NVidia Ada Speculation, Rumours and Discussion

[DegustatoR]

NVIDIA GeForce RTX 4090/4080/4070 get new rumored specs, RTX 4070 with 7168 cores and 10GB 160-bit memory? - VideoCardz.com

Yet another rumored GeForce RTX 40 specs NVIDIA hardware leaker @kopite7kimi suggests specifications change for GeForce RTX 40 series. RTX 4090 with more cores According to Kopite, the RTX 4090 would get more cores than previously reported. Instead of 16128 CUDAs, the new specs list 16384 cores...

videocardz.com

So a simpler 4070 on a considerably more advanced process and with G6 instead of G6X will presumably consume as much as 3080 10GB?
These watts continue to not make any sense to me.

 

[trinibwoy]

NVIDIA GeForce RTX 4090/4080/4070 get new rumored specs, RTX 4070 with 7168 cores and 10GB 160-bit memory? - VideoCardz.com

Yet another rumored GeForce RTX 40 specs NVIDIA hardware leaker @kopite7kimi suggests specifications change for GeForce RTX 40 series. RTX 4090 with more cores According to Kopite, the RTX 4090 would get more cores than previously reported. Instead of 16128 CUDAs, the new specs list 16384 cores...

videocardz.com

So a simpler 4070 on a considerably more advanced process and with G6 instead of G6X will presumably consume as much as 3080 10GB?
These watts continue to not make any sense to me.

kopite is all over the place on Ada rumors. So much for being a reliable leaker.

 

[Jawed]

The rumoured count of SMs is more important than the count of active lanes in a product. After all, NVidia fiddles with active lanes pretty much willy-nilly.

 

[DegustatoR]

The rumoured count of SMs is more important than the count of active lanes in a product. After all, NVidia fiddles with active lanes pretty much willy-nilly.

Same. Considerably less SMs compared to 3080.

 

[xpea]

oh oh an OEM little bird just told me that, from latest IHV marketing material, AD104 has better claimed efficiency than Navi 33. I know, it's the opposite of what the leakers are saying but my source is solid. Don't ask for more, that's all I have for you today....

 

[Seanspeed]

If we're talking just 'claimed' efficiency....remember.

That said, there's nothing hard to believe about it on the surface. Nvidia are gonna be making an enormous process leap with Lovelace, and have a great history of pushing architectural efficiency when it's a priority. Perhaps things will look different at the top end if AMD truly are going with multiple graphics tiles(which is in question now), but in terms of these midrange products, I can see AD104 winning the efficiency crown over Navi 33.

Most of the rumors of RDNA3 having better efficiency than Lovelace tended to stem from talk about the high end flagship parts, where Navi 31 was supposed to be this twin GPU beast, and Nvidia would have to push power and clocks on AD102 to ridiculous new levels to try and stay competitive in performance due to still being monolithic. Heck, even by the old rumors, Navi 33 was not only supposed to be monolithic like Lovelace, but it was also said to only be TSMC 6nm. I'm not treating that as gospel, but if that were to be the case, then there should be no question at all that AD104 will be more efficiency than Navi 33.

 

[DegustatoR]

Worth remembering that Ampere destroys RDNA2 in raytracing efficiency. So without knowing what is used for comparison all these manufacturer sourced claims - from any side - are pretty much useless.

 

[DavidGraham]

Kopite is saying Ada is operating at "much" higher clocks than 2.8GHz

https://twitter.com/x/status/1542387472902098944

 

[trinibwoy]

Kopite is saying Ada is operating at "much" higher clocks than 2.8GHz

https://twitter.com/x/status/1542387472902098944

It looks like he’s saying much higher clocks than Ampere, not 2.8Ghz.

 

[Krteq]

Twitter snips are confusing sometimes, you need to check whole thread to get context

 

[DavidGraham]

It looks like he’s saying much higher clocks than Ampere, not 2.8Ghz.

Nope, he is responding directly to the question of much higher than 2.8GHz?

https://twitter.com/x/status/1542730895181824001

 

[trinibwoy]

Nope, he is responding directly to the question of much higher than 2.8GHz?

https://twitter.com/x/status/1542730895181824001

I saw that but “it can be done” isn’t a confirmation of the intent of his prior post. Either way it seems real info is a few weeks away.

 

[xpea]

I saw that but “it can be done” isn’t a confirmation of the intent of his prior post. Either way it seems real info is a few weeks away.

No silicon has been sent out to any AIB yet. Only preliminary technical material for PCB design, power delivery, mechanical mounting and thermal solution. That's where the current leaks come from. Before the official architecture announcement, don't except much more...

 

[trinibwoy]

No silicon has been sent out to any AIB yet. Only preliminary technical material for PCB design, power delivery, mechanical mounting and thermal solution. That's where the current leaks come from. Before the official architecture announcement, don't except much more...

Does that mean no Q3 launch?

 

[DegustatoR]

Does that mean no Q3 launch?

Launches are fully supplied by the chip makers usually, especially in the high end. That's why AIB cards tend to become available a month or two later. So that in itself doesn't preclude a Q3 launch.

 

[TopSpoiler]

https://twitter.com/x/status/1543844605531418625

 

[DegustatoR]

High clocks would somewhat explain the TDPs.
Assuming a similar boost clock for 4070 that would put it at ~40TFlops or about 2X of 3070.
It would still be a sizeable perf/watt gain (+40% or so) on 3070 even at 300W.
Would also put it at about the same performance as a 3090Ti (450W).

 

[Qesa]

The 4070 and 4080 would have to be clocked much higher than the 4090 for the TDPs to make sense.

 

[pcchen]

These memory interface numbers, of course, could be due to that it's difficult to get more memory bandwidth for now, but the regression on 4080 and 4070 compared to 3080 and 3070 seem to suggest either 3080 and 3070 aren't using these memory bandwidth efficiently, or 4080 and 4070 are better at memory bandwidth efficiency (for example, with larger cache).

 

[TopSpoiler]

https://twitter.com/x/status/1543887719729807360

+8.6% clock with +22% power. Seems 450W is the marginal sweet spot.
I think "2 points" implies the 4090 would have multiple power limits depending on the power cable connected.
i.e. A single 12VHPWR can boost up to >3GHz, while 3x8-pins (via adapter) boost up to 2.8GHz.