Next gen graphics and Vista.

[kyetech]

With sources 'confirming' a december release of vista. Does this mean that WGF 2 / DX10 cards will be scheduled for a fall 2006 release?

http://www.anandtech.com/news/shownews.aspx?i=24827

Intel / IBM will be producing 65nm chips in early 2006, Apparently the transition from 90 into 65 will be easier than 130 into 90.

I wonder if we may see 65nm graphics chips in late 2006? or will it be a mature 90nm process?

Apologies for making more questions than input !

 

[Carl B]

I was actually wondering along the same lines recently. With Vista set for the end of next year basically, will the card manufacturers release DX10 compliant cards a couple of months early in order to be prepared, or a couple of months later in order to ride their present architectures a bit longer?

I imagine NVidia might be more inclined to do the later, whereas ATI - with an architectural foundation already more or less prepared - might be more willing to go with the former. And of course each company's choice might apply pressure to the others'.

Knowing more or less what R500 is, are indications that G80 will be DX10 compliant, a new architecture, and/or an evolution of NV40/G70?

 

[seahawk]

Both have one "old" tech refresh up their sleeves - G80 and R580, which should out 1Q06.

What will come after that is everybodies guess, but I think it is obvious.

 

[Carl B]

Yeah but, Q1 '06? That would give the G70 architecture only two/three quarters as NVidia's flagship technology. R580 I can understand as it's already been developed concurrently with R520 - and it shares the same 'R5xx' generational designator. So R520/R580-->R600 I think makes a lot of sense for ATI. For G80, I can't really see it coming on until Q2 at the earliest. Of course that said, a lot depends also on whether G80 is NVidia's first DX10 chip or whether it's just a G70 extension.

 

[Tim Murray]

I think there have been some implications (by Wavey, perhaps) that DX10 might not ship with Vista.

 

[Carl B]

Well if so, that would certainly lend NVidia some breathing room to extend the present architecture a little longer should G80 simply be a G70 refresh. Would allow them the standard year-long gen with it. For ATI, they'd probably still go with R600 come fall of next year I imagine, and tout the DX10 functionality same as NVidia was doing with SM3.0 for the past year.

 

[Dave Baumann]

The latest update from Microsoft (I met with them over the GDCE period a few weeks back) is that DX10 will ship with Vista, at the point in time. I suggested, then, that its logical to assume that MS, at least, would have hardware beforehand in order check the API out correctly…

As for G80/R600, first off I’m not sure why anyone would assume that G80 would be a refresh of G70 – G80 would, to me, suggest a new architecture; if there is to be a G70 refresh then I would expect to see it named something along the lines of G75 or G78. At this point in time my personal expectation for G80 would be a non-unified, but highly threaded, DX10 part. It introduction can probably be timed dependant on whether there is indeed a G70 refresh (i.e. if a refresh comes early next year, G80 could well be Q3/Q4, if it doesn’t then Q2/Q3 may be the timing for G80).

WRT R600, I wouldn’t necessarily expect to see this as a simple “yep, here it isâ€. Yes, Xenos has a unified architecture, which on the surface would appear to lend itself to DX10, it is not close to compliant. If fact, there may be other hurdles that it has to overcome as, according to Orton, R600 will primarily be done via the Valley and Orlando, not Marlborough, who has done the first implementation of the unified architecture and who I assume were the primary architects of it – there are some critical differences between R520 and Xenos that will have raised some discussions as to how to get round them.

 

[_xxx_]

Apart from any guesswork, the engineers in all those places have access to the work of other groups in the company. They'll also surely have some educational gatherings now and then to keep everyone updated.

Will say, it doesn't really matter which team does what and where, they're all surely equaly qualified and up to date.

 

[Carl B]

Thanks for the update Dave. So G80: DX10, but non-unified (possibly).

To clarify though when I was using 'refresh' I didn't mean an intra-generational refresh, a la NV38, but rather an inter-generational refresh - more along the lines of R300--->R420. Basically just wondering if G80 is an extension of G70 the way G70 was of NV40, or whether it's something 'fresh and different.'

 

[Jawed]

I just want to splash this nice diagram here:

Jawed

 

[Nick]

At this point in time my personal expectation for G80 would be a non-unified, but highly threaded, DX10 part.

What does that mean exactly? The vertex, geometry and pixel shader pipelines are still separate but execution units can be freely assigned to different shader execution threads (much like Intel's HyperThreading)?

Doesn't DirectX 10 require so much overlapping functionality between vertex, geometry and pixel shaders that a unified architecture becomes almost a necessity? I mean, next-next-generation games could use a very inbalanced combination of vertex, geometry and pixel shaders, so that one of them is always a big bottleneck.

 

[Jawed]

NVidia seemed pretty determined that unification of vertex and fragment shader pipelines was a fair way off. "Jack of all trades, master of none" was the gist of it.

Jawed

 

[Dave Baumann]

Apart from any guesswork, the engineers in all those places have access to the work of other groups in the company. They'll also surely have some educational gatherings now and then to keep everyone updated.

Will say, it doesn't really matter which team does what and where, they're all surely equaly qualified and up to date.

Errr, yes, of course they are going to be up to date. What I'm getting at is that there are certain elements that are in this generation that may not necessarily fit nicely in a unified architecture because of some rather rigid costs associated with it. Things still have to be solved in the unified architecture it a.) hit DX10 functionality and b.) have the same capabilities as the current generation.

To clarify though when I was using 'refresh' I didn't mean an intra-generational refresh, a la NV38, but rather an inter-generational refresh - more along the lines of R300--->R420. Basically just wondering if G80 is an extension of G70 the way G70 was of NV40, or whether it's something 'fresh and different.'

Actually, if you look at NVIDIA's development, from Riva128 to NV30 there is a fairly clear development path all the way up, JHH pretty much states that himself in our interview with him; ATI had previously been the ones that "throw the baby out with the bathwater" each generation by starting from a much cleaner slate (obviously not entirely clean). These roles have changes somewhat over the past few years with ATI displaying a clearer development path from R300->R520 (although this stopping with the unified architecture), whereas I think NV40 is NVIDIA's single largest architectural overhaul since NV1, so basically, who knows.

Doesn't DirectX 10 require so much overlapping functionality between vertex, geometry and pixel shaders that a unified architecture becomes almost a necessity?

Not according to David Kirk, in numerous past questioning with him (from myself and others) - IIRC it wasn't so long ago that he was saying that even under a unified shader API the balance of operations between the VS and PS were such that units dedicated and tuned to those tasks would, in their opinion, still be of more importance. Now, in more recent interviews that stance appears to have softened to the point where, IIRC again, he said unified probably would be a necessity at some point in time, which I why I think they will go that route eventually, but given the pevious comments not for their first iteration of DX10 hardware.

 

[_xxx_]

Errr, yes, of course they are going to be up to date.
...
I think NV40 is NVIDIA's single largest architectural overhaul since NV1

I had the impression that you're hinting that a particular team is more capable than the other

Care to elaborate on that second sentence? There were huge overhauls in between (32 bit color, fixed T&L, programmable T&L etc.) IMHO.

 

[Nick]

Thanks for the info! So R600 vs. G80 is going to be really interesting.

The one thing that concerns me about a unified architecture, is that so much data will have to be centralized. I'm no electronics engineer, but with dedicated pipelines you have one flow of data and little congestions (an oversimplification, I know). While with a unified architecture you have to keep track of where every data element has to go and frequently do some form of scheduling. It's more software than hardware. So a unified architecture would be close to a software renderer running on a multi-core CPU with graphics-specific instructions. While I'm a fan of software rendering (swShader), we all know it has serious limitations. "Jack of all trades, master of none" describes it perfectly.

Anyone got the same idea or am I completely off track?

 

[kyetech]

Perhaps they are hedging their bets, with a 'suck it and see' approach. Using the tried and tested hardware philosophy until proven otherwise?

Do we have any further clarification as to what seperates DX10 from DX9 and SM4 from SM3 ? (barring whats already mentioned in the DXnext article written here)?

Does Xenos hint at that, or is the design still too far removed from DX10, to gauge anything useful about the next gen API?

 

[Jawed]

Anyone got the same idea or am I completely off track?

Xenos does the complex scheduling that you're alluding to, in hardware.

Jawed

 

[Ailuros]

What does that mean exactly? The vertex, geometry and pixel shader pipelines are still separate but execution units can be freely assigned to different shader execution threads (much like Intel's HyperThreading)?

Something like "massively multi-threaded" I guess (whatever that could mean)

Question: what is more likely for R600, a completely unified shader core for PS/VS and GS or an architecture that builds/expands on Xenos with a still (temporarily) separate GS unit? Frankly the latter doesn't make all that much sense to me, but I thought I'd still ask.

As for NVIDIA - while we have seen Xenos and we somewhat have an idea where ATI is heading - there it's still a large question mark. How about though (and yes that's 100% speculative) having a separate PS units (temporarily) and unify VS and GS in units?

Finally too little is known about IMG's Eurasia, but I'd guess that they have unified all "3 types of shaders" into one single pipeline in a highly relative sense.

 

[Ailuros]

Do we have any further clarification as to what seperates DX10 from DX9 and SM4 from SM3 ? (barring whats already mentioned in the DXnext article written here)?

If they'll actually call it SM4.0 and not just simply "Shaders", then one way to describe it would be SM3.0+Geometry Shader (actually WGF2.0 claimed "unlimited resources" but it's "unlimited" down to 640*480@10fps and NV already claims "unlimited resources" since the NV40).

 

[Headstone]

How similar will DX10 be to the specialized version in xBox360? Maybe someone who has a line with a developer could get us a little inside track on the functionality and benefits/drawbacks to the new setup (aside from Gabe's lament at having to program for the possibilty of no HDD).