The New and Improved "G80 Rumours Thread" *DailyTech specs at #802*

I think it's because it's mentioned in the description of the videos he linked.

[edit] sorry I guess someone was faster than me with this answer

 

Since when has that site ever been the official Crysis page?! From their about page:

This site is an Unofficial Crysis website and has no association with EA or Crytek.

 

What am I missing? I can't find a link to the video.. is there a need for some kind of registration?

[edit] oops... I sorry about that.. I found the link

 

First you have to wait for 25 seconds, then you have to enter a code and click on a link.
They are on the lower half of the page.
Search for "Enter code" and "Download the file".

 

Can't open that thing...
It keeps saying the code is invalid and resets the counter back to 30 seconds.

 

Individual PCs per thread are necessary to manage branching. But that doesn't necessarily imply that diverged threads can run simultaneously, does it? It would just reduce performance (drastically). Isn't that how things were anyway in previous architectures?

 

So I guess on my MRX700 I get to see it like the zip links.

Personally I prefer the visuals of the first link, but I assume the gameplay is supposed to tell us it's open-ended sandbox play, and what better to show that than a drunken buggy-ride to the beach followed by a swim and then jumping off a moving boat.

It'd really be nice to see a full DX9C/DX9L vs DX10 comparo, but I have a feeling it's subtle things like the waves, the better diffused lighting like coming out of the canopy, and the improve smoke & fog.

 

Thanks Micron, It's looking good despite the low res, couldn't help laughing at the underwater shadow about 34s in though, it bore no relation to the actions of the character swimming.

Great demonstration for a non techy like myself to see quite vividly that shadows are generated entirely separately from the charcater model.

Not having a go at crysis tho, presumably still early code.
[/offtopic]

 

No problem GrapeApe who posted above you is the one who turned me on to the goodness

 

Dont say that...

Hopefully games like this will lure you away from the notebook kick you've been on this last couple of years.

 

Damn have to tear myself away from those Crysis vids, ahem...

Well I'm the one who can never actually bring myself to stay awake for the entire patent, so...

But it seems in this context that a "thread" is defined by PC, not by "fragment" (or "sample", if you prefer). So, apologies for the misdirection on that.

It looks explicit, but I'm not convinced. I'm actually trying to avoid reading the bloody patent I'm afraid.

When a triangle, say, is rasterised, it results in multiple batches being created. All the fragments are going to execute the same shader, but they will not execute in lockstep, because they're each in their own batch. So fragments in batch 37 can be executing instruction 48 while fragments in batch 38 are waiting to execute instruction 12.

In older GPUs (e.g. NV40) batch 38 can't start until batch 37 has completed execution of the entire shader as there's one PC across the entire GPU (actually on G7x batch 38 might be executing on a different quad-pipeline in parallel to 37 - G7x supports one PC per quad-pipeline).

Similarly, batch 37 might execute loop 3 in the shader, while batch 38 doesn't. This might cause batch 38 to overtake 37 (if the loop is full of tortuous texturing).

This is the nub of the problem. Predication is just a workaround for having a SIMD pipeline. If shading was through truly MIMD pipelines then predication wouldn't exist.

But the question of predication is not relevant to this patent because under predication the PC progresses according to the longest execution path. This patent isn't about PC progress, it's only about the assignment of threads (PCs) and how to track thread state (which includes keeping track of the PC between context switches).

Yeah, that's always going to be the killer. It seems like a scaling problem, fundamentally. When your register file is 200KB, then one technique works. When it's 2MB, you need to change tactics...

I'm going to blame that on patent vagueness - not saying more than the absolute minimum. It's why the patent is so yawn-inducing because the bloody concept is so simple and nothing truly interesting happens at the detail level to explain why this stuff is patent-worthy. It's bland, generic and not remotely novel.

You can safely read it as thread=batch, though.

It's incredibly generic language designed to cover all possibilities in one go.

Jawed

 

I didn't like the Crysis video that much myself. Yes, it had a great-looking environment, but I was really put off by the character shadows, particularly in the water. The shadow didn't match swimming animation we could see in first-person mode at all. Hopefully they'll fix the little details like that before the game ships.

 

I think leaving it where Demirug took it is sufficient, actually
The patent is three years old....

Yep, but, MIMD is expensive....

Yeah, I'm going to go off and give this some thought. It seems to merely be a matter of an appropriate projection space.... Of course, the alternative is simple -- bind the register state to the execution unit in question (instead of sharing the register state at a higher level, which is kind of silly, really, given that each fragment is bound to a particular stream given its position within the batch even now <shrug>). Still, the projection problem or even the reapportionment of scheduler and dispatch functionality could be fun to think about

Anyways, that's well off topic...

-Dave

 

dnavas - and just to make things even more difficult for the MIMD approach - the gradient instructions will require access to registers allocated to adjacent fragments and thread synchronization

 

A thread is defined by its context. This may include a PC; but that does not mean that the PC defines the thread. Fragment threads can't look at other threads' registers (ie: context), even if they run in parallel at the same time and have identical PCs (at least not in the programing model), so from a programmers perspective, you write a program controlling a single thread of execution for a single fragment. Repeat the above for vertices.

The fact that the PCs may be physically shared in hardware is an optimization.

News to me.

 

If the NVIO thingy actually does SLI on multi-monitors, R600's got a new checkbox to fill..

 

480mm2 Scorchio, no wonder there's a heat spreader.

http://www.theinquirer.net/default.aspx?article=35406

Although, saying that, if you assume G71 transistor density then you get that kind of size for 680M transistors. Makes me wonder if one or the other figure isn't, in fact, derived like this.

Jawed

 

Yeah, I think that their recent official "walkthrough" movie looked much better.

 

According to the german PCHG magazine, G80 weighs 681M Transistors.

 

What's it going to do for Nvidia's famed chip margins? That's a big chip, and with a second chip to deal with too for the Nvivo/Ramdac.