AMD: R8xx Speculation

I'm not very good at coming up with creative user names

 

Any word as far as Hemlock is concerned? Isn't it supposed to launch next week? How come we haven't seen a single bench yet?

 

There's more than just Hemlock coming out this year.. take your time.

 

Yes I know. I am equally interested in Fermi too!

I would just like to know what kind of performance we can expect (better or worse than two 5870s/5850s) so I can decide if I wait for Nvidia or move to the new X2!

 

Did we see this before

http://www.hardware.fr/articles/775-7/dossier-amd-radeon-hd-5770-5750.html

That's a comparison of Juniper with HD4890 at "the same bandwidth". In the past we've seen "anomalies" with GDDR5 that's been seriously underclocked, not sure in this situation.

Overall Juniper comes out ahead. Though Battleforge benefits from D3D11.

I dare say drivers are still playing their part here. And of course this is the kind of test that'll never be repeated when the drivers have settled.

Jawed

 

The BW is "same" as to Juniper, but the GDDR5 access latency is hammered twice in the RV790 case for this scenario... if it's all that important, anyway.

 

Stock HD4890 has 62.5% more bandwidth than the underclocked HD4890 - and on those terms the best advantage it has is 38% in Batman (with control panel MSAA - or 34% with AA off) and "averages" 27%. So, overall, bandwidth is not a huge problem and it seems reasonable to expect the lower performance cards (Juniper-based or lower) in a generation to be more bandwidth-limited.

Tests with 8xMSAA would have been useful, too. Particularly at lower resolutions like 1680, 8xMSAA should be used. Though you might argue that budget cards shouldn't be expected to do 8xMSAA.

Jawed

 

I think with half the clock but twice the bus width you're at a disadvantage due to latencies. That said, it's certainly interesting that Juniper is sometimes slower regardless, and certainly that's pretty much the same tests where the 5750 loses against the 4850 which was discussed in the HD 57xx review thread.

 

The NFS:Shift advantage is pretty huge as well. As always, game title selection will play its role, as will testing methods. HardOCP got a 5% advantage for the 5770 over the 4870 in that game, not 15%. Some games show a small deficit compared to an equal clocked RV790, and that's enough to make the 5750 lose to the 4850 now and then.

That review has a FANTASTIC set of data. If anyone is willing to transcribe that into an Excel file (48xx, 4890 underclocked, and 5770 scores for each game at each resolution w/ 4xAA), we can get very good estimates of how much time is spent limited by CPU, setup, bandwidth, and the rest.

 

It's probably noteworthy that this benchmark showing the largest increase outside DX11 optimizations (no matter what numbers you use) with juniper over rv770 is also where the rv770 pretty much fared worst against nvidia's chips. So the 5750 might be 20% faster than 4850 but it's still not quite enough to catch the GTS 250, and even the GTX 275 puts up a good fight against the 5850 something which it very rarely does. Would be interesting to know why it got faster.

 

Supposedly Shift has some teething issues with ATI hardware, that are about to be patched (per the game dev and ATI in a statement to [H]). So current numbers may suffer something of a shift( ) in the future.

 

Just thought it might be of interrest. Some Uningine DX11 Benchmark video.

Short video with comparison, Tessellation vs none.

http://www.youtube.com/watch?v=bkKtY2G3FbU

And a long vid with none of that, but still damn pretty.

http://www.youtube.com/watch?v=9F6zSgtRnkE&feature=related

Enjoy!

 

Apparently you can download the engine benchmark at the following...

http://unigine.com/download/

Haven't had a chance to check it out, hadn't even heard of this until now. Looks like this is what AMD used to demo some of the Dx11 features?

Their servers must be getting hammered, download is sloooooow.

Regards,
SB

 

Wow! Adaptive (LOD) tessellation in action -- looks promising!

But I'm still resisting pouring money for a 5870 upgrade -- believe me, it's not easy.

 

OK, finally got it installed, takes a while to start up, almost thought the application was hung.

Pretty damn impressive. All I've got to say is WoW. The cobblestones, roof shingles, etc were very impressive.

Although it was a bit odd, in places you would expect to see tessellation, you didn't see any. For example chimney pipes have a round hood over them to prevent rain from going down them. You'd expect tessellation there to make them round, but they are very low poly. It sort of clashes with tesselated roof shingles all around them.

It serves as a demo, but you can also take manual control and pretty much go where ever you wish and look at things in detail.

I wonder if there's any games in the works using this engine? Just going through it made me wish this was an RPG that I could roam around in.

If this is a sign of the direction tessellation will take in software titles, I can't wait. Seriously impressive stuff.

You also have the option to run it in Dx10 or Dx9. Much much less impressive obviously since tessellation is disabled. If I hadn't viewed the Dx11 version first, the Dx10 version would have been pretty interesting. But while the cobblestones have the illusion of depth (Dx10), it's pretty blah when compared to the tessellated version (Dx11). Especially once you start moving around.

Regards,
SB

 

Two new DX11 demos on AMD's dev page

 

Played with Unigine some more. Particulary impressive to watch the lighting on tessellated surfaces as you change the time of day.

Regards,
SB

 

Damn, I wish Oblivion to get that kind of makeover!

 

Yeah, I get excited just thinking about watching a fireball move over realistically lighted cobblestones at night.

A year ago if someone told me I would find Cobblestones exciting, I'd have looked at them strangely.

My god. Someone NEEDS to make an updated Dungeon Keeper with tessellation. SOMEONE MAKE IT HAPPEN.

Regards,
SB

 

*twitch*

I would play that more than BadCompany2.