Time to install Vista on my spare (test dummy) HDD. Looks awesome, but I highly doubt I'll get good framerate. But who knows. I got pretty decent framerates with Crysis. Wonder how it'll be optimized, or is its sole purpose to make you feel like you need a huge upgrade?

DX10 is a joke right now. It seems to exist mostly to destroy framerates.

Most "DX10 Software" isn't DX10. It is DX9 with some low-hanging goodies which eat up extra performance. You want more eye candy, that costs. Most of the performance gains from DX10-based software aren't implimented.

This was the same story with DX9. It effectively killed Matrox and NV had horrible performance in games that dared to even use the slightest bit of DX9 features. Extra goodies come at a cost in design as well as hardware. DX9 is still the sweet spot and until the DX10 market grows, and some killer features get shown (chicken egg, egg chicken) there won't be a huge push.

I would be remiss if I didn't point out that DX9 shipped in 2002. The first games with notable DX9 features were 2004 and the first DX9-only game was 2006.

^^ Yup. DX10 is definitely the joke.

Then you don't understand what DX10 is.

This isn't like DX6-8 with a hoard of fixed function effects. DX10 is about increasing flexibility of GPUs as well as efficiency improvements for common techniques. In both cases you will need to design with DX10 in mind to tap the results.

The "DX10 only" stuff people are seeing are easy to impliment techniques that are expensive on any architectures (the difference being the flexibility of DX10 allows them at reasonable costs). Doing advanced motion blur, for example, isn't an architecture issue--it just is very resource intensive.

While there are some shortcomings in each of the designs, the performance issues seen in "DX10 software" aren't really related to those (e.g. one of the GPUs has poor streamout bandwidth--but no game has hit this hurdle yet).

See above on the timeline for DX9 software.Call it Vista... or recognize that some of these "DX10 add ons" devs have thrown in are (a) not really efficient and (b) are very resource intensive.

To put this into perspective: DX10 GPUs are extremely flexible, and while not quite to the level of being CPUs you can run very-non GPU algorhythms on them or graphic techniques that are far outside the scope of the current performance envelope. In the past the improvements to DX were mainly inline with their ability to adequately execute fixed function effects; when you open up your resources and flexibility you run into a new paradigm which puts the weight of decision making on developers: is this technique a cost effective solution for the GPU?

No one expected 6000+ line shaders to run well on DX9 GPUs (even though the HW could do it), so I am not quite sure why people would expect some of the stuff devs are tacking on to magically run great.

Currently, the new GPUs trash the old DX9 GPUs, a clear indication they are faster. And synthetic tests show that they both do as expected with the new stuff (i.e. no NV FX style cheats and mis-designs).Not following you here. While the R300 was designed to run DX9 SM2.0 code at "full speed" it was a bit faster in DX8 software.The problem with the FX was pretty simple (w/o reading the article):

NV tried to strong arm MS and went with a DX9 compliant card but went with a different "sweet spot" instead of focusing on accelerating the API to the "t". e.g. 32bit instead of 24bit for shaders and clinging to 16bit. So you got full speed 16bit (outside the DX9 spec for most stuff) and half speed 32bit (which ment less realestate working on 24bit code). They had HQ but slow AF and made a tragic mistake of not making the compiler with the design.

Essentially NV went with nailing the DX9 check boxes, focused on great DX8 performance, and tried to use new features that were non-DX9 to segment the market. And it totally backfired on them.Not really. NV's problems were from Fall 2002 to Summer 2004--nearly 2 full years of "ouch" (although their "friends" in the media covered their ass a lot, as did big marketing and game branding to offset the fact their cards sucked).

ATI had pretty big problems with R600. But the HD3*50 are relatively competitive and price slotted to match any disparity. i.e. good bang for buck. And they have the edge in features right now (DX10.1). Whereas NV was getting *killed* in early 2004 by FX revisions -- we are talking 50% less performance at playable framerates -- ATI is nowhere near that sort of craptitude.

ATI's problems have to do with being tied to a sinking ship, a loss in R&D funding, and NV/Intel snatching up the Physics API market (bye-bye AGEIA and Havok). ATI also is losing out on top tier moneys by basically not competiting at the top end.Or...

Option 3: DX10's focus isn't fixed function "click to enable" effects but is a focus in versatility/flexability as well as design efficiencies...

So a title can (a) tack on some expensive, not available before, effects that Drop DX10 performance because it is doing a lot more that cannot be done in DX9 with any sort of performance or (b) they can design their games with the efficiency changes to make better use of the platform.

B takes *years* as shown above and requires a sizable market share to motivate this shift. Interestingly Crysis on DX10 GPUs does show a number of nice effects that are **very slow** on DX9 hardware that run with a **relative** minimal performance hit on DX10 hardware (when you enable hacks to get them on XP).

DX10 isn't about IHVs screwing up; it is about a new way of thinking of GPUs. The trend within MS is shifting GPUs directly at general purpose resourses that are far more flexible than before. You can do more stuff but you need to design with it in mind and be responsible with the hardware performance limitations.

My GPU cannot do FFSW level graphics with 64x st. stitch AA and 16xHQAF at 4000x2000 so it is a failure of my GPU :|

The MB example is a perfect example of why you don't understand the issue. DX10 GPUs are faster at the old techniques and can do new techniques at reasonably levels. SM3.0 could do these new tricks, but they aren't even playable. e.g. LP has non-silhouette based MB. The praise of DX10 is a) the flexibility to do such techniques with b) reasonable performance tradeoffs.

The fact techniques, like certain raytracting tricks, can work on a DX10 GPU -- albeit slowly -- isn't a failure of the GPUs but a nod to their success. GPUs aren't the ol' school fixed function tools you seem to cast them. It isn't even an issue of techniques they accelerate.

They are general resource pools with resource limitations. It is up to developers now to properly manage the resource. Just because you CAN run something on a GPU doesn't mean you SHOULD. And just because something is slow doesn't mean it is the hardware's fault.

That is like saying a 3GHz quadcore sucks because it cannot handle 50k sphere-sphere fully dynamic physics objects in gameplay. It means then idiot programmer didn't properly assess the resource limitations/ Just because new hardware can allow something doesn't mean you should do it because the caps were taken off.

Final Reality was more of a scene demo than a benchmark program, wasn't it?