the second (or third or fourth) core if its available and a dedicated CPU crunching the numbers for physics calculations will be plenty enough for anyones gaming needs. It's just maths after all.

To me, these PhysX cards are a dead born baby. I've never felt that current games lack alot in terms of physics calculations. Perfect water and hair/fur motion only adds so much to a game experience.