Reddoguk:

New twist - Maybe it's intentional so crypto farmers don't want them. If next gen is 600w+ i think a big crypto farm would melt or they'd need some incredible form of cooling. Now imagine racks of these things and the currently increasing energy prices to boot. Might not be worth it even if 4XXX is double 3XXX in performance. This high number of watts is only about 3-5 hours a day while gaming, the rest of the time the pc hardly uses much energy. So it isn't as bad as people think. That's just me personally, you might game longer and use other intensive software.

It doesn't matter how many watts an individual card uses. As long it is more efficient than the previous generation (which is 100% guranteed), your mining efficiency/watt will improve. You will have less cards in your crypto farm, mining faster, using the same power. On the other hand, an enthusiast home user wanting to play major title games will be using twice as much power as they did 5 years ago. So gaming is a lot less efficient than it used to be and energy costs are much higher.

Dragam1337:

Argh, i really hope that this isn't the start of a new trend... PLEASE let the aib's continue to use standard 8x pcie connectors...

I'll take one single cable vs 3 separate 8x pins

Other than that 12+4 pin, what are the others? Floppy power, Molex, SPDIF and analog for your CDROM drive? What a waste and confusing mess.

The 3090Ti is a Titan-level card. Why are people discussing it as if it were a mainstream product? It isn't. It's meant for those with more money than common sense. Gamers don't need a 3090Ti. They don't need a 3090. Heck, most don't need a 3080, but at least that one makes sense for gaming. If you're an enthusiast, then this is your hobby, and you can easily afford one. If you aren't, then what do you care about an insanely overpriced and power-hungry contraption? The people that will buy them already own a cottage, two luxury cars and are probably eyeing a Cessna.

user1:

makes the MCM rumours more plausible, 2x the gpu means 2x the power! same efficiency

not exactly. Nvidia is and has been addicted to massive monolithic dies, simply because they have no fab experience. on the other hand AMD (who sold off their fabs to finance Ryzen) has massive experience that has allowed them to work hand-in-glove with their fab partners (mainly TSMC) at the cutting edge. of course they paid for this and partially underwrote the new fabs (7nm at the time). but this allowed them the freedom to work around Moore's Law. every decrease in the node pitch increases efficiency and yield. a MCM design purposefully uses the smallest available node for both. an MCM design (depending on the i/o controller - typically a larger node) can deliver 2x performance without 2x power and heat of the previous generation. which is not to say it "sips" the kilowatt hours, but it will not "chug" them like the rtx 3090ti. a large part of the ridiculous prices for the RTX3070ti and up are because of the low yield of massive monolithic dies.

tunejunky:

not exactly. Nvidia is and has been addicted to massive monolithic dies, simply because they have no fab experience. on the other hand AMD (who sold off their fabs to finance Ryzen) has massive experience that has allowed them to work hand-in-glove with their fab partners (mainly TSMC) at the cutting edge. of course they paid for this and partially underwrote the new fabs (7nm at the time). but this allowed them the freedom to work around Moore's Law. every decrease in the node pitch increases efficiency and yield. a MCM design purposefully uses the smallest available node for both. an MCM design (depending on the i/o controller - typically a larger node) can deliver 2x performance without 2x power and heat of the previous generation. which is not to say it "sips" the kilowatt hours, but it will not "chug" them like the rtx 3090ti. a large part of the ridiculous prices for the RTX3070ti and up are because of the low yield of massive monolithic dies.

Lol, Nvidia produced lots of GA100 at TSMC 7nm too, probably even more than RDNA2, they have ton of experience working with TSMC cutting edge fab. This is a quote from Nvidia Q4 2021 earnings

Data Center Fourth-quarter revenue was a record $3.26 billion, up 71 percent from a year ago and up 11 percent from the previous quarter. Fiscal-year revenue rose 58 percent to a record $10.61 billion. Announced that Meta is building its AI Research SuperCluster with NVIDIA® DGX™ A100 systems.

Mind you AMD's revenue for CPU+GPU in Q4 2021 was 2.6bil Next gen Ada Lovelace will have an insane jump in efficiency, since they cram another 80% more cores into the same area (AD 102 will be ~600mm2, same as GA102). MCM is only for halo product since AMD can use 2-3 smaller die with great yield to compete with AD102, so probably no defective chip left for cut down GPU like 6800/6800XT for next gen (increasing profit margins). The midrange RDNA3 will have a hard time against midrange Ada Lovelace though, but since Nvidia command sky high profit margin (65%), that would leave some space for AMD to price their mid-range RDNA3 (well that's duopoly for ya)

tunejunky:

not exactly. Nvidia is and has been addicted to massive monolithic dies, simply because they have no fab experience. on the other hand AMD (who sold off their fabs to finance Ryzen) has massive experience that has allowed them to work hand-in-glove with their fab partners (mainly TSMC) at the cutting edge. of course they paid for this and partially underwrote the new fabs (7nm at the time). but this allowed them the freedom to work around Moore's Law. every decrease in the node pitch increases efficiency and yield. a MCM design purposefully uses the smallest available node for both. an MCM design (depending on the i/o controller - typically a larger node) can deliver 2x performance without 2x power and heat of the previous generation. which is not to say it "sips" the kilowatt hours, but it will not "chug" them like the rtx 3090ti. a large part of the ridiculous prices for the RTX3070ti and up are because of the low yield of massive monolithic dies.

Why stop at 2x the performance when you can have more? usually dual gpu cards struck a balance to keep costs down and keep things 'sane', however if you are targeting maximum performance, then the tdp can very well hit 2x power and beyond. in the past there were such cards, albeit not from nvidia directly, (good example is the asus mars 2)[SPOILER]https://tpucdn.com/review/asus-mars-ii/images/power_average.gif [/SPOILER] obviously such a product is not for everyone, but I can see nvidia adapting a quadro or tesla product with a high tdp , extreme tdps aren't out of the question, especially for a halo product.

tunejunky:

Nvidia is and has been addicted to massive monolithic dies, simply because they have no fab experience.

lmao, this is wrong.

Krizby:

Lol, Nvidia produced lots of GA100 at TSMC 7nm too, probably even more than RDNA2, they have ton of experience working with TSMC cutting edge fab. This is a quote from Nvidia Q4 2021 earnings Mind you AMD's revenue for CPU+GPU in Q4 2021 was 2.6bil Next gen Ada Lovelace will have an insane jump in efficiency, since they cram another 80% more cores into the same area (AD 102 will be ~600mm2, same as GA102). MCM is only for halo product since AMD can use 2-3 smaller die with great yield to compete with AD102, so probably no defective chip left for cut down GPU like 6800/6800XT for next gen (increasing profit margins). The midrange RDNA3 will have a hard time against midrange Ada Lovelace though, but since Nvidia command sky high profit margin (65%), that would leave some space for AMD to price their mid-range RDNA3 (well that's duopoly for ya)

you are missing (deliberately?) the point. GA100 is a monolithic die. @ 7nm one chiplet (from Ryzen in this case - the only on the market mcm as we speak) has vastly greater yield than GA100 at over 6:1

Astyanax:

lmao, this is wrong.

Laugh all you want - name ONE fabrication plant that Nvidia has owned or operated. this is true and correct. just because Nvidia is freaking amazing at uArch does not translate to manufacturing just as a Physics teacher is out of place at a nuclear power plant.

tunejunky:

Laugh all you want - name ONE fabrication plant that Nvidia has owned or operated. this is true and correct. just because Nvidia is freaking amazing at uArch does not translate to manufacturing just as a Physics teacher is out of place at a nuclear power plant.

I don't believe that matters - companies hire manufacturing process engineers with experience. It's not like owning a fabrication plant suddenly imbues everyone in the company with the knowledge of the process there - you have specialized people for that and when those people leave (foundry splits off) they take most of that experience with them. If Nvidia poached every top engineer from TSMC, I don't think we'd be sitting here saying AMD has more knowledge because they onced owned a fab a decade ago. I think a better argument for AMD would be all the cutting edge process technology they work on - HBM/3DCache (silicon vias work)/MCM/Semi-Custom/etc. They clearly spend a large chunk of their R&D budget on gambles in fabrication in order to maintain a competitive edge and you'd presume this would continue - but I don't think this has to do with them once owning a fab plant, I think it's them choosing to allocate R&D expenses on fabrication vs Nvidia who's R&D is heavily spent on software development.

tunejunky:

you are missing (deliberately?) the point. GA100 is a monolithic die. @ 7nm one chiplet (from Ryzen in this case - the only on the market mcm as we speak) has vastly greater yield than GA100 at over 6:1

Huh? Since when did I mention that Nvidia is going to compete with RDNA3 with MCM design? Yes the point behind MCM is that by utilizing smaller dies with good yield (like 90% yield rate with 350mm2 die), AMD can make more halo product from a single wafer. Let say Navi33 is 350mm2, with 90% yield rate there are 166 perfect die from a single wafer, that means AMD can make 83 MCM from a wafer and sell them 1200usd a pop, making 99.6k USD revenue from a single wafer. Meanwhile Nvidia can make 92 AD102 chip from a wafer (~600mm2), but only 45 of them are 4090 (50% yield) and 47 defective chips are salvaged as 4080, selling 4090 @ 1500usd and 4080@700usd would net Nvidia 100.4k usd from a wafer. For midrange GPU, MCM is pointless as 350mm2 and smaller die get 90% yield anyways, no point coupling 2x 200mm2 die then spend extra for the Infinity Cache/ IO chiplet. Here Nvidia definitely has the edge from superior uarch.

Krizby:

Huh? Since when did I mention that Nvidia is going to compete with RDNA3 with MCM design? Yes the point behind MCM is that by utilizing smaller dies with good yield (like 90% yield rate with 350mm2 die), AMD can make more halo product from a single wafer. Let say Navi33 is 350mm2, with 90% yield rate there are 166 perfect die from a single wafer, that means AMD can make 83 MCM from a wafer and sell them 1200usd a pop, making 99.6k USD revenue from a single wafer. Meanwhile Nvidia can make 92 AD102 chip from a wafer (~600mm2), but only 45 of them are 4090 (50% yield) and 47 defective chips are salvaged as 4080, selling 4090 @ 1500usd and 4080@700usd would net Nvidia 100.4k usd from a wafer. For midrange GPU, MCM is pointless as 350mm2 and smaller die get 90% yield anyways, no point coupling 2x 200mm2 die then spend extra for the Infinity Cache/ IO chiplet. Here Nvidia definitely has the edge from superior uarch.

1) agreed Nvidia is the King of (gpu) uArch 2) in all forms of uArch, due to the realities of manufacturing the larger the die the lower the yield as you agreed. 3) you supposed Nvidia Lovelace 3090ti replacement with a mcm. salvaging gpus is something all do. i even made the point the other day about Nvidia potentially peeing on AMD's RDNA 3 launch with cut-downs. and your math on the yields is not even guessing and it's presented as an argument. which it is not, it is an illustration and i get your point. i do know the ballpark of the yields and you are off. what i will say is this - higher yields are always better not least because Fab prices Are Not Equal and never have been. AMD has a lower cost of fabrication than Nvidia by contract through next year and will until Intel grabs that spot in '24 AMD knows this, although they hoped Intel wouldn't be aggressive

Agent-A01:

I'll take one single cable vs 3 separate 8x pins

Right? Can't for the life of me understand why anyone would want to stick to the previous connectors...

mackintosh:

The 3090Ti is a Titan-level card. Why are people discussing it as if it were a mainstream product? It isn't. It's meant for those with more money than common sense. Gamers don't need a 3090Ti. They don't need a 3090. Heck, most don't need a 3080, but at least that one makes sense for gaming. If you're an enthusiast, then this is your hobby, and you can easily afford one. If you aren't, then what do you care about an insanely overpriced and power-hungry contraption? The people that will buy them already own a cottage, two luxury cars and are probably eyeing a Cessna.

All the people who bought a 3090 at MSRP for gaming are laughing at you right now. Hold my beer.