The Official G84/G86 Rumours & Speculation Thread

[jimmyjames123]

Here's some food for thought: it's obvious in so many ways why there would be such a big difference in performance and cost to produce 8600 GTS 256MB vs 8800 GTS 320MB. This begs the question: is there an 8600 GTX and/or 8600 Ultra in the works to fill the gap between 8600 GTS and 8800 GTS?

 

[MulciberXP]

Here's some food for thought: it's obvious in so many ways why there would be such a big difference in performance and cost to produce 8600 GTS 256MB vs 8800 GTS 320MB. This begs the question: is there an 8600 GTX and/or 8600 Ultra in the works to fill the gap between 8600 GTS and 8800 GTS?

Unless it has a wider bus, I just don't see how.

 

[Jawed]

Yeah, I think that's quite a pickle.

It's down to architecture now. I think it's notable that 8600GTS isn't embarrassed by 3DMk06 - a benchmark that NVidia has always considered home turf. Now, if only I knew what architectural features 3DMk06 points to... (Or is it just pointing to the drivers?)

I wonder if AMD did/will rethink the switching to GDDR4 for certain parts, or instead use faster GDDR4 or slower GDDR3 than what's optimal cost-wise to separate products.

I don't think there's much choice. In the same way that G84 appears to leave a gap for a G87 (or whatever the hell it is between it and 8800GTS: 256-bit, 16 ROP), I expect a RV670, 256-bit. Initially, perhaps, AMD will release a X2900GT based on R600 to mop-up slow/faulty cores. Then there'll be an RV660 with a faster GDDR4, but 128-bit. You can have fun permuting ROP/ring-stop count and bus-width, much like the fun RV560/570 provide. (Can AMD produce a 384-bit R6xx?)

Do you happen to have any idea of pricing of 1ghz gddr3 vs. 1.1ghz GDDR4?

No. You can bet 700MHz GDDR3 is seriously cheap (because its used in both PS3 and XB360). It's also worth noting that 700MHz was cheap enough to be used on X1600XT right back, way back when, when its price crashed. I suspect it's been silly cheap for a long long time.

I've always been curious about that considering it appears they perform approx the same (OC-wise) at default voltages, which are only .1v apart.

GDDR4 is at the bottom of a range that supposedly reaches up to around 3.5GHz. GDDR4 is internally clocked at half GDDR3's rate, so you could say it's coasting right now. It should be fairly mature now, at least in its lowest speed.

Jawed

 

[Jawed]

Here's some food for thought: it's obvious in so many ways why there would be such a big difference in performance and cost to produce 8600 GTS 256MB vs 8800 GTS 320MB. This begs the question: is there an 8600 GTX and/or 8600 Ultra in the works to fill the gap between 8600 GTS and 8800 GTS?

I think the die sizes tell you all you need to know there... I wouldn't be surprised if G84 and RV630 both end up looking much like RV530 - somewhat over-priced on introduction, and obviously destined for half their price. AMD has the march with 65nm, even if it's one month+ behind. Actually, I take that back, RV630 is quite a bit bigger than we were expecting, ouch. That should translate into performance but does it (how can it with a 128-bit bus?)?...

Jawed

 

[Jawed]

Comparing 8600GTS against 8800GTS 320MB

Since I can, I might as well do a quick comparison with 8800GTS 320MB.

8600GTS has a

• bandwidth of 36.64GB/s
• with 8 ROPs clocked at 745MHz, i.e. a fillrate of 5960MP/s
• AA fillrate 23840MP/s
• zixel rate of 47680MP/s
• presumably 16 TMUs for a bilinear rate of 11920MT/s
• trilinear rate of 23840MT/s
• 64 SPs @1500MHz (guess) for 192GFLOPs (or 96GFLOPs if 32 SPs)

Compared with 8800GTS 320MB, which has:

• 64GB/s bandwidth
• fillrate of 10000MP/s
• AA fillrate of 40000MP/s
• and a zixel rate of 80000MP/s
• bilinear rate of 12000MT/s
• trilinear rate of 24000MT/s
• 96 SPs @ 1200MHz for 230.4GFLOPs

So in percentages, 8600GTS v 8800GTS:

• bandwidth = 57%
• fillrate = 60%
• AA fillrate = 60%
• zixel rate = 60%
• bilinear rate = 99%
• trilinear rate = 99%
• GFLOPs = 83% or 42%

For the games (averaged across the 3 resolutions):

• SC:CT HDR with AF = 71%
• SC:CT HDR with AA/AF = 67%
• FEAR AF = 60%
• FEAR AA/AF = 61%
• Quake4 AA/AF = 75%
• Company of Heroes = 59%

Quake4, true to form seems mostly texture-bound. SC:CT, given all that HDR, seems to be doing relatively well despite its bandwidth shortfall. Are FEAR and CoH both fillrate limited? Or is it merely bandwidth?

Arguably, 8800GTS is more bandwidth-constrained.

Jawed

 

[jimmyjames123]

Unless it has a wider bus, I just don't see how.

Yes, it would probably need a wider bus than what 8600 GTS has, but that shouldn't be a big issue for them I would think based on past designs with bus width in that range.

 

[Unknown Soldier]

8600GTS has a

• bandwidth of 36.64GB/s
• with 8 ROPs clocked at 745MHz, i.e. a fillrate of 5960MP/s
• AA fillrate 23840MP/s
• zixel rate of 47680MP/s
• presumably 16 TMUs for a bilinear rate of 11920MT/s
• trilinear rate of 23840MT/s
• 64 SPs @1500MHz (guess) for 192GFLOPs (or 96GFLOPs if 32 SPs)

According to DailyTech

GeForce 8600 GTS -- 256 MB GDDR3, 675 MHz core clock, 1000 MHz memory clock
GeForce 8600 GT -- 256 MB GDDR3, 540 MHz core clock, 700 MHz memory clock
GeForce 8500 GT -- 128 to 256 MB DDR2 or GDDR3, 450 MHz core clock, 700 MHz memory clock
NVIDIA claims these three cards will be available at launch. The 8600 GTS will fill the $199 to $229 price point, followed by the 8600 GT ($149 to $159) and the GeForce 8500 GT ($89 to $129).

News Source: DailyTEch

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

Jawed, ur specs for the 8600GTS seem over the top. If you see what this lucky aussie dewd did on his overclock(std. air) then you might have to revise your list.

http://forums.whirlpool.net.au/forum-replies-archive.cfm/715719.html

 

[Jawed]

I took those "specs" from the Hardware Zone "review":

http://www.hardwarezone.com/articles/view.php?cid=3&id=2231

which is for a "factory overclocked" 8600GTS. So, obviously, a stock 8600GTS is going to have a harder time.

Earlier, I said "drivers are out of whack". Well, if they are, then they're also out of whack for the 8800GTS-320MB, because 8600GTS's performance in those games doesn't seem out of whack according to the theoretical percentages.

• NVIDIA GeForce 8600 GTS 256MB Overclocked (ForceWare 100.95 beta)
• NVIDIA GeForce 8800 GTS 320MB (ForceWare 97.02)
• NVIDIA GeForce 7950 GT 512MB (ForceWare 93.71)
• ATI Radeon X1950 XT 256MB (Catalyst 7.2)
• ATI Radeon X1950 PRO 256MB (Catalyst 7.2)

Jawed

 

[trinibwoy]

Well ChrisRay said that there is some magical driver set on the way that should "surprise". Now since we expect there to be no bugs stability and compatibility shouldn't be considered surprises (or have we lost all faith now and we expect drivers to be crap?). So hopefully it will bring some kind of performance gain as well. The 320MB GTS still performs lower than it should compared to 256MB cards in some situations.

 

[Jawed]

The 320MB GTS still performs lower than it should compared to 256MB cards in some situations.

I have to admit that's a topic I've not been sucked into...

Jawed

 

[Nick]

Any chance a GeForce 8600 GX2 could fill the gap?

This brings me to the question; has SLI matured since Series 7? I'm not talking about the software (driver stability), but the hardware. With a unified shader architecture it seems that migrating tasks from one chip to another is more straightforward. Is this the case or does multi-GPU still have inherent issues?

Anyway, even if there's no real mid-range GX2, dual-card SLI is still an interesting option to fill any gaps. They could even sell them in pairs...

 

[Razor1]

highly unlikely, for a mid range card to have the same issues that the 7950gx2 came across (all SLi problems) it wouldn't do so well as a higher selling part.

 

[seahawk]

Look at how big G80 became compared to G71 and then compare 8600 to 7600, it is obvious that there is a card planed for in between in the future. Today high-end-Chips have become to big and expensive to fill the gap between them and the mainstram chips with only those chips. A dedicated performance chip will be the norm very soon and 1950pro showed the worth of it.

 

[ChrisRay]

Well ChrisRay said that there is some magical driver set on the way that should "surprise". Now since we expect there to be no bugs stability and compatibility shouldn't be considered surprises (or have we lost all faith now and we expect drivers to be crap?). So hopefully it will bring some kind of performance gain as well. The 320MB GTS still performs lower than it should compared to 256MB cards in some situations.

I think you may have read a bit too much into what I said. I dont believe there is any magic. Though I do believe it is much improved from the current set at nzone. I just said the bug fixes in the next revision should be a surprise to some. No matter how long the list of fixes in a driver set that occurs there will always be issues that dont get solved or arent prioritized as much as people think they should be and which ends up as unhappy customers. But the next official revision will pobably fix alot of issues certain users are having and is one of the larger advancements I've seen happen to drivers from a stability and bug fix PoV in a long time. SLI users do have certain things to look forward too as well. I'm optimistic from what I have seen. But grounded expectations are not something we should lose sight of either.

Chris

 

[trinibwoy]

Ah gotcha, so the number of fixes is the surprise. Although, to be fair the real surprise is that there have been so few fixes up to this point. But I'm sure G80 owners will be happy - better late than never. And it's in Nvidia's best interest to have the best possible showing against R600 when it ships. A couple percentage points here and there won't count for much if AMD really delivers on its promise of good drivers.

Another thing, do you have any insight into why 8800GTS-320 performance tanks unexpectedly in some situations?

 

[ChrisRay]

Ah gotcha, so the number of fixes is the surprise. Although, to be fair the real surprise is that there have been so few fixes up to this point.

Do you have any insight into why 8800GTS-320 performance tanks unexpectedly in some situations?

I updated my post for clarity. So tired I kinda rambled along there losing sight of what I trying to say.((*Edit* Sometimes its difficult to say what you wanna say when you know you have some constraints )) I have no clue about the 8800GTS 320 ((I admit I have not payed much attention to it, I'm sure it'll see some performance improvements. Keep in mind history performance variables between 256-512 cards for Nvidia)) I'm sure alot of people have big expectations from Nvidia's next set. And I think they know that. ((Talk with them sometime. They are taking the Vista drivers very seriously right now)). Theres alot of SLI fixes and some much needed SLI feature's coming back to the control panel as well.

 

[trinibwoy]

Cool, thanks Chris. As always the heads-up is appreciated.

 

[Blazkowicz]

32 ALUs according to http://my.ocworkbench.com/bbs/showthread.php? threadid=61776 (shader frequency specs are given too)

 

[stevem]

I'm not really liking the "mid-range" so far... A 64SP 256bit SKU would usefully sit under the 96SP 320bit 8800GTS. I wonder if performance would be too close, though.

 

[trinibwoy]

32 ALUs according to http://my.ocworkbench.com/bbs/showthread.php? threadid=61776 (shader frequency specs are given too)

Only 1/4 G80 for the midrange? That doesn't sound right unless Nvidia vastly overestimated the strength of their architecture.

That would put it at 5400 MT/s which is less than the 7600GT....I don't buy it.