It seems to me up until now all the most highly acclaimed graphics technology has been very "traditional" in nature. All the "untraditional" designs seem to be very troubled (NV3x everyone seems to hate, Playstation 2 has lower performance than it was supposed to, Power VR was successful in Dreamcast but did not hold up well against the competition in the PC space). The most popular tech right now (R350) is traditional with a few modifications (early Z etc.). R400 was supposed to be something wildly different but it was canceled in favor of R420 which is supposedly a heavily modified R300. I think until the limits of silicon are reached these "traditional" designs will be the technology leaders.
are "untraditional" architechtures a bad idea?
- Thread starter Josiah
- Start date
Freak'n Big Panda
Regular
When the kyro 2 was first released it was a very competitive part. I think the PS2 has proven its ability to produce nice looking graphics as well. The Kyro was cheaper and faster than the direct competion (GF2 MX), if its done properly any type of architechture can suceede in this market.
I don't think that "non-traditional architectures" have an inherent disadvantage, except that developers which already know "traditional" architectures will need to think differently. Won't comment on PS2/Dreamcast, as I hardly know anthing about them...
The NV3x, however, IMHO doesn't suffer per se because it uses a less traditional approach. nv3x just suffers because the design violates the KISS principle (keep it simple, stupid). The nv3x chips can do quite a bit more than is needed by both major graphic APIs, it has features which you could call "bloat" (two float formats (plus ints) in the PS, branching in PS, native sin/cos support etc. all adds considerable design complexity (and considerably increases transistor count) for not much reason).
And I'm still puzzled why the nv3x has such low register count, hurting ps performance even more. This just feels so wrong - clearly nvidia must have known that the x86 architecture is suffering because of the low GPR count and all newer chips have added lots of tricks like register renaming to get around that, yet still on the brand-new nv3x instruction set nvidia only has 2 registers??? But I disgress...
In the end, I think there is nothing preventing you from developing a good non-traditional architecture - just make sure the design makes sense, just being non-traditional is not enough. Cost is likely to be higher than a more traditional design though I think, so you should make sure that there are actually benefits from using a non-traditional approach.
The NV3x, however, IMHO doesn't suffer per se because it uses a less traditional approach. nv3x just suffers because the design violates the KISS principle (keep it simple, stupid). The nv3x chips can do quite a bit more than is needed by both major graphic APIs, it has features which you could call "bloat" (two float formats (plus ints) in the PS, branching in PS, native sin/cos support etc. all adds considerable design complexity (and considerably increases transistor count) for not much reason).
And I'm still puzzled why the nv3x has such low register count, hurting ps performance even more. This just feels so wrong - clearly nvidia must have known that the x86 architecture is suffering because of the low GPR count and all newer chips have added lots of tricks like register renaming to get around that, yet still on the brand-new nv3x instruction set nvidia only has 2 registers??? But I disgress...
In the end, I think there is nothing preventing you from developing a good non-traditional architecture - just make sure the design makes sense, just being non-traditional is not enough. Cost is likely to be higher than a more traditional design though I think, so you should make sure that there are actually benefits from using a non-traditional approach.
I keep reading similar argumentations lately over and over again. IMHO NV3x is too traditional to be a "less traditional approach", even if it sounds like an oxymoron.
At the end of the day do you really think that the end users cares what each architecture can do theoretically on paper and on paper alone? Both IHVs proposed their ideas of a dx9.0 design to Microsoft and the latter opted for the one that seemed more attractive. Besides the end user is not only turned away by lacklustering dx9.0 performance alone; NV has been feeding it´s customerbase with the same old tired AA/AF algorithms for generations now and while one may argue pro/contra about AF implementations at ATI´s side, ATI did actually improve their AF implementation quality wise while sustaining the same performance levels at least; au contraire NV is doing nothing else than potentially degrading it´s AF quality as time goes by.
Yes the NV3x has probably very good presuppositions to scale with less modifications into a PS/VS3.0 design, but they´ve quite an amount of work to do too to reach the competition in other departments.
Sum it all up and I´d say that the end-user has gotten today a very traditional GeForce, as the name actually implies.
Actually the problem with PS2 is that (in some ways) it's too traditional. The rasterization chip is basically something like 8 Voodoo Graphics chips bolted together with 4MB eDRAM, in some ways even simpler. On the other hand, the transform + lighting units (VU0 +VU1) are difficult to manage but are more flexible than either the Flipper's T+L or even the NV22.5's twin VS units! Odd.
One of the downsides to "non-traditional" architectures is that it seems the costs of a less than stellar implementation are becoming more and more prohibitive. I'm beginning to wonder if some of what Nvidia is reportedly doing isn't somehow related to some desperate pleading from the accounting department.
In addition, I don't know how well firmed up intellectual property standards are. If a company were to produce a more "traditional" approach, who's to say it won't wind up infringing on another "traditional" architecture? The legal costs from dealing with just the allegation would be huge.
Along those lines, the overhead associated with maintaining and developing expertise high performance 3d architectures is probably only going to grow higher with each product generation. While there's no shortage of ideas, there are only so many optimal solutions and finding new ones isn't going to get easier.
Nvidia and Ati both show signs that the air is getting kind of thin at the top, since their high end product lines both have substantial roots in IP purchased from other companies. In the case of 3dfx and Gigapixel, it's starting to look like being innovative may be great for the evolution of the field, but is becoming a less than viable business model.
It may be that the future of 3d technology will become dominated by several monolithic 3d commodity companies whose income is derived from IP they can appropriate from small-cap companies and "published" through their economies of scale much like how game franchises are becoming commodities for software publishers, with all but a few large developers becoming contract labor.
In addition, I don't know how well firmed up intellectual property standards are. If a company were to produce a more "traditional" approach, who's to say it won't wind up infringing on another "traditional" architecture? The legal costs from dealing with just the allegation would be huge.
Along those lines, the overhead associated with maintaining and developing expertise high performance 3d architectures is probably only going to grow higher with each product generation. While there's no shortage of ideas, there are only so many optimal solutions and finding new ones isn't going to get easier.
Nvidia and Ati both show signs that the air is getting kind of thin at the top, since their high end product lines both have substantial roots in IP purchased from other companies. In the case of 3dfx and Gigapixel, it's starting to look like being innovative may be great for the evolution of the field, but is becoming a less than viable business model.
It may be that the future of 3d technology will become dominated by several monolithic 3d commodity companies whose income is derived from IP they can appropriate from small-cap companies and "published" through their economies of scale much like how game franchises are becoming commodities for software publishers, with all but a few large developers becoming contract labor.
PVR_Extremist
Newcomer
Precisely my thoughts.
What IMGTEC would give for a licensee like nVidia.
IMO The more nVidia fall behind the more likely (albeit still a small likelihood) that they will look to more radical approaches (like PoweVR) to claim back the performance crown.
What IMGTEC would give for a licensee like nVidia.
IMO The more nVidia fall behind the more likely (albeit still a small likelihood) that they will look to more radical approaches (like PoweVR) to claim back the performance crown.
PVR_Extremist said:Precisely my thoughts.
What IMGTEC would give for a licensee like nVidia.
IMO The more nVidia fall behind the more likely (albeit still a small likelihood) that they will look to more radical approaches (like PoweVR) to claim back the performance crown.
nVidia got GigaPixel technologies from 3Dfx buy.
Maybe they'll find them some use.
Isn't the NV3x the first 3Dfx-nVidia chip ? (if so no wonder why it failed so miserably ;p [I never liked 3Dfx])
Only when they fail.
One note I'd like to add is that if a "non-traditional" architecture succeeds, it's going to seem less traditional as time goes on simply because it will become more common and people will start thinking of it as "normal". The only things that are still non-traditional after they've been around for a while are things that either:
A. Things that are good designs but can't be duplicated for legal reasons or lack of know-how.
B. Things that aren't good enough for people to want to duplicate it.
Because of this, what we think of as non-traditional these days are those that either didn't succeed, or are restricted in some way. It's the same way with CPUs. Does RISC seem non-traditional at this point?
Nite_Hawk
A. Things that are good designs but can't be duplicated for legal reasons or lack of know-how.
B. Things that aren't good enough for people to want to duplicate it.
Because of this, what we think of as non-traditional these days are those that either didn't succeed, or are restricted in some way. It's the same way with CPUs. Does RISC seem non-traditional at this point?
Nite_Hawk
I'm not too crazy on untraditional architecture.
Untraditional architecture means that developpers will have to work extra to make the games work on such architectures or else face the risk of exposing millions of bugs.
Do developpers really have the time to waste on such things? I think not.
Untraditional architecture means that developpers will have to work extra to make the games work on such architectures or else face the risk of exposing millions of bugs.
Do developpers really have the time to waste on such things? I think not.
Ailuros said:
I pretty much agree with this, and have never seen that the "non-traditional" moniker ever authentically applied to nV3x, and it more often than not was a sentiment used to connect the dots between nVidia's very inconsistent statements about the architecture (which turned out to merely be an attempt to portray its 4x2 organization as 8x1, by deliberately confusing ops with pixels, to make it sound more like the R3x0.)
I think that the increasing useage/dependence on pixel shaders might be termed pretty "non-traditional" in some key respects, too. I think we're past the point where non-traditional can only signify something like TBR. And if indeed it is true that nV3x is actually non-traditional for different reasons, I can only say that it seems to be a negative mutation for doing what 3d chips, regardless of architecture, are expected to do.
Simple as it sounds, someone has to make the first move. If everyone just played safe the entire time, progress would eventually grind to a halt. Sure, the NV3x experiement didn't work, but maybe it'll pay dividends for them in the future.
Then again, maybe it won't. Time will tell.
Then again, maybe it won't. Time will tell.
Similar threads
