NVIDIA GF100 & Friends speculation

[Silus]

The cash is one thing but at $299 how could they ever hope to keep it in stock? If there's any truth to this rumour I'm putting a stake in the ground that the 470 is a 384-shader part and there will be no 512-shader part on launch day.

Where's the GTX 470 showing @ $299 ?

EDIT: Ah it's in that tweaktown article!

 

[Mindfury]

http://www.guru3d.com/article/nvidia-geforce-470-480/

Last week NVIDIA's board partners finally got samples of Fermi based products. This means a finalized package - with perhaps some clock changes at best. Most of the partners received a GeForce GTX 470. Now here's the difference with the aforementioned website, we know the majority specs and some performance results, however I'm not about to share them, I mean come on ... give NVIDIA some credit here.

What I will tell you is that the clock frequencies on these boards surprised me, the GeForce 470 seems to be clocked at roughly 650 MHz, that's lower than I expected. And that indeed will have an effect on performance. I think it's safe to that the GeForce 470 and 480 will be worthy competitors to towards the Radeon HD 5850 and 5870. Will it be a knock-out ? I doubt it very much. But is it important for NVIDIA to deliver a knockout to the competition ? Well they would hope so, but no .. not really, as the current performance levels that ATI for example offers simply are superb already. Being six months late to the market does pose an issue, ATI will already be respinning and binning their products, clocked higher and they could match NVIDIA in either price or performance.

 

[CarstenS]

I like the "you save : $ -70.00" bit in the screen shot for 470

I specifically like 384-Bit-Cards with one or two Gigs of VRAM. Now I know the reason for the delay: custom made memory modules!

 

[rpg.314]

What I will tell you is that the clock frequencies on these boards surprised me, the GeForce 470 seems to be clocked at roughly 650 MHz, that's lower than I expected.

Strike 1 in favor of Charlie's claims of clock speeds.

PS: I know it's off by 25MHz, but IMHO, that's a minor detail.

 

[trinibwoy]

Charlie was referring to half hot-clock. Is that what Hilbert is referring to as well?

 

[KimB]

Charlie was referring to half hot-clock. Is that what Hilbert is referring to as well?

Must be. There's no conceivable way that's the shader clock.

 

[Jawed]

In contrast it seems that fermi maintains a 2:1 mul ratio which requires adding significantly more hardware into the design. AKA, a DP mul requires roughly 4x the area of a SP mul, where as for add its only roughly 2x.

Fermi has a float multiplier and a 32-bit int multiplier, side by side, so that basically does your half-rate DP. Larrabee is the same (ALUs can do 16-wide float mult and 16-wide int32 mult) and the two chips are prolly very very similar, because they both do half-speed DP.

Fermi also has the monster "denormal" adder on the end so that the fused multiply add maximises precision. FMA doesn't exist in Larrabee's instruction set (unless it is de-facto fused?). Well, not the version that Intel made public. So Fermi's DP is more advanced.

Jawed

 

[chavvdarrr]

I specifically like 384-Bit-Cards with one or two Gigs of VRAM. Now I know the reason for the delay: custom made memory modules!

 

[trinibwoy]

Must be. There's no conceivable way that's the shader clock.

No I'm asking whether or not Hilbert is referring to the ROP/core clock.

 

[rpg.314]

Charlie was referring to half hot-clock. Is that what Hilbert is referring to as well?

EDIT: ! If it is the core clock then half the hot clock could be ~730 MHz, not too off from GTX 285. (using the 12-14% margin between core and half hot clock).

 

[CarstenS]

Rhethorical Question: How much half hot clock would you need to completely "saturate"* a base clock of 650 MHz. *w00t*

Or the other way around: How embarrasingly low would base clock have to be, if half hot clock is at 650 MHz. Somehow, I do not buy into this "it's 48 ROPs this time, because it aligns nicely with the memory controller, despite it's 32 ppc max for the chip".

*i know, it's not the base clock that get's saturated, but the associated units.

 

[GZ007]

I specifically like 384-Bit-Cards with one or two Gigs of VRAM. Now I know the reason for the delay: custom made memory modules!

1024/12 MB gddr5 chips It seems to be fake. With those prices and the confirmed low clocks from guru3d noone would buy those cards.

 

[Jawed]

What is the area cost of an int32 multiplier vs fp32 multiplier? The former was added in fermi just to do dpfp more cheaply. int32 multiplication happens only in the t unit in ati, so clearly not something that is terribly useful for graphics.

ATI now has 24-bit multipliers on the x,y,z,w ALUs. Though there's no sign of this in IL so far, only in ISA as MUL_UINT24, MULADD_UINT24 and MULHI_UINT24.

Seems curious that 24-bit integer multiplication got deleted from Fermi.

Generally you want decent throughput for integer multiplication because of pointer arithmetic. 24-bits obviously isn't going to be enough for huge surfaces and byte addressing hurts even more.

Jawed

 

[Jawed]

If AMD can pull another "4 chips in <6 months" in 2H10, (or even by 1Q11) they will be in a shitload of trouble.

I think the rumour is for 3 chips, isn't it? Redwood, Juniper and Cypress appear to be lined-up for the felling squad, leaving Cedar standing until well into 2011.

Jawed

 

[ap_]

ATI now has 24-bit multipliers on the x,y,z,w ALUs. Though there's no sign of this in IL so far, only in ISA as MUL_UINT24, MULADD_UINT24 and MULHI_UINT24.

Seems curious that 24-bit integer multiplication got deleted from Fermi.

Generally you want decent throughput for integer multiplication because of pointer arithmetic. 24-bits obviously isn't going to be enough for huge surfaces and byte addressing hurts even more.

Jawed

I'm not sure what you mean, doesn't 24-bit integer multiplication come for free in a fp32 multiplier? That means you can't really delete it, right?

 

[Jawed]

With the next shrink ATI wil reach 3200 SP-s easyli.

That shrink is not due until summer/autumn 2011 though. A smaller chip on the next process node is likely to come before that, but Cypress's ALUs won't be doubled until then I reckon.

Jawed

 

[rpg.314]

I'm not sure what you mean, doesn't 24-bit integer multiplication come for free in a fp32 multiplier? That means you can't really delete it, right?

Yes it comes for free. Jawed is saying that it is now exposed to compiler as a separate instruction.

 

[Jawed]

I'm not sure what you mean, doesn't 24-bit integer multiplication come for free in a fp32 multiplier? That means you can't really delete it, right?

It should be "free", which is why I don't understand why Fermi doesn't have it. But then again, why has it only just now appeared in ATI?

Jawed

 

[rpg.314]

It should be "free", which is why I don't understand why Fermi doesn't have it. But then again, why has it only just now appeared in ATI?

I think that is because Cypress is their first chip which can walk the gpgpu talk.

 

[GZ007]

That shrink is not due until summer/autumn 2011 though. A smaller chip on the next process node is likely to come before that, but Cypress's ALUs won't be doubled until then I reckon.

Jawed

They wont need it with nvidias GF100 lineup. They can go away with much less and a dual gpu card.
Dual GF100 will need to wait for a shrink it seems like the gtx295.