NVIDIA GF100 & Friends speculation

[KimB]

As I said earlier, it is strange to me that they didn't show any numbers from real games against Cypress if their product is supposedly much superior.

It's standard practice in the GPU business to not show any benchmarks until the card is ready to be launched, which it isn't just yet.

 

[Arty]

Not only that, if what Fudzilla are reporting is true (mainstream GF100 parts in June) then Nvidia are clearly a lot further along with their fab process than Charlie likes to make out. Harvested GF100s will only be useful for GTX360s nothing below that.

To be fair, Fuad insisted, insisted that GT300 will be out in Q3 then Q4 and finally now Q1.

Also can you and Charlie have a got at each other via PM? I just want to deny digi from using his popcorn smiley.

 

[TKK]

Not only that, if what Fudzilla are reporting is true (mainstream GF100 parts in June) then Nvidia are clearly a lot further along with their fab process than Charlie likes to make out.

Compared to Cypress, that would be 9 months late. Compared to Juniper, still 8 months late.

Besides, even if GF100 exceeded expectations and should end up being on average ~50% faster than the 5870 and ~equal to 5970 (which, based on what was leaked/shown so far, is rather optimistic), that would still mean they needed ~550mm² (let's just assume the number is very close to reality for the time being) @ 40nm to accomplish that.

And here's the catch, and Nvidias real problem:
Half a GF100 @40nm would then be ~275mm² if you could cut everthing into half, but you can't (and for ROPs/mem interface it also doesn't make sense, see * ). Several things, like the Gigathread engine, Display output and Video Processor can't be downscaled, obviously. Maybe you can cut down the L2-cache to 256KB instead of 384, but I'm not sure if that even makes sense. So at a minimum, a 256 CudaCores/32 TextureUnits/32* ROPs/256*-bit derivative would be close to 300mm² (probably bigger), and therefore very close to Cypress (or even the same/bigger, who knows?). Do you think with those specs it will be able to compete with Cypress? I highly doubt it, not even with full 256SP/32TMU and higher clockspeeds compared to GF100.
Sure, the fastest part -might- be able to compete with a 5830/5850, but after 9 months of reaping nice margins, AMD can easily afford to lower prices into regions that make this a low-margin part from the beginning.

*24 ROP/192-bit is unlikely IMO, 768 MB is bad for marketing and hi-res performance, 1.5 GB over 1 GB negates the cost advantage of the chip itself, and of course performance would be lower too (less pixel fillrate, less memory bandwidth)

And that problem of course trickles down to mainstream/low-end parts as well. If the above chip is ~300mm², a 128SP/16TMU/16ROP/128bit-part would likely end up somewhere around ~160-170mm². That would be approx. the size of Juniper, for a part that - based on specs - would maybe land in the ballpark of 4850/GTS250/5750, performance-wise. And I think that's again a rather optimistic assumption.


To cut a long story short, it would be better (or less bad...) for Nvidia to do what AMD did with RV610/630 and go for 28nm right away. @40nm, the combination of being much later to market and most likely losing on perf. per mm²/watt would just make these parts obsolete from day one. The performance mainstream-part @40nm -might- still happen simply because they're in desperate need for a GTX260/275 replacement in the 150-250$ segment, but mainstream- and low-end DX11-parts from Nvidia @40nm? Would be highly surprised to see that.

 

[gtr]

On the other hand, they just got $1 billion from Intel and they don't have to pay to upgrade foundries any more.

Yes, AMD will likely to use $750mm to pay down the debt. In that case, the total debt should be about $4bn. The basic theme remains true though.

gtr: Maybe you should start checking up on AMD's (not GloFo's, not a year ago) long term debt - should be around half your number. But I really wonder what this is doing in the GF100 speculation thread, we got a whole thread for that here: http://forum.beyond3d.com/showthread.php?t=41627

My numbers include GloFo spin-off. These numbers are from AMD's most recent 10-Q filings, which is two months old only.

The point I was trying to make is that AMD may not have much room to reduce price (as opposed to many assumed here) and Nvidia has a lot of room to compete on price due to their other more profitable business lines.

All in all, I don't think GF100 will be at much cost disadvantage at all compared to 5800.

 

[Silus]

I don't get it. There is a separate thread for that. Isn't pre-release speculation about speculation over lack of information, and rumor mongering, etc?

That's not really what Charlie is doing though. He's turning this forum into "lessthanaccurate"...

Going back to my earlier post, I think that the 1 GPC chip (128 SPs) can indeed compete with Juniper. Given the architectural improvements and taking into account that a GTS 250 was quite competitive with a HD 4850, I see no reason for it not to be. Another win some, lose some scenario, with the edge going to this GeForce 340 (?) in DX11 titles.

 

[Jawed]

Going back to my earlier post, I think that the 1 GPC chip (128 SPs) can indeed compete with Juniper.

Really? 8 fragments per clock rasterisation rate is enough to keep up with 16?

Jawed

 

[leoneazzurro]

That's not really what Charlie is doing though. He's turning this forum into "lessthanaccurate"...

Going back to my earlier post, I think that the 1 GPC chip (128 SPs) can indeed compete with Juniper. Given the architectural improvements and taking into account that a GTS 250 was quite competitive with a HD 4850, I see no reason for it not to be. Another win some, lose some scenario, with the edge going to this GeForce 340 (?) in DX11 titles.

Redo the calculations, please. There are substantial disadvantages of the 1 GPR chip Vs Juniper, practically anything from triangle rate to texturing and pixel fill rate. And Juniper is normally more than half the performance of Cypress. I think GF100 is a great architecture, but can be somewhat difficult to scale down over a certain point.

 

[TKK]

That's not really what Charlie is doing though. He's turning this forum into "lessthanaccurate"...

Going back to my earlier post, I think that the 1 GPC chip (128 SPs) can indeed compete with Juniper. Given the architectural improvements and taking into account that a GTS 250 was quite competitive with a HD 4850, I see no reason for it not to be. Another win some, lose some scenario, with the edge going to this GeForce 340 (?) in DX11 titles.

A 1 GPC part would have only 16 TexUnits, I doubt that's enough to match a GTS 250 in texturing performance no matter how efficient they are, and in most other aspects it should be roughly equal to the GTS250. and since 1 GPC means only 1 rasterizer and 4 instead of 16 Geometry units, the "4-6 times" geometry advantage Nvidia claims for GF100 over Cypress should shrink down to 1.0-1.5 times, since Juniper matches Cypress in this regard.

So to sum it up, Juniper=Cypress for tesselation (at least on the computational side, not sure how much of a role bandwidth and so on play, but that would affect the 1 GPC part as well)), and 1/2 Cypress for everything else.

A 1 GPC-part would be 1/3 GF100 for ROPs/MC (assuming 16 ROP/128bit), and 1/4 for everything else. Maybe it will be less of a difference in reality due to higher clocks and less/no disabled Cores/TUs, but still, performance should be at most 1/3 of the fastest GF100 SKU under DX11, where Juniper is more than half a Cypress.

Reaching 5750 performance might be possible, but the 5770 should be a tough nut to crack with those specs.

 

[Groo The Wanderer]

When you are talking about cost, i assume you meant it in financial terms. Then, you have to consider all costs, not just the manufacturing side of the equation.

Look at AMD's balance sheet, they have a whopping $5bn in long term debt. Given their credit rating (below investment grade), annual interest cost is about 8%. The interst cost alone will be around $400 mm for AMD every year.

What about Nvidia? None, they have no outstanding long term debt.

This $400mm annual interst cost is massive given AMD's annual sales of only $4bn. That's why it has been bleeding big time for the past two year. Just FYI, the year the AMD made the most money EVER is $158 mm in 2005. (At that time, they have only $1 bn long term debt.)

It is AMD who can fight the price war, not Nvidia.

If you want to generalize it to the whole company, I agree with your logic. That said, it was specifically stated that GF100 could sell for whatever _BECAUSE_ it also sold as a GPGPU part. My counter was exactly what you stated, basically that you can subsidize any single part with any other part, or none at all, just sell it for less than cost. No magic.

-Charlie

 

[Jaaanosik]

Software Tessellation.

So what's an input into the PolyMorph Engine? Is it only data?
Anandtech:

While the PolyMoprh Engine may sound simple in its description, don’t let it fool you. NVIDIA didn’t just move their geometry hardware to a different place, clone it 15 times, and call it a day. This was previously fixed-function hardware where a single unit sat in a pipeline and did its share of the work. By splitting up the fixed-function pipeline like this, NVIDIA in actuality created a lot of work for themselves. Why? Out of order execution.

It seems there are going to be some instructions (software) within the Polymorph Engine.
One would say it's a software tessellation after all, though very fast.

 

[willardjuice]

Yikes, I take a nap for a few hours and the place goes to hell. Remember folks this is a Fermi thread. Just keep repeating that to yourself over and over if you have to.

 

[Dave Baumann]

A 1 GPC part would have only 16 TexUnits, I doubt that's enough to match a GTS 250 in texturing performance no matter how efficient they are, and in most other aspects it should be roughly equal to the GTS250. and since 1 GPC means only 1 rasterizer and 4 instead of 16 Geometry units, the "4-6 times" geometry advantage Nvidia claims for GF100 over Cypress should shrink down to 1.0-1.5 times, since Juniper matches Cypress in this regard.

Well, from a technical point of view it would have to clock the same as Juniper for it have the same setup rate (not that it would make a blind bit of difference!).

 

[Sontin]

It seems there are going to be some instructions (software) within the Polymorph Engine.
One would say it's a software tessellation after all, though very fast.

From ChrisRay:

Polymorph Tessellation Engine:
Vertex Fetch -> Tessellator -> Viewport Transform -> Atribute setup -> Stream Output
The tessellator units are hardware. But the entire pipeline is programmable.

http://www.rage3d.com/board/showpost.php?p=1336139459&postcount=1

 

[Squilliam]

On top of that, is internal costing relevant? I think the more important question is what NV sells the parts for, and that has zero to do with cost.

-Charlie

I see what you mean but I don't think its worthwhile to contradict you over a minor technicality when you have bigger fish to fry.

So any predictions on the mainstream Fermi derivatives?

 

[sethk]

The thing I'm most curious about regarding Fermi derivatives (i.e. 'half-Fermi' parts or smaller, as opposed the salvage parts for the top end Fermi) is whether they will be ready to launch simultaneously with the top end Fermi part, or if they will have to be launched later.

 

[mczak]

The thing I'm most curious about regarding Fermi derivatives (i.e. 'half-Fermi' parts or smaller, as opposed the salvage parts for the top end Fermi) is whether they will be ready to launch simultaneously with the top end Fermi part, or if they will have to be launched later.

I think someone said q2/q3 for them, given we've heard absolutely nothing about them I'd guess that's credible...

 

[DemoCoder]

It seems there are going to be some instructions (software) within the Polymorph Engine.
One would say it's a software tessellation after all, though very fast.

DirectX11 requires "software tessellation". The fixed function tessellation stage does not produce vertices, it produces (u,v) points which must be processed by a Domain Shader to create vertices.

 

[Arty]

That's not really what Charlie is doing though. He's turning this forum into "lessthanaccurate"...

Going back to my earlier post, I think that the 1 GPC chip (128 SPs) can indeed compete with Juniper. Given the architectural improvements and taking into account that a GTS 250 was quite competitive with a HD 4850, I see no reason for it not to be. Another win some, lose some scenario, with the edge going to this GeForce 340 (?) in DX11 titles.

Lot of wishful thinking there, even trini admits they've got a scaling problem with a Fermi derivatives.

 

[trinibwoy]

Really? 8 fragments per clock rasterisation rate is enough to keep up with 16?

Well if we were to put aside for a moment the prevailing assumption that Nvidia's engineers are morons I suspect they won't produce Fermi variants that are terribly unbalanced.

I'm losing track in all the spin though (not from you specifically Jawed). On one hand geometry processing isn't an issue even on high-end cards yet at the same time it would be a travesty if Nvidia scales back geometry throughput on downmarket parts. And both angles are coming from the same people. So which one is it, cause it sure can't be both.

Lot of wishful thinking there, even trini admits they've got a scaling problem with a Fermi derivatives.

Lol, what do you mean by "even trini admits"? Didn't realize I was on trial

 

[3dcgi]

This is an absolute non-truth, especially when talking about professional and commercial graphics. There are lots of markets for them, I've had all kinds of companies, from digital signage to Aquariums(!) asking for boards! Precisely the same reason Matrox recently delivered an entire line of multi-display cards as well.

When I was at Matrox in the early 2000s we were told that the multi display products alone brought in enough revenue to cover payroll. So I'm not surprised this has kept them alive. Though I wonder how long they can keep it going.

At one point Matrox was working on a big deal with McDonalds, but I don't know if they won the contract. Just noting that this is another data point for how varied the multi display market is.

So you blame DX11 for NV delay?

This question wasn't addressed to me, but I'm not sure if it was answered. Nvidia has indirectly pointed to DX11 being the delay. If they increased geometry performance because of tessellation and implemented tessellation because of DX11 then DX11 is to blame.