The RTX 3080 TI isn't going to be- EDIT - No We Were All Wrong!

g67575 · 25 Jul 2020 at 22:43

I wonder if a hypothetical new power connector could eventually be used to power a dual GPU Nvidia graphics card? I think that would make sense, considering the last Nvidia dual GPU card had a TDP of 300 W, even with just ~~3.1 TFs~~ 6.26‬ TFs of processing power.

Also, people with the $$$ to buy a dual GPU are also not likely to be bothered about needing to buy a new high end powersupply.

The dual GPU GTX 690 had a similar RRP on launch ($1000), to the RTX 2080 TI which shows you how much more profitable the graphics card market has become for Nvidia.

Grim5 · 25 Jul 2020 at 23:26

g67575 said:
I wonder if a hypothetical new power connector could eventually be used to power a dual GPU Nvidia graphics card? I think that would make sense, considering the last Nvidia dual GPU card had a TDP of 300 W, even with just ~~3.1 TFs~~ 6.26‬ TFs of processing power.

Also, people with the $$$ to buy a dual GPU are also not likely to be bothered about needing to buy a new high end powersupply.

The dual GPU GTX 690 had a similar RRP on launch ($1000), to the RTX 2080 TI which shows you how much more profitable the graphics card market has become for Nvidia.

Considering the work going into MCM I don't think we'll see any dual chip cards until NVIDIA's Hopper MCM architecture is ready in 2 years.

ahttps://en.m.wikipedia.org/wiki/Hopper_(microarchitecture)

Hopper looks like it's going to be a banger of generation if it's works out - major change in Graphics Card architecture and built on 5nm - we'll probably also have GDDR7 memory by then and PCIE5.

And if MCM does to GPUs what it did to CPUs were in for a treat that I can't wait for. If not surprised if we get gaming GPUs with 10,000+ cores

g67575 · 25 Jul 2020 at 23:35

Unless they are planning Ampere GPUs for 2020, and hopper MCM GPUs for 2021 (don't see why they couldn't be 7/8nm instead). Who knows when 5nm will be ready for mass production, given the problems Intel had with 10nm (and still no sign of 5nm CPUs, with only 7nm planned by AMD or Intel).

Grim5 · 25 Jul 2020 at 23:46

g67575 said:
Unless they are planning Ampere GPUs for 2020, and hopper MCM GPUs for 2021 (don't see why they couldn't be 7/8nm instead). Who knows when 5nm will be ready for mass production, given the problems Intel had with 10nm (and still no sign of 5nm CPUs, with only 7nm planned by AMD or Intel).

The most up to date leaks we have suggests this

RTX3000 and 3000ti cards Q3 2020 (Samsung 8nm/TSMC 7nm)

RTX3000 Super card's Q3 2021 (TSMC 7nm+)

Hopper MCM Card's Q2/Q3 2022 (TSMC 5nm)

TSMC 5nm is ready for mass production of small processors, right now they are producing millions of Apple A14 chips on 5nm. Desktop large CPUs will come next year from AMD and large graphics cards will follow that.

g67575 · 26 Jul 2020 at 00:00

Meh, might not matter if the transistor/ logic density is high enough on 7/8nm. Sometimes, the fab. process is just used to market the latest tech. Also, transistor density can be very similar on different nodes, like 10nm / 7nm. 10nm is a big set up from 14nm, which could be why Intel has struggled so much to produce 10nm desktop CPUs.

Here's some transistor density figures I found on the rather useful wikichip website, for nodes that have been produced (not estimates):

14nm - 37.22 million transistors per square mm
7nm - 95.08 million transistors per square mm (2019)
10nm - 100.76 million transistors per square mm (just 52.51 million in 2017)
6nm - 112.79 million transistors per square mm
5nm - 126.89 million transistors per square mm

Links :
https://en.wikichip.org/w/images/thumb/e/eb/5nm_densities.svg/600px-5nm_densities.svg.png

https://en.wikichip.org/w/images/thumb/5/5c/7nm_densities.svg/700px-7nm_densities.svg.png

So Samsung (so far) is furthest ahead with their 5nm fab. process.

7nm isn't looking that great really, but Intel is estimating much greater densities in the future, which I think look too optimistic.

I know achieving high clock speeds can be a problem on new nodes, so I don't know how this will translate to desktop CPUs and GPUs (if at all)

Phixsator · 26 Jul 2020 at 08:42

Grim5 said:
The most up to date leaks we have suggests this

RTX3000 and 3000ti cards Q3 2020 (Samsung 8nm/TSMC 7nm)

RTX3000 Super card's Q3 2021 (TSMC 7nm+)

Hopper MCM Card's Q2/Q3 2022 (TSMC 5nm)

TSMC 5nm is ready for mass production of small processors, right now they are producing millions of Apple A14 chips on 5nm. Desktop large CPUs will come next year from AMD and large graphics cards will follow that.

What makes you think that TSMC will be handing off 7nm+/5nm Fab capacity to a company who has publicly blamed them in the past and tried to bully them into lower prices when TSMC already has a profitable and solid relationship with its competitor who btw is supplying the server market, enthusiast market, mainstream market, laptop market, and the living room, in other words, could eat up all the fab capacity TSMC is willing to give them. The only reason would be that the other big players have moved on IMHO.

BigBANGtheory · 26 Jul 2020 at 08:45

Phixsator said:
What makes you think that TSMC will be handing off 7nm+/5nm Fab capacity to a company who has publicly blamed them in the past and tried to bully them into lower prices when TSMC already has a profitable and solid relationship with its competitor who btw is supplying the server market, enthusiast market, mainstream market, laptop market, and the living room, in other words, could eat up all the fab capacity TSMC is willing to give them. The only reason would be that the other big players have moved on IMHO.

Money

Phixsator · 26 Jul 2020 at 08:55

BigBANGtheory said:
Money

There are no shortcomings of tech companies who want to utilize TSMC's latest and greatest, so no not necessarily. Nvidia has a reputation for being a horrible company to work with. So if the money is secured from other companies who are less volatile and who can gobble up almost your entire fab capacity if you allow them too then why take money from an unpredictable source? simple you won't unless it was your last option.

BigBANGtheory · 26 Jul 2020 at 09:05

What tends to happen in big companies is people change around so lets say they had a big falling out which we don't really know and infact they are still building silicon for Nvidia... then in 2-3yrs time that all gets forgotten real quick in the pursuit of making $$$. Also aren't TSMC looking to build more factories in Japan and the US because of the Chinese threat to supply lines? That basically means more capacity at some point. Its unusual for two companies to fallout so badly they never ever do business again and if they did what you'd see is one party pulling out forcibly with legal challenges to contract etc... I'm seeing a lot of noise in the press but not those realities play out

Phixsator · 26 Jul 2020 at 11:09

BigBANGtheory said:
What tends to happen in big companies is people change around so lets say they had a big falling out which we don't really know and infact they are still building silicon for Nvidia... then in 2-3yrs time that all gets forgotten real quick in the pursuit of making $$$. Also aren't TSMC looking to build more factories in Japan and the US because of the Chinese threat to supply lines? That basically means more capacity at some point. Its unusual for two companies to fallout so badly they never ever do business again and if they did what you'd see is one party pulling out forcibly with legal challenges to contract etc... I'm seeing a lot of noise in the press but not those realities play out

Ohh I'm certain Nvidia has business with TSMC, I just don't see them(Nvidia) getting priority on the latest and greatest TSMC nodes. Sure it may change if Nvidia decides to pony up some serious money but then there are other things to consider such as current contracts which cannot be voided which in my mind would be a factor in Nvidia's next silicon design as it makes little sense to design something for a node you don't have access too or very little access to. I admit this is just speculation of course on my part, I could be wrong, but more and more companies and people are complaining about Nvidia's way of business these days. Would make sense to be skeptical of them if they(Nvidia) wants to do serious business with you(TSMC) and if more stable options are available from a source that cannot afford to act up, aka AMD due to MS and Sony, then why not go with the safer option.

g67575 · 26 Jul 2020 at 14:42

Pretty sure Nvidia will use 7nm for ampere GPUs, including a hypothetical high tdp dual GPU. The reasons are, the transistor density is decent / good enough and it's been used in CPUs + GPUs before.

8nm has a significantly lower density (apparently, just 61.2 MTr/mm² in 2018) so far, so it doesn't make sense to use it. Also, as far as I know, it hasn't been used or tested in many chips yet.

6/5nm are still too new (first nodes produced in 2020) and are probably only suited to mobile chips for the next year. Also, the 5nm transistor density is only about 25% higher than 7nm so far, and Intel estimates that ~200 MTr/mm^2 is possible in the future with 7nm.

Also with the 7nm fab process, companies can use either TSMC or Samsung.

I doubt it matters much, because the transistor density is similar between them (91.2 vs 95.08 MTr/mm^2).

Phixsator · 26 Jul 2020 at 16:50

..just pause for a second.... go 80 odd years back and the first transistor was the size of a hand. Now it's possible to stuff 100-130 million of them in 1 mm2? pretty insane. Yet all we do here is bicker about GPU performance

. Perhaps it's time to appreciate what is possible right now no matter if you swing green, red, blue, or something else.

g67575 · 26 Jul 2020 at 17:47

I agree, what tech. companies can build with 7/10nm is already an impressive improvement over say 14nm.

My dad still has a 1st gen i7 (45nm), which apparently has 3.3 million transistors per square millimetre. So, about 33 times this amount of transistors can be crammed into the same space with 10nm Intel laptop CPUs, probably desktop chips next year too.

Sometimes you gotta swing both ways

BigBANGtheory · 27 Jul 2020 at 08:32

Phixsator said:
..just pause for a second.... go 80 odd years back and the first transistor was the size of a hand. Now it's possible to stuff 100-130 million of them in 1 mm2? pretty insane. Yet all we do here is bicker about GPU performance . Perhaps it's time to appreciate what is possible right now no matter if you swing green, red, blue, or something else.

Ok I have completed my 1min reflection can we get back to price & performance

CuriousTomCat · 27 Jul 2020 at 08:41

If smaller nm means lower voltages for the GPU and lower voltages for the CPU, why are voltages and power requirements going up?

Nvidia's thicker wires, Intel's decision to abandon 3v and 5v, even power to the CPU has gone from 4 pin to 8 pin :confused:

Joxeon · 27 Jul 2020 at 09:32

Grim5 said:
Considering the work going into MCM I don't think we'll see any dual chip cards until NVIDIA's Hopper MCM architecture is ready in 2 years.

And if MCM does to GPUs what it did to CPUs were in for a treat that I can't wait for. If not surprised if we get gaming GPUs with 10,000+ cores

Will probably cost £10,000 by then too.

gpuerrilla · 27 Jul 2020 at 09:39

CuriousTomCat said:
If smaller nm means lower voltages for the GPU and lower voltages for the CPU, why are voltages and power requirements going up?

Nvidia's thicker wires, Intel's decision to abandon 3v and 5v, even power to the CPU has gone from 4 pin to 8 pin

It generally does but if you keep adding transistors or ramping up the clockspeed etc. not to mention the cost investment for the shrink - its something not worth rushing into. I feel its poor decisions lately rather than anything else. Might be a good thing that has the competition level out so hopefully we can force a price war southwards as theres nothing between them to dictate one is better than the other.

melmac · 27 Jul 2020 at 14:19

Phixsator said:
What makes you think that TSMC will be handing off 7nm+/5nm Fab capacity to a company who has publicly blamed them in the past and tried to bully them into lower prices when TSMC already has a profitable and solid relationship with its competitor who btw is supplying the server market, enthusiast market, mainstream market, laptop market, and the living room, in other words, could eat up all the fab capacity TSMC is willing to give them. The only reason would be that the other big players have moved on IMHO.

TSMC have already confirmed that they have pre booking from Nvidia for their 5nm Hopper cards. It was reported a few months ago.

And for companies it's all about business first. You see it all the time multinationals suing each other one on hand while partnering with each other on some other venture.

g67575 · 27 Jul 2020 at 17:20

The TSMC N5P 5nm node (could be upto 171.3 million transistors per mm²) is gonna be used in mobile devices this year. Perhaps it will be ready for desktop CPUs and GPUs (like Hopper) by 2h 2021? If not, don't see why a Hopper GPU couldn't be built with 7nm by next year. Especially if the design is basically just dual ampere GPUs, rather than a completely new design.

link:
https://www.techpowerup.com/264994/...transistor-density-gain-over-current-7nm-node

The N5P 5nm node could also be used in socket AM5 / Zen 4 CPUs, which should help AMD to compete with Intel's desktop 10nm CPUs, planned for release in 2021.

Im a bit surprised to see how quickly the 5nm fab. process is being developed... Maybe Samsung and TSMC are gonna focus their development on 5nm, and just leave 7nm?

Grim5 · 27 Jul 2020 at 20:01

g67575 said:
The TSMC N5P 5nm node (could be upto 171.3 million transistors per mm²) is gonna be used in mobile devices this year. Perhaps it will be ready for desktop CPUs and GPUs (like Hopper) by 2h 2021? If not, don't see why a Hopper GPU couldn't be built with 7nm by next year. Especially if the design is basically just dual ampere GPUs, rather than a completely new design.

link:
https://www.techpowerup.com/264994/...transistor-density-gain-over-current-7nm-node

The N5P 5nm node could also be used in socket AM5 / Zen 4 CPUs, which should help AMD to compete with Intel's desktop 10nm CPUs, planned for release in 2021.

Im a bit surprised to see how quickly the 5nm fab. process is being developed... Maybe Samsung and TSMC are gonna focus their development on 5nm, and just leave 7nm?

It's a significant change

Trying to say MCM is like sticking two Ampere cards together is like saying Ryzen is just two FX Bulldozer CPUs stick together.

MCM is not a dual chip gpu like the old days that utilised SLI, regardless of the number of chip modules it will be recognised by windows as a single GPU and requires a new high speed link similiar to infinity fabric between the modules. I also doubt they will just take two monolithic dies and stick on there, it's more likely to be several hardware accelerator units, eg a module that just does BVH intersections etc