[B3D Analysis] R600 has been unleashed upon an unsuspecting enthusiast community

That's a pretty old chip you've got there Rys.

07 - 04 .. it even has all the info on top of the die instead of the shim.

but it sure is an enjoyable read!

 

I am struggling to understand how you arrived at that conclusion based on anything that was published today so far.

R600 has a mixed record againt GTX is mixed, but what really would make it shine is 40 less ALUs, 3/4ths narrower bus, half the TMUs and 3/4ths less ROPs?

 

Sorry, don't understand that but it may be just me. Hope you can expand this properly.

 

Well, I don't see how there's a routing hit vs. binding a texture unit to a single 16x SIMD array (imagine the 80 ALUs laid out vertically, the buses to texture units laid out horizontally). Whether or not there is a caching hit depends on how quads are assigned to threads, but regardless I suspect that Enos is right - the way things are arranged it looks like one would get more even utilization of texture units. For instance you could schedule completely different thread types to different SIMD arrays (say a texture heavy pixel shader, and a vertex/geometry shader with zero texturing instructions) and still fully utilize the texture units.

If I understand G80 correctly, you'd have to take care to have each cluster executing a mix of all the thread types to achieve the same TMU efficiency (and that would lower cluster I-cache efficiency).

 

Hmm... Were we really an 'unsuspecting enthusiast community'?

I haven't done a lot of checking, but is there any word on when R600 will be available in Australia?

 

I hope AMD has something to say in the mainstream segment, but then they still cling on to NDA's even for that segment. Things do not look bright on the high end. Bad thing is AMD has to sell their top card for so much less, this would greatly reduce profit margins as we would expect to see greatly reduce prices for the entire family (we would want the mainstream segment priced accordingly)

We can also commend NVidia for coming up with state of the art filtering performance. Oh boy.

AMD tried to play the HDMI, HDCP card well I guess. But then again, how many wants to watch blue-ray, hd-dvd movies on their monitors? I don't think many people would be connecting their pc's to their plasma TV's in their living room. Media PC's would surely benefit, but media pc concept might not be as popular as xbox360's or PS3's with HD playback capabilities, or dedicated players, for that matters.

Is it possible to enhance the filtering performance of this arcitecture, so we can see increased performance on lower nm parts (semi-next gen) or will we have to wait for the next architecture?

 

Is there any way to explain in as laymen terms as possible what makes these 320 superscalar ALUs different from NVIDIA's scalar ALUs, if they can still all work on separate instructions?

In other words what binds them together as 4+1?

 

Magic

 

Imho

Based on the bench numbers I've seen so far, ATI/AMD or DAAMIT should have

A.) Released these cards last year and work on driver issues all along like Nvidia has

or

B.) Extended out the NDA more to get the drivers better as the ones that have been supplied have a 2nd Tier Highend(if XTX was around difference would be amount of memory and mem/core clocks) trading blows with a 2nd tier highend that has 3/4 the ALUs, 5/6 the bandwidth and memory of its bigger brother.

So Either the G80 design is just simply that good or ATI screwed the pouch. 3-4 more months may tell a different story after several driver releases, but right now, for me and I'm sure others aswell, the HD2900XT is nothing but a product with a hopeful future. It will be seen if it is a promising one. Glad I got my BFG8800GTS OC when I did. I've been gaming in bliss for about a month now. Glad I didn't wait to see the what if today, I'd have been pissed.

 

Such a shame, if HD2-series card ever ends up in my computer, narrow & wide tent AA modes are out of the question, edge detect is the only possibly usable CFAA mode, at least there's still 2 to 8 MSAA modes aswell.

 

Right, I understand that a SIMD will feed into a single texture unit, but why the preference of SIMDs across different clusters sharing a texture unit versus SIMDs from the same cluster sharing a texture unit? Is this just to evenly mix the workload going to the texture units which could potentially be coming from different clusters doing drastically different types of work?

 

Dito , though there is still things i like on ati's side. EATM and performance/quality aAA modes.

 

One issue I will take with "waiting for the better drivers" suggestion is that it almost assumes that while 2900 drivers improve drastically, G80 will lag behind. This, however, is hardly a replay of G70 Vs R520, which many seem to be drawing the erroneous parallel with, where former was an development of a well-established architecture while the latter was much more novel design.

I would argue that now, the opposite is actually the case: at this point, AMD software engineers should be much more familiar with R600's inner workings (after all, this is "the second generation USA" and the ring bus has been around for two years now) than their counterparts at Nvidia, because on many fronts G80 represents a much bigger departure from their respective preceding designs than R600. In other words, if you want to aplly R520/G70 parrallel, G80 would seem to be a better candidate for the part of R520.

 

Some here: a decidedly mixed bag, at this point.

Geeforcer, IIRC, R5x0 made the greatest advances with AA thanks to purported work on the MC. I'm guessing we'll see the same again, though this time they're not starting from as level a theoretical playing field WRT pixels exiting the ROPs per clock, so really no way they'll catch G80 GTX in ROP-limited scenarios. But G80 seems to use the same basic crossbar MC as before, while R600, though it's a second gen "ring bus," still seems to have changed somewhat from R5x0. So I'm thinking R600 will show more improvements with AA, but still won't reach G80GTX levels.

But if you're talking about shader or scheduler efficiency, isn't the argument that R600's VLIW superscalar approach means a greater onus on the compiler, whereas G80's scalar ALUs may enjoy greater (natural) efficiency from the start? I guess the next question is how much efficiency is left to exploit for each GPU, and how exploitable is it (tho R600 may have more room for improvement, does AMD have the manpower/time to ameliorate this sooner rather than later in R600's life cycle?).

 

Actually, I think the driver does use EATM in place of AAA. I'll check and make sure, but I think the AAA perf fix was to enable EATM instead.

 

Guru3D also has some DX10 benchmarks with ATi's Call Of Juarez demo
http://www.guru3d.com/article/Videocards/431/17/

In short, 2900XT performs prety equal with 8800GTS, and 8800GTX is ahead.

 

http://guru3d.com/article/Videocards/431/17/

 

So the driver no longer supports SS AA on alphas?

 

oh ok.

 

I will give the R600 this, it seems to be chewing up OpenGL games much better then any of the previous cores, all the way back to the R300, which i doubt is a coincidence. AOE3 performance is vastly improved as well which is also another title that caused all their previous cards to stumble. The performance seems very hit and miss which really strikes me as strange.