NVIDIA GeForce RTX 40 Series "AD104" Could Match RTX 3090 Ti Performance

[Lei]

Well, regardless of how anyone views this incoming generation, I think we can all agree that the TDP of these GPUs are becoming a bit concerning. In two years from now, are we going to be staring at a 750w RTX 7060? It sure as hell looks that way right now. And I highly doubt that there will be significant improvements to Power Grids ...

Nope, 60 series will be multiple chip module. They can't shrink transistor size any more.
A silicon atom is 0.2 nanometer. So power draw won't go up with MCM

4090 would probably be less than 450w of 3090Ti
And 4090ti won't have 48gb vram
Kopite7kimi doesn't have many followers despite how widely he's referred to, so he grabs a beer watches a Liverpool match then tweets 800 and watt not (что нет)

 

[ppn]

4070/10 160bit compared to a 3080 Ti 1199, $399 is asking too much for a 160 bit card, so it's very likely that we get 2GB less at 1/3 cost. This is why they had to sell 3080/12 as fast as they could at $729-799. and it's now gone.
10752 Cuda 256 bit and 7680/192 bit cards make perfect sense, very proporional and linear scaling. if the chances for defective channels are so high nvidia should have designed it with 256 bit just like 2070. but why so much cutting, probably 192 bit can allow for a low production cost and low hash rate.

Nice, but, Knowing Nvidia it is possible it will also approach the price of a 3090 ti.

3070 starting at only 499 compared to 2080 Ti 1199 is what we know.

 

[wolf]

Disappointed if true,

Unlike so many commenters that effectively claim to own a crystal ball, I'll wait for the real products. I really doubt they're launching any card that matches a 3090Ti and has a board power of 400w.

 

[TheoneandonlyMrK]

4070/10 160bit compared to a 3080 Ti 1199, $399 is asking too much for a 160 bit card, so it's very likely that we get 2GB less at 1/3 cost. This is why they had to sell 3080/12 as fast as they could at $729-799. and it's now gone.
10752 Cuda 256 bit and 7680/192 bit cards make perfect sense, very proporional and linear scaling. if the chances for defective channels are so high nvidia should have designed it with 256 bit just like 2070. but why so much cutting, probably 192 bit can allow for a low production cost and low hash rate.

3070 starting at only 499 compared to 2080 Ti 1199 is what we know.

I mean you say Know, Huang upped price on the actual day of announcement last time so Know,, might be overstated.

 

[ppn]

Unlike so many commenters that effectively claim to own a crystal ball, I'll wait for the real products. I really doubt they're launching any card that matches a 3090Ti and has a board power of 400w.

N5 technology provides about 20% faster speed than N7 technology or about 40% power reduction

You don't need a crystal ball for this. it's 20% faster OR 40 % power reduction.
All we get is a shrinked version of 3090 Ti. At 98 Mtr density the 300mm2 die packs 29000M transistors.

unfortunately for N7 and N10 it says AND, so it's probably still OR, a mistake was made.

~20% speed improvement, and ~40% power reduction

~15% faster speed and ~35% less power consumption

 

[Sisyphus]

[♥...]Only a complete ignoramus would think that enabling RTRT doesn't increase GPU workload and thus increase power usage.

If a GPU is throttled, rt used, RT cores doesn't increase power usage, power is channeled off from cudacores to rt-cores (cudacores clocked down). Rt cores are about 10 times more efficient in ray tracing than cudacores, its a positive tradeoff. If the GPU is not throttled, rt used, I would estimate about 5-10% higher power usage, as 5-10% of the chip surface are rt-cores. Workload in applications like games are changing quickly, there is a dynamic relationship between cuda-core, rt-core and tensor-core power distribution. It has to be measured for every application separately or roughly estimated. I am using a lot of RT. In this case, the GPU is always in throttling mode at max temp or max TBP. Undervolting is recomended.

To 192 bit bus. The higher the frequency, the more problems with a broad bus. They calculated, how many resources are spent for higher clock rates, how many for 160/190/256 bit. It was cheaper to gain performance, using higher clock rates/memory compression technologies/cashing, than invest in stabilizing 256 bit bus/PCB production. These are gaming cards, not hpc, where other prices are paid.

 

[lexluthermiester]

To 192 bit bus. The higher the frequency, the more problems with a broad bus. They calculated, how many resources are spent for higher clock rates, how many for 160/190/256 bit. It was cheaper to gain performance, using higher clock rates/memory compression technologies/cashing, than invest in stabilizing 256 bit bus/PCB production. These are gaming cards, not hpc, where other prices are paid.

I don't care what the reasoning is. Memory makes or breaks a GPU. I expect high-clocks AND wide bus bandwidth for anything upper midtier and above. Won't touch it otherwise. As I said elsewhere, the leasons learned from the GTX460 and GTX970 are important to remember.

 

[efikkan]

Well, regardless of how anyone views this incoming generation, I think we can all agree that the TDP of these GPUs are becoming a bit concerning. In two years from now, are we going to be staring at a 750w RTX 7060? It sure as hell looks that way right now. And I highly doubt that there will be significant improvements to Power Grids around the world, to handle such power draw.

How many computers will you be running in your basement?
And do you think this will matter compared to people using induction cooktops which typically draw 6-7 kW in bursts? Or people charging their cars?

"But I do wonder though, how does this affect the frame rate consistency?"
It's better with UV, because you usually don't hit power limits and temperature limits which can result to frequency drops and hence frame rate consistency.

Undervolting is best thing you can do now with cards on the edge of their stability because of auto oc. There is no risk of crashing during movies because u are changing 3d frequency. Usually you are able to find stable UV during first hour and that' it. Now you have stable card with the same performance which is colder, quieter, has lover consumption and longer life span.

Undervolting is a very short-sighted "solution", as the silicon ages it will become more unstable. Computers have enough annoyances as it is, I'm not going to introduce new problems like that.

 

[Lei]

How many computers will you be running in your basement?
And do you think this will matter compared to people using induction cooktops which typically draw 6-7 kW in bursts? Or people charging their cars?

EV cars would have otherwise use fuel and pollute the air.
What's the point of having power hungry gpu?

We'll have one in our basement, but town will have total of millions of cards.

 

[nguyen]

Undervolting is a very short-sighted "solution", as the silicon ages it will become more unstable. Computers have enough annoyances as it is, I'm not going to introduce new problems like that.

Undervolting would actually be slowing down the silicon aging since it reduces power consumption and lower thermal.

Once you know the ballpark for undervolting, it takes a minute. I just bought my nephew a 3070 and it took me 1 minute to set up an undervolting profile for him, lowering stock power consumption from 210W down to 120W, my nephew has been playing with the new 3070 for 3 days now and no complaint about instability. Basically it can be set&forget.

 

[Sisyphus]

I don't care what the reasoning is. Memory makes or breaks a GPU. I expect high-clocks AND wide bus bandwidth for anything upper midtier and above. Won't touch it otherwise. As I said elsewhere, the leasons learned from the GTX460 and GTX970 are important to remember.

If you think so, no problem. I am not interested in technical details, but system performance, stability and software support.

 

[cellar door]

Correct. Power draw is tested with RT off. Also the "regular" game tests are with RT off, so they can be compared to other cards. The RT on tests are on the "Ray Tracing Performance" page.

It doesn't matter anyway. All Ampere cards run in their power limits during gaming, so RT on doesn't increase power draw. You can test this easily, just look at GPU-Z power numbers while playing with the settings

Umm, that is not correct. Sorry @W1zzard

Of course RT ON increases power draw - anyone running an undervolted Ampere card (for various reasons) will easily notice that enabling RT adds a huge amount of power consumption, as all of a sudden a large portion of the chip now is active. There are literally a couple billion transistors now consuming power (while otherwise they sit idle/inactive).

Anyone can test this easily for themselves.

Example:
- RT Off, card is capable of 1900mhz within a set power limit.
- RT On, the card can lose 100mhz easily to stay within that same power limit.

That is because the gpu needs to lower the voltage and therefore frequency. Power scales quadratically with voltage - just look at any voltage/frequency graph that also shows power consumption.

 

[lexluthermiester]

If you think so, no problem. I am not interested in technical details, but system performance, stability and software support.

Those are good things to be interested in, but do remember, you can't have good performance without good specs. 192bit memory bus is not good for an upper midtier. Such a limitation artificially gimp's performance and IMPO is unacceptable.

Undervolting would actually be slowing down the silicon aging since it reduces power consumption and lower thermal.

This and it's a very good reason to do so. Undervolting a easing back the clocks will improve GPU longevity.

 

[W1zzard]

Of course RT ON increases power draw - anyone running an undervolted Ampere card

Of course. I never said RT ON doesn't increase power draw ever, I said it doesn't increase power draw for cards running in the power limit (like 99.9% of Ampere cards out there). Sorry if I wasn't clear enough on what I meant

 

[efikkan]

EV cars would have otherwise use fuel and pollute the air.
What's the point of having power hungry gpu?

We'll have one in our basement, but town will have total of millions of cards.

Well done, you totally missed the point. Did you read the line I underlined?
The question was whether the power grid would handle it, I mentioned some other loads which are much more concerning.

And how big is your town if there are millions of cards gaming at the same time?

Undervolting would actually be slowing down the silicon aging since it reduces power consumption and lower thermal.

Once you know the ballpark for undervolting, it takes a minute. I just bought my nephew a 3070 and it took me 1 minute to set up an undervolting profile for him, lowering stock power consumption from 210W down to 120W, my nephew has been playing with the new 3070 for 3 days now and no complaint about instability. Basically it can be set&forget.

Wow, no complaints in three days and you think it's stable.
Microchips have a safety margin to account for chip variance and chip degradation. If you remove this margin, then over months or years it becomes gradually more unstable. I'm fine with people taking a conscious and informed decision to experiment, but undervolting should never be a recommendation to cope with power draw, as it will lead to a lot of undefined behavior which is hard to troubleshoot for even most enthusiasts.

 

[Lei]

Well done, you totally missed the point. Did you read the line I underlined?
The question was whether the power grid would handle it, I mentioned some other loads which are much more concerning.

And how big is your town if there are millions of cards gaming at the same time?

The power grid can not handle it. See the bitcoin banning. It did in fact cause blackouts in numerous places.

My town is 2.6x bigger than Toronto. With population more than 20 of the European countries.
Our province is around the size of Spain

There are blackouts specially in summer in many countries because people use more AC or whatever reason. You don't want to add 800w salt of gpu to the injury.

 

[nguyen]

Wow, no complaints in three days and you think it's stable.
Microchips have a safety margin to account for chip variance and chip degradation. If you remove this margin, then over months or years it becomes gradually more unstable. I'm fine with people taking a conscious and informed decision to experiment, but undervolting should never be a recommendation to cope with power draw, as it will lead to a lot of undefined behavior which is hard to troubleshoot for even most enthusiasts.

Lol there is no such thing as safety margins for chip degradation (only chip variance or ASIC), GPU will crash when it degrades (most likely after warranty end) and you would have to downclock/undervolt it when RMA is no longer an option

Undervolting would greatly extend the lifetime of your GPU since it removes clock fluctuation (which is really bad for VRM), reduce power and thermal (which are main factors of chip degradation).

After undervolting my 3090 for 1.5 years, I think I know the ballpark for undervolting Ampere, and I left quite a good margin for stability (I only spent a minute setting up the undervolt LOL)

 

[Sisyphus]

Of course RT ON increases power draw - anyone running an undervolted Ampere card (for various reasons) will easily notice that enabling RT adds a huge amount of power consumption, as all of a sudden a large portion of the chip now is active. There are literally a couple billion transistors now consuming power (while otherwise they sit idle/inactive).

This is just a misunderstanding. RT-cores use power, activate them draws power, yes. If the card reached its maximum temp or board power consumption, activated RT cores power usage has the same amount, as the cudacores are throttled. If you sum it up, the power consumption of the card stays the same.

If RT is needed, the tradeoff is positive, the frame rates would go significantly down, if you use cudacores for raytracing instead of throttling cudacores and activate rt-cores.

Those are good things to be interested in, but do remember, you can't have good performance without good specs. 192bit memory bus is not good for an upper midtier. Such a limitation artificially gimp's performance and IMPO is unacceptable.

If clock speed, caching, compression are maxed out, its quite possible, we will see broader memory bus again. It really depends on what technology is advancing rapidly. PCB production is conservative, slow advancing.

 

[medi01]

30 sents. If it used to be 10 sents

In which country was it 10 cents in 2021, but 30 cents in 2022?

 

[Lei]

You're trying to dimiss his leaks with those tweest but that first tweet is actually pretty accurate. The only thing he didn't know was the SM layout and that the CUDA cores per SM doubled. He predicted the raw specs of Ampere more than a year before launch.

He accurately leaked the raw layout of all Ampere chips and that it was using Samsung more 1.5 years before launch. That was the most impressive leak imo.

He also leaked TSE scores back then for GA102 just like he did this time.

I'm pretty sure those SKUs were being tested. They just didn't release. Companies test many potential products that never release

This guy is sincerely making fun of us. Have a look:

He says 16gb he say 12gb. these are his accurate 4080 rumors.

The whole thing will be announced 15 days after his tweet, and he still doesn't know the vram.

 

[john_]

In which country was it 10 cents in 2021, but 30 cents in 2022?

The cost in Greece jumped from around .12 per kWh to over .40.
I guess other European countries have seen major price increases, others lower than ours, others maybe even higher.

This war affects Europe, in a very very significant way.

 

[napata]

This guy is sincerely making fun of us. Have a look:

View attachment 261043

He says 16gb he say 12gb. these are his accurate 4080 rumors.

The whole thing will be announced 15 days after his tweet, and he still doesn't know the vram.

He has no clue about what a 4080 or 4070 will be. He just knows the potential test boards but not what they'll be. That's a reply on the leak from someone else that there will be 2 4080 models: 16 & 12GB.