AMD: R7xx Speculation

[Jawed]

Just a wild guess here but could a 512bit bus restrict what they can do in terms of multiple core GPU configs? I.e. perhaps R700 (2xRV770) wouldn't be possible if RV770 utilised a 512bit memory interface?

Too much complexity on the PCB perhaps?

I think it's worth considering what happens when RV770 gets refreshed for 45nm - I doubt a 512-bit bus would be possible on RV770 and it just gets even more dubious on the 45nm version. "RV" does mean something in the end.

As far as the PCB is concerned I presume it's a matter of adding layers. Then packing the memory chips tight - MXM appears to use tight packing for memory chips ...

Jawed

 

[Shtal]

http://www.rage3d.com/articles/assassinscreed/index.php?p=5

1. Altair has certainly come bearing gifts-based on what we've seen above, it seems rather clear that Assassin's Creed is the first title to implement DX10.1.

2. The impact that DX10.1 has is quite frankly greater than expected. A 20% improvement in AA performance is very impressive.

I wonder about upcoming ATI RV770 under DX10.1 mode, it would probably destroy - GF9800GTX using same game test since nvidia G92 supports DX10 only

 

[Rangers]

Weird, why does AMD suck so much in DX10, and now get a boost from 10.1? Or do all cards get boost from 10.1?

And I thought 10.1 wasn't ever going to be used? Or was that just Nvidia fud?

According to this article http://www.dailytech.com/NVIDIA+AMD+Set+to+Square+Off+Once+Again+This+Summer/article11451.htm

AMD partners claim Radeon RV770 will make its debut this summer under $300.

That would seem AMD is shooting for a similar market segment as 3870, aka just off the top. If the prices are true though, it's nice, it should force Nvidias hand and maybe we'll keep seeing great solutions in that 200/sub 200 range (but of course even better, being next-gen parts) as we do now with the 8800GT, 3870, 3850, and 9600GT all being exceptional values.

 

[AlphaWolf]

Weird, why does AMD suck so much in DX10, and now get a boost from 10.1? Or do all cards get boost from 10.1?

And I thought 10.1 wasn't ever going to be used? Or was that just Nvidia fud?

According to this article http://www.dailytech.com/NVIDIA+AMD+Set+to+Square+Off+Once+Again+This+Summer/article11451.htm


That would seem AMD is shooting for a similar market segment as 3870, aka just off the top. If the prices are true though, it's nice, it should force Nvidias hand and maybe we'll keep seeing great solutions in that 200 range (but of course even better, being next-gen parts) as we do now with the 8800GT, 3870, 3850, and 9600GT all being exceptional values.

The 3870s are all pretty much under $200 now aren't they?

 

[Tchock]

Why 512-bit?
With 256 you can scale the chip's memory modules. GDDR3 for the mainstream, 4 (probable) for a presumed initial highend, 5 for the "refresh" (by darn marketing).

Flexibility from an overall standpoint is a torrent of win compared to a single place margin look.

PCB is simple enough for mobile, even 2-chip mobiles. I doubt that the GT200 ever gets into a notebook unless they get another MXM standard in...

 

[AnarchX]

Hmm, anyone willing to believe RV770, per chip, will utilise ~128GB/s?

Doubled Z over RV670?

 

[CarstenS]

Doubled Z over RV670?

GF9600 GT get's about 48M Zixels through its ROPs every second. And that is with a mere 57.6 GB/s of bandwidth.

A doubled z-fill for a rumored 900-MHz-RV770 with 16 ROPs would yield 57,6 GZix/sec - you'd need about 70 GB/s to saturate that (on G94-arch). So, 128 GB/s seems to be a bit too much (R600 anyone?)

 

[Jawed]

Doubled Z over RV670?

If the TUs don't bottleneck things when Z doubles, then yeah that could be possible But I'm dubious - unless there's an architectural change that also helps the TUs. Maybe I'm placing too much faith in rumours of "upto 50% faster".

Jawed

 

[Humus]

Weird, why does AMD suck so much in DX10, and now get a boost from 10.1? Or do all cards get boost from 10.1?

Only DX10.1 cards it seems. These results don't surprise me too much. Maybe a little bit larger than expected, but nothing unbelievable. DX10.1 adds access to multisampled depth buffers, and I think that this is either the entire performance increase or at least the majority of the performance increase. If you're going to combine antialiasing with access to the depth buffer you need to do have a separate render target on DX10 to store your depth values into, which multiplies the bandwidth needs several times for the pre-z pass (keep in mind that the regular depth buffer can utilize z-compression, whereas a color render target cannot).

And I thought 10.1 wasn't ever going to be used? Or was that just Nvidia fud?

Yes. I think more games are going to see this effect soon. Access to a multisampled depth buffer is such a useful feature. It's a shame it didn't make it into the original DX10 spec. Access to the depth buffer is something that more or less every modern game require, and since you likely also want antialiasing support the case for DX10.1 is an easy one to make. Especially since so many games target the consoles too, who have been able to do this for a while now.

 

[trinibwoy]

If you're going to combine antialiasing with access to the depth buffer you need to do have a separate render target on DX10 to store your depth values into, which multiplies the bandwidth needs several times for the pre-z pass (keep in mind that the regular depth buffer can utilize z-compression, whereas a color render target cannot).

Are you sure it's bandwidth related? The 2900XT has 45% more bandwidth than the 3870 and they perform identically with AA enabled. How do you explain that?

 

[AlexV]

Are you sure it's bandwidth related? The 2900XT has 45% more bandwidth than the 3870 and they perform identically with AA enabled. How do you explain that?

They perform identically in which titles?

 

[Humus]

Are you sure it's bandwidth related?

No, I'm just speculating. Could also be the loss of double-Z.

The 2900XT has 45% more bandwidth than the 3870 and they perform identically with AA enabled.

Where are you getting that from? The article doesn't test any 3870 AFAICT, only the 3870 X2, and its score is not identical to the 2900XT or even close for that matter.

 

[trinibwoy]

Where are you getting that from? The article doesn't test any 3870 AFAICT, only the 3870 X2, and its score is not identical to the 2900XT or even close for that matter.

I wasn't talking about that article or 3870X2 numbers. This was in context of your inference on DX10.1 reducing bandwidth needs when doing AA on R6xx. If that was the case the 2900XT would be considerably faster than the 3870 which is certainly not the case based on http://techreport.com/articles.x/13603/1

It seems likely that the performance gain is due to having access to the multi-sampled depth buffer but the bottleneck it alleviates is probably not bandwidth related.

 

[Humus]

Well, this of course only applies if you're using the depth buffer in the first place. I don't know if any of the games in that review do, and even then I would be very careful about concluding anything about Assassin's creed performance based on performance in other games.

 

[ECH]

Weird, why does AMD suck so much in DX10, and now get a boost from 10.1? Or do all cards get boost from 10.1?

And I thought 10.1 wasn't ever going to be used? Or was that just Nvidia fud?

According to this article http://www.dailytech.com/NVIDIA+AMD+Set+to+Square+Off+Once+Again+This+Summer/article11451.htm


That would seem AMD is shooting for a similar market segment as 3870, aka just off the top. If the prices are true though, it's nice, it should force Nvidias hand and maybe we'll keep seeing great solutions in that 200/sub 200 range (but of course even better, being next-gen parts) as we do now with the 8800GT, 3870, 3850, and 9600GT all being exceptional values.

It was fud (fear, uncertainty and doubt through and through). If ATI continues to expand their developer relationships they should be able to promote the use of DX10.1 into more titles. This should help promote the usefulness of the R700 considerably and as a side effect help increase the grow of Vista. It's already been assumed that the GT200 would use DX10.1 therefore, there was never any reason why anyone would doubt seeing it's use. All we need to see now is more games using it.

 

[trinibwoy]

Well, this of course only applies if you're using the depth buffer in the first place. I don't know if any of the games in that review do, and even then I would be very careful about concluding anything about Assassin's creed performance based on performance in other games.

Maybe I'm confused but I thought this whole thing was about ATI increasing AA performance by leveraging the new features in DX10.1 to directly access the multisampled depth buffer. Why is this app specific? Isn't a multi-sampled depth buffer a requirement for MSAA?

 

[jimmyjames123]

That's an impressive gain! So would all DirectX 10.1 cards show a speedup, or just the ATI DX 10.1 cards? How difficult is it to modify a DX10 game for DX 10.1?

 

[AlexV]

Maybe I'm confused but I thought this whole thing was about ATI increasing AA performance by leveraging the new features in DX10.1 to directly access the multisampled depth buffer. Why is this app specific? Isn't a multi-sampled depth buffer a requirement for MSAA?

How would it be but app specific?You need to have the 10.1 shader(s) taking advantage of the full acces the new DX provides-if you've got only DX9/10 code,how would that happen?

 

[AlexV]

That's an impressive gain! So would all DirectX 10.1 cards show a speedup, or just the ATI DX 10.1 cards? How difficult is it to modify a DX10 game for DX 10.1?

The only other 10.1 cards on the market are S3s,and they seem to still suffer from limited availability. In theory,they should benefit as well

 

[Humus]

Maybe I'm confused but I thought this whole thing was about ATI increasing AA performance by leveraging the new features in DX10.1 to directly access the multisampled depth buffer. Why is this app specific? Isn't a multi-sampled depth buffer a requirement for MSAA?

Yes, but it's not a requirement to be able to sample it as a texture. That's the feature DX10.1 brings to the table, and the app needs to write DX10.1 code to use it. For an application that needs the scene depth for effects like depth-of-field, soft particles etc. that's highly valuable. Without being able to sample the depth buffer you need to come up with workarounds like outputting the depth into a separate render target. But that's all assuming the app requires the depth in the first place.