Nvidia GeForce RTX 40x0 rumors and speculation

1 GPC is disabled on 4090 which means minus 16 ROPs in addition to removed SMs, etc.


Has been for some time actually even before ROPs were moved to GPCs - 970 anyone?
But here it is also possible to just cut down the size of each L2 partition because the size of L2 allows this in Lovelace.
Thank you.

As for the 970, I'd say if the memory channel can't be interleaved with the rest of them then it is effectively lost.
 
Is there a good resource for non-neurologist on how our eyesight works? I don't believe we can talk about images before the brain.

AFAIK, there is no evidence to support that assertion. There is however, evidence that compression artifacts can and do mislead the human visual system into seeing things that aren't there or misinterpreting what was actually captured on video.


Thus, any video containing any compression artifacts beyond a certain threshold require a qualified video forensic analyst if it is to be used in a court of law. And even then it may not be admissible as evidence.

While the human visual system is good at filling in details that do not exist in an image, it can often fill in those details incorrectly.

Regards,
SB
 
Might give some answers.


that video confuses me, i can't grasp what was it trying to say. halfway it drifts into laptop frame/design and cpu.

oh and its not an rtx 4000 laptop in the thumbnail

at least from what i can understand from the video. so imight be very wrong
 
any news/rumors about mobile rtx 4xxx? and usually when nvidia starts launching their mobile parts?

CES, so probably at the GeForce Beyond event mentioned in the post above yours.

As for what to expect, probably shifting every die in the stack upwards one compared desktop naming. Eg. 4090 - desktop ad102, mobile ad103, 4080 - desktop ad103, mobile ad104, etc.
 
Any hopes for a desktop 75W/no external power connector 4050 this gen? With these efficiency gains a 24 SM (full AD107), 2.4GHz in game clock (down from ~2.7-2.8GHz for efficiency) would be 2x more efficient than the 1650 by my rough estimation which was the last PCI-e power only card off the top of my head. Rest of the stack could be 4070Ti-4060Ti (AD104, 60-52-44SM), 4060-4050Ti (AD106, 36-30SM) which works, only problem is the performance increase vs the 3050 will be small so maybe 90-100W Ada, 75W Blackwell.
 
Any hopes for a desktop 75W/no external power connector 4050 this gen? With these efficiency gains a 24 SM (full AD107), 2.4GHz in game clock (down from ~2.7-2.8GHz for efficiency) would be 2x more efficient than the 1650 by my rough estimation which was the last PCI-e power only card off the top of my head. Rest of the stack could be 4070Ti-4060Ti (AD104, 60-52-44SM), 4060-4050Ti (AD106, 36-30SM) which works, only problem is the performance increase vs the 3050 will be small so maybe 90-100W Ada, 75W Blackwell.

If Ada 4050 needs 100w there’s little chance a Blackwell 5050 only needs 75w. I would really like an Nvidia 75w hdmi 2.1 card but they don’t seem to be interested in making one.
 
Why? That's pretty much every generation. Ampere 3070 was on par in gaming with 2080ti, whilst also being $700 less MSRP.
3070 was a 256bit card and really only 1 step down from the top (I pretend GA103 didn’t exist since that’s what NVIDIA did). This would be more like the 3060 being on par with 2080Ti.
 
AD104 being priced close to GA102 will also be incredible for different reasons.
AD104 has 35.8 billion transistors (assuming TPUs numbers are correct). GA102 had 28.3 billion.

Cost per transistor is increasing (or flat at best), and it’s made worse by the fact that Nvidia went from a cheap Samsung process to a very expensive TSMC process. So it’s very likely that it costs Nvidia *more* to make that AD104 chip, which is completely the opposite of historical trends. We’re going to feel that pressure on the consumer side whether we like it or not.

I don’t fully understand the entire economics behind it all but as a relative layman it seems Nvidia may have been better off snipping Ada off at AD103 and continuing to make Ampere below that, taking advantage of gradually improving yields and cost-reduced board designs to provide cheaper GA10x products, perhaps with a rebadging. It’s been done before. I suspect the reason it’s not happening is because foundry contracts and other decisions were made years in advance and all that capacity was pre-sold.
 
AD104 has 35.8 billion transistors (assuming TPUs numbers are correct). GA102 had 28.3 billion.

Cost per transistor is increasing (or flat at best) ...
Transistors ain't transistors

AD104 has half the memory bus of GA102, instead replacing it with an 8x larger L2$. It's stripping out the least dense part of a chip and replacing it with the most dense. Not even getting into things like different cell libraries, decap cells or design rule restrictions.

You can work off estimates of wafer cost and die size, but simply comparing transistor count between different devices is unlikely to lead you anywhere accurate
 
AD104 has 35.8 billion transistors (assuming TPUs numbers are correct). GA102 had 28.3 billion.

Cost per transistor is increasing (or flat at best), and it’s made worse by the fact that Nvidia went from a cheap Samsung process to a very expensive TSMC process. So it’s very likely that it costs Nvidia *more* to make that AD104 chip, which is completely the opposite of historical trends. We’re going to feel that pressure on the consumer side whether we like it or not.

I don’t fully understand the entire economics behind it all but as a relative layman it seems Nvidia may have been better off snipping Ada off at AD103 and continuing to make Ampere below that, taking advantage of gradually improving yields and cost-reduced board designs to provide cheaper GA10x products, perhaps with a rebadging. It’s been done before. I suspect the reason it’s not happening is because foundry contracts and other decisions were made years in advance and all that capacity was pre-sold.

It seems like their HW team has spent so long disregarding cost optimization in favor of adding new features that they've institutionally forgotten how to control costs. If AMD or Intel (or... Qualcomm?) can actually pick their efforts up Nvidia might be in a lot of trouble.

Of course with the launch of Arc and the current implementation of RDNA3 that's currently not happening. Well, we'll see what 2023 brings.

Transistors ain't transistors

AD104 has half the memory bus of GA102, instead replacing it with an 8x larger L2$. It's stripping out the least dense part of a chip and replacing it with the most dense. Not even getting into things like different cell libraries, decap cells or design rule restrictions.

You can work off estimates of wafer cost and die size, but simply comparing transistor count between different devices is unlikely to lead you anywhere accurate

True, it's not 1:1, but it's not like SRAM has shrunk appreciably. The idea that there's not an appreciable increase in perf/cost for Ada seems sound, since Ada has more transistors even if those transistors have shrunk. Especially when you factor in that Nvidia had to go buy TSMC without any prior relationship, TSMC likes charging a whole lot for that.
 
It seems like their HW team has spent so long disregarding cost optimization in favor of adding new features that they've institutionally forgotten how to control costs. If AMD or Intel (or... Qualcomm?) can actually pick their efforts up Nvidia might be in a lot of trouble.

Of course with the launch of Arc and the current implementation of RDNA3 that's currently not happening. Well, we'll see what 2023 brings.

Or their strategy is to keep pushing features to stay ahead of the competition to justify ever higher asking prices. Wafer cost gets a lot of attention but it’s a small part of the overall cost and pricing equation.
 
Back
Top