[Dave Baumann]

Following the Games Convention in Leipzig golem.de is reporting that following the launch of GeForce 6600 NVIDIA will move the NV4x range into the low end segment, replacing the 5200 series with an NV4x based version, and an notebook NV4x part will soon to be introduced on the NVIDIA MXM notebook video board format - a look at the PCI Express Integrators List shows NV43M, indicating that the mobile part coming first is a version of the 6600. However, evidently NVIDIA have signalled the end to the 12 month architectural cycle.

In statement that echoed ATI's previous sentiments, NVIDIA are suggesting that due to chip complexities and lithography cycle times architectural innovation times will be pushed out to about 18 months. Although NVIDIA are suggesting similar things to ATI now, if we look at NVIDIA's NV2x and NV3x generations, they have been on greater than 12 month cycles for the past two generations anyway (taking into account that NV30 was launched in November 2002, which was probably close to its intended introduction). This being the case, given NV40's announcement and wide scale availability, it would suggest that the NV50 series will not be announced until late 2005 with possible wide scale availability in 2006. NVIDIA have already stated to their investors that the low end NV4x parts will last for up to 3 years, indicating that the NV5x range will remain the domain of the high end segments and that NV4x will stay around until "Windows Graphics Foundation" in Micorsofts next geneteration OS, Longhorn.

The release of Longhorn will also be critical to both NVIDIA and ATI's plans - given the timing and the fact that there are not to be any low end NV5x, this may suggest that NV5x is set to be an extended Shader 3.0 architecture, which would indicate that the further Longhorn moves into 2007 the better it would suit NVIDIA's architectural innovation cycle. Presently the expectation is that ATI will introduce their Shader 3.0 part, suggested to be primarily developed by the R300 architectural team, in mid 2005 - ATI may be a little off their 18 month cycle as they chose not to innovate as much this cycle in order to hit the PCI Express transition - and that would also suggest that "R600" based parts would come in the late 2006 / early 2007 period. Regardless of longterm exrapolated timescales, though, its likely that both vendors will attempt to hit as close to the introduction of Longhorn as possible.

[vb]

Originally Posted by DaveBaumann

... and that would also suggest that "R600" based parts would come in the late 2006 / early 2007 period.

ETA early 2007 for a architecture that started developement in 2001...

[epicstruggle]

Originally Posted by DaveBaumann

Regardless of longterm exrapolated timescales, though, its likely that both vendors will attempt to hit as close to the introduction of Longhorn as possible.

Whats interesting is that both nv and ati are aiming at longhorn, which is a moving target. I dont think even MS has an accurate guess as to when it will definetly ship. Does either company have a plan if longhorn gets hit with major delays?

epic

[vb]

Originally Posted by epicstruggle

Whats interesting is that both nv and ati are aiming at longhorn, which is a moving target. I dont think even MS has an accurate guess as to when it will definetly ship. Does either company have a plan if longhorn gets hit with major delays?

What is really interesting is that between first gen sm2.0 lines and future longhorn suposed SM4.0 both vendors have two generation of products, so the SM 3.0 battle will be fought twice.

[Dave Baumann]

Originally Posted by vb

ETA early 2007 for a architecture that started developement in 2001...

I has a touch of "Ramageitus" about it, doesn't it? However, first purchasable implementation should turn up in the XBox2.

Originally Posted by epicstruggle

Whats interesting is that both nv and ati are aiming at longhorn, which is a moving target. I dont think even MS has an accurate guess as to when it will definetly ship. Does either company have a plan if longhorn gets hit with major delays?

They'll both be working fairly close to MS to know whats going on. I would guess that if Longhorn is even further out than 2007 then they may push for a new version of DX without it.

Originally Posted by vb

What is really interesting is that between first gen sm2.0 lines and future longhorn suposed SM4.0 both vendors have two generation of products, so the SM 3.0 battle will be fought twice.

IMO its likely to be (at least in the PC space) 1 for ATI and 2 for NVIDIA.

[vb]

Originally Posted by DaveBaumann

IMO its likely to be (at least in the PC space) 1 for ATI and 2 for NVIDIA.

I was counting R420 as a new gen product mainly because it is competing with NV40 and it is doing rather well at it.

While i am looking forward a "second take" implementation fight between R520 and NV50, i miss a Unified architecture SM3.0 from ATI competing with NV50. that would have been quite a show. but, I guess, longhorn/DXnext will be a more impartial battleground.

[Maintank]

Judging by both companies inability to get their cards out nearly 4-5 months after launch. This doesnt surprise me in the least. Realistically it looks like Sept for Nvidia before they possibly can fill back orders and get supply into the channel. And Oct for ATI or possibly even Nov for the X800XT-PE.

That leaves them with 4-5 months of milking the cow for this generation before they have to launch a new generation. Not enough time. 18 months is much more reasonable and honestly in the future wouldnt be surprised if it is extended to 24 months. Instead of new archs they will start pushing clock speeds like the CPU manufacturers do.

[DemoCoder]

Now you see some of the additional motivation to support dual PCI-E. If you've got to scale performance through 2005, but no new architectures are coming, except for some small process tweaks, you need something extra to fill the gap.

[PatrickL]

Depends if you think also that newer software will or will not take full advantage of the new bus.

[Anonymous]

I dont think democoder was referring to the bus potential but the fact with a dual PCIE board you can almost double your gpu performance by adding in a 2nd Nvidia GPU.

BTW I am under the impression that you arent limited to using the same exact card. So for instance you may beable to add in a 6600GT and when the price of the 6800GT comes down you can add it to the system and the driver will divy up the rendering duties.

Is this correct or am I blowing smoke?

[Dave Baumann]

Originally Posted by Anonymous

I dont think democoder was referring to the bus potential but the fact with a dual PCIE board you can almost double your gpu performance by adding in a 2nd Nvidia GPU.

The point is of some relevance though, since it a bit unclear how the PCIe bus is used with SLI and how the boards will apportion the tasks. Present boards have the full 16 lanes going to one board and just 4 to the second, while the next nForce board is said to switch between 16 lanes on one and none on the other or 8 lanes on both. In both cases either one or both of the boards will fall below the specified number of lanes that PCI Express graphics are supposed to have access to, so the question is what will happen to apps that use the full bandwidth - is the SLI connection itself of high endough bandwidth to cope with 8G of transfer so that it can act as the PCI Express bus for the second board?

BTW I am under the impression that you arent limited to using the same exact card. So for instance you may beable to add in a 6600GT and when the price of the 6800GT comes down you can add it to the system and the driver will divy up the rendering duties.

I recall at the launch you had to have the same board at the same clock speeds - theoretically you could have different boards but NVIDIA haven't implemented that in this solution as yet. The Alienware solution allows you to have to completely separate boards.

[Rockster]

If I'm understanding things correctly, ATI will launch the R520 next spring and NVidia will launch the NV50 next fall. The R520 will likely have an architecture similar to the R420 but support SM3.0. Wasn't the NV40 actually the NV50, so what exactly will the NV50 be now? I seem to remember David Kirk questioning whether or not a unified architecture is even the correct strategy to persue. So what areas do you guys think will receive the focus for improvements. And will Nvidia be releasing any NV4x variant to counter the R520?

[Ailuros]

I´m curious about performance ratings of multi-board setups. Even if performance nearly doubles (which I doubt at this point), I can see it only as good news for the IHVs and less for the consumer. If I am to upgrade a graphics accelerator in a say 12-15 months period I´ll be normally able to sell the former GPU for at least half it´s initial price.

Blah on the other hand development cycles of games haven´t exactly decreased either...

[DemoCoder]

Originally Posted by DaveBaumann

The point is of some relevance though, since it a bit unclear how the PCIe bus is used with SLI and how the boards will apportion the tasks.

The PCIe bus isn't used to share data between the cards. NVidia's solution would work equally well if you had dual AGP motherboards. The cards share a private bus over a connector cable. The fact that it is PCIe is mostly irrelevent, since we are talking about using an SLI solution to accelerate current game architectures, not some hypothetical game that takes advantage of extra bandwidth.

What PCIe gives you that AGP doesn't, in regards to SLI, is a bus standard that allows atleast 2 or more buses with AGP8x upload bandwidth buses so you can plug in 2 cards, and new standards for power delivery as well. Before PCIe, the only way to get two AGP cards into a machine was ti use an AGP2PCI bridge.

In both cases either one or both of the boards will fall below the specified number of lanes that PCI Express graphics are supposed to have access to, so the question is what will happen to apps that use the full bandwidth - is the SLI connection itself of high endough bandwidth to cope with 8G of transfer so that it can act as the PCI Express bus for the second board?

An interesting theoretical proposition. But are there any games even on the drawing board today that require this?

NVidia's SLI solution really exists for one reason: so two Ultras or flagship cards can be combined, and post tremendous benchmarks this fall for the enthusiast crowd. It's a way to scale performance while RAM, process, and architecture catches up.

I really doubt that any games will be released in the next 2 years which require full PEGx16 to play.

[Sxotty]

I was hoping they would allow two different boards in sli as well, that would be awesome.


On another topic (the one at hand), what is the likelyhood of the big 2 releasing cards every other year, i.e. ATI release in December '05, Nv in december '06, and so forth, is this never going to happen b/c of fear of the OEM market loss? It seems then they could trade the enthusiast market every year.

[Anonymous]

The PCIe bus isn't used to share data between the cards.

That's not true. Render-targets and vertex buffers are copied between the two cards. In order to maintain predictable performance, NVidia imposes the identical card restriction. Check Chapter 6 from the link below.

http://download.nvidia.com/developer...ming_Guide.pdf

Back to the subject of the news post, I understand where Dave is getting most of his assumptions from, but I personally don't see ATI getting to .09u (ie. R520) that much ahead of the NVidia (ie. NV50). ATI has typically be cautious of targetting new processes for its complex parts, and NVidia has been more agressive. Couple that with the fact that they are both at the mercy of the same foundries, and it doesn't seem the gap will be that large (ie. spring to fall). But you never know.

[Ailuros]

As a sidenote there's a reason why it's worth reading sites like B3D and stay as much away from sites like this one:

http://www.theinquirer.net/?article=18029

I'm still not sure if I want to laugh or cry while reading that one above....can someone please tie me up before I hurt myself?

------------------------------------------------------------------------------

NVidia's SLI solution really exists for one reason: so two Ultras or flagship cards can be combined, and post tremendous benchmarks this fall for the enthusiast crowd. It's a way to scale performance while RAM, process, and architecture catches up.

Simple example: I currently own an AGP slot platform and a 6k8nonU. I don't see much reason why I shouldn't replace it with a 6800GT later on, instead of buying another 6k8nonU. If you encount the possibility that I could resell the 6k8, I'll pay far less with that sollution than with two similar boards.

For the Ultras there's also one more consideration: power consumption. There's no chance in hell that for 2 Ultras a 450-500W PSU will be enough, which immediately translates into more added cost. The competition will have to have a sollution with a better price/performance ratio than that. Shouldn't there be any than NV will have a winner; in any other case I don't see any clear advantages with that idea either to be honest.

[John Reynolds]

Originally Posted by Ailuros

As a sidenote there's a reason why it's worth reading sites like B3D and stay as much away from sites like this one:

http://www.theinquirer.net/?article=18029

I'm still not sure if I want to laugh or cry while reading that one above....can someone please tie me up before I hurt myself?

That article was written, according to its author, to punish someone at ATI for their attitude. Great journalism, eh?

[Tim Murray]

The point is of some relevance though, since it a bit unclear how the PCIe bus is used with SLI and how the boards will apportion the tasks. Present boards have the full 16 lanes going to one board and just 4 to the second, while the next nForce board is said to switch between 16 lanes on one and none on the other or 8 lanes on both. In both cases either one or both of the boards will fall below the specified number of lanes that PCI Express graphics are supposed to have access to, so the question is what will happen to apps that use the full bandwidth - is the SLI connection itself of high endough bandwidth to cope with 8G of transfer so that it can act as the PCI Express bus for the second board?

But, we're dealing with cards that are fundamentally AGP thanks to our good buddy the bridge chip. So, as long as you have eight lanes, won't that match the bandwidth of AGP8x to each card? I don't see that as a limitation. I'd be surprised if we ever encounter a bus bottleneck on these cards; I'd bet they'd be memory bandwidth or fillrate limited far before the bus ever came into play.

[Reverend]

Originally Posted by John Reynolds

That article was written, according to its author, to punish someone at ATI for their attitude. Great journalism, eh?

Ignoring the tone of the author, I found some truths in that article however.

[Mariner]

Originally Posted by Reverend

Ignoring the tone of the author, I found some truths in that article however.

...mixed in with many inaccuracies/errors and plenty of FUD.

[Ailuros]

Originally Posted by Reverend

Ignoring the tone of the author, I found some truths in that article however.

Whereby the persentage of "truths" takes up how much of the whole nonsensical drivel? 10%? 20%? I'd trust even a half way ignorant teenager to hit a better rate than that.

Given the fact that competition between ATI and NVIDIA is a constant and ongoing process I don't see any clear winners as of yet; especially all future prospects encounted. Sales and market penetration figures do tell one side of the story too and the least I can say is that ATI has been "owned" as of yet. Tables turn way too fast in the markets those two are addressing.

John,

Journalism? I guess people like that would be better off at some low quality trash celebrity gossip magazines. At least they could yield there a couple of heavy lawsuits for the publishing company; bad publicity is still publicity they say.

[Anonymous]

And I'm probably get married by that time, so it won't matter to my wife

[Nyo_S23]

If 520 architecture is going to be pixel shader 3.0, than shouldn't nvidia release nv45 since its 3.0 shader architecture. It really depends on what features Microsoft wants to implement into direct x or its longhorn version. Why bring out nv50 when you can use NV45?????????????????

[Megadrive1988]

It sounds like Nvidia to possibly ride out all of 2005 with NV4X products. I don't think NV50 is gonna be out until early 2006 or spring 2006.

R520 is obviously coming Q2 2005 and Xbox2 (w/ R500) in Q4 2005

SM3 capable R520 in 2005 to fight SM3 capable NV45 ?

SM4 capable R600 sometime in 2006 to fight SM4 capable NV50 ?