NVIDIA Fermi: Architecture discussion

[Tchock]

That seems more like GF108 to be honest. Not GTX460 definitely (unless the 336 rumors are totally invalid, which still makes little sense on where nVidia is positioning this)


(And seems right for ATI to focus SI on its lowest end since GF108 seems to be packing quite a big punch, a first so far for mid/lower-end mobile chips; Redwood and Pine feel rather lethargic for the power they're using)

 

[rpg.314]

7SM's in the lowest end part would be quite a jump from it's immediate predecessor. GF106 is more likely then gf108, imo.

 

[neliz]

Official scores will shows that GTX460=GTX465 in general, heck the 460 will score better in some benches (texturing, I'm looking at you!)

I.O.W. you'd be crazy to buy a GTX465.

 

[Picao84]

Official scores will shows that GTX460=GTX465 in general, heck the 460 will score better in some benches (texturing, I'm looking at you!)

I.O.W. you'd be crazy to buy a GTX465.

If that was to be true with those specs (224) it would be epic win for nVIDIA

 

[neliz]

If that was to be true with those specs (224) it would be epic win for nVIDIA

Now, if only the card would be a lot cheaper than a HD5850, it would be "epic", but alas.

 

[Picao84]

Now, if only the card would be a lot cheaper than a HD5850, it would be "epic", but alas.

I meant epic win because of a 224 shader chip beating a 352 one
It would mean nVIDIA had transformed a terribly ineficient architecture in an eficient one. Or at least maintained GT200 level of efficiency, which thinking a bit more, was not brilliantly efficient also

 

[TKK]

Is it guaranteed that GPU-Z can really detect the correct number of cores per SM? Maybe it can only really detect the number of active SMs and the 32 cores-per-SM was a guess by the developer of GPU-Z when that version of the tool was developed.
But whatever, what matters in the end is performance, price, noise, power-consumption. I'm really looking forward to all the reviews.

 

[KimB]

I meant epic win because of a 224 shader chip beating a 352 one
It would mean nVIDIA had transformed a terribly ineficient architecture in an eficient one. Or at least maintained GT200 level of efficiency, which thinking a bit more, was not brilliantly efficient also

Not efficient in what sense? The GF100, like the GT200 before it, are vastly more efficient in terms of making use of their compute power than ATI's hardware. They are less efficient in terms of performance per watt, granted, but when we're talking about a change in the number of shaders versus performance, we're talking about the former type of efficiency, where nVidia currently rules supreme.

 

[Picao84]

Not efficient in what sense? The GF100, like the GT200 before it, are vastly more efficient in terms of making use of their compute power than ATI's hardware. They are less efficient in terms of performance per watt, granted, but when we're talking about a change in the number of shaders versus performance, we're talking about the former type of efficiency, where nVidia currently rules supreme.

Yes I was talking about energy efficiency.

 

[no-X]

Not efficient in what sense? The GF100, like the GT200 before it, are vastly more efficient in terms of making use of their compute power than ATI's hardware.

Depends on the point of view. Theoretical GFLOPs compared to gaming performance puts nVidia's GPUs in better light. But theoretical Z and theoretical geometry numbers compared to gaming performance lead into opposite conclusion.

 

[TKK]

Not efficient in what sense? The GF100, like the GT200 before it, are vastly more efficient in terms of making use of their compute power than ATI's hardware. They are less efficient in terms of performance per watt, granted, but when we're talking about a change in the number of shaders versus performance, we're talking about the former type of efficiency, where nVidia currently rules supreme.

If adding additional compute resources adds less transistors and power consumption then adding logic to make more efficient use of existing ressources, then I think the former is the more efficient approach.

I mean, what is it worth that Nvidias shader cores are 4 times more efficient if they consume 6 times as much power and die area?

 

[Tchock]

Is it guaranteed that GPU-Z can really detect the correct number of cores per SM? Maybe it can only really detect the number of active SMs and the 32 cores-per-SM was a guess by the developer of GPU-Z when that version of the tool was developed.
But whatever, what matters in the end is performance, price, noise, power-consumption. I'm really looking forward to all the reviews.

GPU-Z doesn't detect anything AFAIK. It's a database with preconfigged, NDAble data.

What's ridiculous is that Coolaler has a card already and refuses to take any pics of the control panel to show stats while flaunting 3D06/3DV results and card pics.

 

[neliz]

What's ridiculous is that Coolaler has a card already and refuses to take any pics of the control panel to show stats while flaunting 3D06/3DV results and card pics.

Since world+dog already has the GTX460 cards in-house I'd expect way more than these small leaks right now. One good thing for GF104 is that it's development time took quite long so you'd see custom cards from day 1, that's good for pricing.

 

[Novum]

So does anyone know for certain how GF104 does supply its three vec16-ALUs with only two warp schedulers?

Or are there really three schedulers and the fancy graphics are at fault?

 

[DegustatoR]

So does anyone know for certain how GF104 does supply its three vec16-ALUs with only two warp schedulers?

Or are there really three schedulers and the fancy graphics are at fault?

Both schedullers of GF104 are able to issue two instructions per clock to SIMDs/LSUs/SFUs.
Look at this thread for some discussion on this topic. Read Anand's GTX460 article also.

 

[iwod]

New Rumors suggest that Nv is working on GTX 490 with Dual GF104 inside. Which would outperform the GF100 based GTX480. That is an interesting move as it is basically doing the same as ATI, Sticking 2 Chips for Enthusiast market with chips build off for performance / mainstream.

Of coz GF100 would continue to live as it Fit into the best Single Chip Performance once its yield and power can be tuned down. As well as the Workstation and GPGPU segment,

The Current GF104 is A01 silicon, which means it is very early production. I suppose with further revision we could get 1 more core unlocked, and higher frequency, to combat the Half Generation ATI chips; South Island.

 

[DegustatoR]

That is an interesting move as it is basically doing the same as ATI, Sticking 2 Chips for Enthusiast market with chips build off for performance / mainstream.

It's not exactly the same because they'll still have GF100 (in it's B revision by the time probably).

 

[rpg.314]

It's not exactly the same because they'll still have GF100 (in it's B revision by the time probably).

Is this thing still coming?

 

[NathansFortune]

Is this thing still coming?

It wouldn't make sense for Nvidia to stick A2/3 and get a neutered chip when a silicon respin would fix a lot of their problems and increase yields...

 

[Silent_Buddha]

Unless they were ditching GF100 entirely as ATI did with R520/R600. And instead there's a new modified chip coming out for the fall. Not saying there is but it's entirely within the realm of possibility.

Regards,
SB