It looks like AMD (Lisa Su) are prepared to lie low until there is a chip shortage again.
Until then it may not make business sense for AMD to release a big Navi card for desktops
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
RTX 5xxx / RX 8xxx speculation
- Thread starter Nebell
- Start date
It looks like AMD (Lisa Su) are prepared to lie low until there is a chip shortage again.
Until then it may not make business sense for AMD to release a big Navi card for desktops
Let us wait & see if AMD give any roadmap update in GamesCom ...
Tamlin_WSGF
2[H]4U
- Joined
- Aug 1, 2006
- Messages
- 3,123
I havent seen this anywhere on the internet besides on the youtube channel. From an AMD employee:
View: https://www.youtube.com/watch?v=2XYbtfns1BU
View: https://www.youtube.com/watch?v=2XYbtfns1BU
DooKey
[H]F Junkie
- Joined
- Apr 25, 2001
- Messages
- 13,683
This only matters if we know AMDs definition of high-end market.
Tamlin_WSGF
2[H]4U
- Joined
- Aug 1, 2006
- Messages
- 3,123
You mean by his definitition. This is of the few direct quotes about this from AMD employees. Appreciate I took the time to share it when I spotted it just now.
After all, its this is a speculation thread ...
It was not directly stated, but it was implied that, if there is no demand, due to crypto-shortage/AI etc. then there is no business case for AMD to make a "big navi" style chip.
Imo, they could still release a nvidia competitor, 6 months to 1 year behind, at a "cheaper" price.
ChronoDetector
2[H]4U
- Joined
- Apr 1, 2008
- Messages
- 2,784
I don't really want to know how much the supposed successor to the 4090 will cost, considering how much PC parts have increased over the years, I stupidly paid more than AUD $3K for my 4090 which is like more than $1900 USD.
Though I wonder if that 512 bit memory bus rumour is true for the 5090 or whatever it's going to be called will most likely increase the card's price, and we all know that NVIDIA loves to charge an arm and a leg for their flagship card. Can't blame them really because their competitor's can't really come up with a product that will match them in performance and they know that suckers will still purchase one regardless if they want the best.
Though I wonder if that 512 bit memory bus rumour is true for the 5090 or whatever it's going to be called will most likely increase the card's price, and we all know that NVIDIA loves to charge an arm and a leg for their flagship card. Can't blame them really because their competitor's can't really come up with a product that will match them in performance and they know that suckers will still purchase one regardless if they want the best.
Armenius
Extremely [H]
- Joined
- Jan 28, 2014
- Messages
- 44,700
The cost per wafer of TSMC 3nm is what will drive the cost, and people won't be happy unless they put all the blame on NVIDIA for the price of consumer cards.
I expect nvidia to revert back to a bimodal pricing strategy.
5080 & above (& probably the 5070 ti too) would be on 3nm & priced high.
5070 & below would be on 4/5nm & priced competitively vs Intel/AMD products in the same range.
I wish they would - then there's more used 4090s on the market - that would mean more supply - so, even the used price would drop a bit. Everything helps. I'd like to get a 4090 but not at the current 'new' price. Ppl who buy new 4090s - will buy them at (I'd like to say any) a high price and are more likely to get a 5090 - some gamers want whatever the best/newest (gen) is.
But, you are right - the 4090 is probably good enough at 4K - but, many ppl do what I said above.
Can I ask 'where are you?' One of the advantages or benefits of getting an Asus or MSI gpu here (Canada) is transferable warranty - so, at least, you can send a 2nd hand card in - just need serial # - and the card/RMA should be processed without a hassle. The question is, how does the RMA go.
There's 'RMA nightmare stories' for Asus products here, too, though - although, usually not too many for the gpus - I think. I am fairly sure that when you rma an Asus mobo - you get a refurbished/used one back. Does that happen with other brands? Probably? Maybe? I dunno. There's a lot of brands that have 'so-so' or suspect RMA / CS nowadays - lots of ppl are dissatisfied, frustrated or angry with their RMA process - across brands and tech companies.
So far, I've read horror stories about - Zotac, Asus, Sapphire, Powercolor, Asrock.... - to do with RMA and gpus.
The 'mixed' reports (good and bad) - MSI, Gigabyte (many bad), XFX (trend has been positive reviews)
This is the Canadian market and my readings are from forums/reddit sites - so, YMMV in your locale.
Cost per wafer might be up, but if you can cut more GPU's out of that wafer, the true cost increase is often times a wash or even reduced. By cost i'm referring to nVidia's cost. Not prices they sell said GPU's at.
Depend on where on the product stack, a 4090 do cost a fortune to make, a 16gb 4060ti current price of $500 is not driven by the Nvidia special 5n cost.
Not that we do know those cost for those special high volume clients deals but even if Nvidia 5 cost them 75% more as what AMD was paying for TSMC 7, navi 22 6700xt was a 335mm die, 192 bits bus, pci 4.0x16, that launched at $480 USD in peak bubble high pricing of March 2021.
188mm and the better yield that come from smaller die could even be cheaper than the 335mm tsmc 7 was and they save on pci express x8, 128 bits, lower power card, the 4080-4070s-4060ti have probably a lot of room price wise, Nvidia current adjusted gross margin above 70% are more those of a software company than an hardware one, obviously booster by Hooper a lot, but still lot of cost are from high margin of a duopoly.
TSMC 3 is rumoured to only be 25% more expensive than TMSC 5 was:
https://www.techspot.com/news/97269-tsmc-may-cut-3nm-wafer-prices-entice-amd.html
Ampere 3070 class card were made on 392mm die, 3080 on 628mm die.
Would 330mm-420mm if they go for a big performance boost next generation cost that much more than those big samsung 8 chips ?
Im in the US. I’ve been burned by asus motherboard rma before. But give the price of these gpus and Microcenter’s reputation on the warranty I just keep buying the extended warranty. If something happens I can drive 30 mins to Microcenter and they will give me whatever I paid for it back to me and I can walk out with a new gpu.
That interesting and maybe in a recent future it will be more regular customer items will stop to be on bleeding edge nodes that will be kept for current cancer, new material research and sellings ads on social media post. This is already a bit the case if the 5xxx series launch in 2025 while Hopper next be in 2024 and if they continue with launch the 5080-5090 first and many months after go down to the 5050-5060 the 3nm will be "old" by then.
Because of how important the laptop world will be and how strong the competition in it will also be, not just in term of raw effort but because of AMD-Intel ability to package the CPU-GPU offer and how big node gains must be in the Laptop world were 100% of the efficacy gains translate in performance gain not just in part or in cheaper cooling.
It is to be noted that new nodes tend to be cheaper or the same price by transistor, more and more it is just same price and not cheaper but it seem to be the same for N3
N3 is rumored to have 1.6 time the logic density than n5, how much money do you save with a bigger N5 cheap over the smaller N3 cheap (and possibly better yield chips because it is a smaller one one it get mature) ?
I imagine that why the 4060 went with a tiny N5 of 160mm instead of just making a refresh 276mm Samsung 8 3060 with better clocks-memory called 4060, because it is cheaper to do.
Last edited:
Axman
VP of Extreme Liberty
- Joined
- Jul 13, 2005
- Messages
- 18,526
And which generation he's referring to. I have no doubts that AMD is working on multi-GPU-die chiplet architectures, but I would be surprised if they have one in the works for this generation.
I mean, it would be nice, I'm just not expecting AMD to deliver. Both AMD and Nvidia are working on this, they have to, since the reticule size of these top-end nodes is too small to have monolithic GPU dies remain competitive.
MistaSparkul
2[H]4U
- Joined
- Jul 5, 2012
- Messages
- 4,025
I believe TSMC's 2nm node has something like a 429mm2 limit which would definitely hinder anymore monolithic designs from Nvidia given that the last 3 flagship dies from team green have been 609mm2, 628mm2, and a crazy 754mm2 on the 2080 Ti. Going to be really really interesting to see what happens beyond RTX 50 series which we know is going to be one last hurrah for monolithic on TSMC 3nm.
Do we have a source for the TSMC 2 node size limit ?
I cannot find aynthing on google, gpt, etc.. about it
MistaSparkul
2[H]4U
- Joined
- Jul 5, 2012
- Messages
- 4,025
Well you won't find anything directly about TSMC 2mn node size limit but you can sort of put 2 and 2 together by these articles:
https://wccftech.com/tsmc-says-2nm-...ed-high-na-machines-will-be-acquired-in-2024/
https://www.reuters.com/technology/...dvanced-asml-chipmaking-tool-2024-2022-06-16/
https://en.wikichip.org/wiki/mask#Reticle_limit
Basically 2nm node should be switching to High-NA EUV, and that is the reason for the reticle limit of 429mm2.
"The next phase in chipmaking will see manufacturers switch to machines with larger lenses. These are dubbed high NA (numeric aperture), and Dr. Mii outlined that his company will receive them in 2024. From this, it appears that TSMC will use these machines for making chips with its 2-nm manufacturing process since the executive also highlighted that this technology will enter mass production in 2025."
"Current i193 and EUV lithography steppers have a maximum field size of 26 mm by 33 mm or 858 mm². In future High-NA EUV lithography steppers the reticle limit will be halved to 26 mm by 16,5 mm or 429 mm² due to the use of an amorphous lens array."
Thanks
Current i193 and EUV lithography steppers have a maximum field size of 26 mm by 33 mm or 858 mm². In future High-NA EUV lithography steppers the reticle limit will be halved to 26 mm by 16,5 mm or 429 mm² due to the use of an amorphous lens array.
Maybe not a big deal for the gaming gpus, but for the Hopper class above 800mm they must have the chiplet working by then, HBM-GH 200 Hopper grace tech being maybe already close to be ready already to group a bunch of them like Apple do with M silicon.
The current sized 4080 already fit well at 379, they can triple the memory bandwith and have 2 generation of node gains and some architecture gains by then
MistaSparkul
2[H]4U
- Joined
- Jul 5, 2012
- Messages
- 4,025
4080 has a decent uplift over the 3090 yes so I'm sure they can come up with something. Just that if it's still monolithtic then we probably cannot expect a crazy 70% uplift again.
Not at the xx90 class level for which it could be an issue. But for anything below, the die are already smaller than that future limit.
But like you said with the 4080, The 3090 had a 66% bigger die with a 50% bigger memory bus, making gaining from it quite the challenge, yet the 4080 is about 30% faster
The 6080 could still be a bit bigger than the 5080 with better memory bandwith (that give the option to us less size of the die for cache if needed), maybe they can still get away with monolithic the next 2 gen for almost all the skus, apparently 2NM is expected to deliver 12% performance boost with 25% less power and 5% smaller die, if you go with similar power-size maybe you get close to a "free" 20% performance boost without any improvement to your gpu and can still get the nice 30% gen on gen boost.
The 4080 delivered a giant 45%, but that type of boost is only needed if you want to increase price or aim to sell to the latest gen buyer which is less and less the case.
MistaSparkul
2[H]4U
- Joined
- Jul 5, 2012
- Messages
- 4,025
Nvidia has a massive RND budget so I have no doubts they will be able to work some magic. Maxwell was made on the same TSMC 28nm node and still provided huge uplifts. As I said, it's going to be some interesting times ahead
Last edited:
Ampere (RTX 3000) was a really good architecture held back by the Samsung 8N process. The reason we saw such an amazing uplift going from the 3000 to 4000 GPUs was mainly due to the process improvement. It allowed Nvidia to push their products much further without incurring the power penalty imposed by Samsung 8N.
Unfortunately, I doubt we will see such an uplift again with the 5000 GPUs.
Unfortunately, I doubt we will see such an uplift again with the 5000 GPUs.
What does he mean by they are currently validating next gen products?
Does that mean we can see RDNA 4 desktop gaming GPUs in H2 2024 or is he referring to pro/compute cards ??
Tamlin_WSGF
2[H]4U
- Joined
- Aug 1, 2006
- Messages
- 3,123
I have no idea. Was written 8 of August, so cannot be anything from current gen (7800XT and 7700XT). For the pro and compute cards, newest announced card is MI300X (announced in June I think), which is a GPU with 8 GPU cores, unlike the MI300 non-X, which have 3 GPU cores and a CPU core. He is talking about highend cards in a video about a consumer card, so it could be consumer cards for prosumer ... Your speculation is as good as mine.
Nobody knows what a highend card is anymore either, after the new pricing models. Price brackets determined this before to a large degree, not performance alone. Some might say 4080 is highend, while others might say that only 4090 is a highend card. Is the 4070 TI (former 4080 12GB) a midrange card now?
Armenius
Extremely [H]
- Joined
- Jan 28, 2014
- Messages
- 44,700
We'll see. N3P is 1.56x denser compared to N4, meaning they could fit up to 130 billion transistors on a similar 609mm² die that is used in the 4090.
NightReaver
Supreme [H]ardness
- Joined
- Apr 20, 2017
- Messages
- 4,144
Only do it if you want it as a hobby.
I sort of do. I'm on the PC every day. I enjoy tinkering. I'm up to four motorcycles now and have sort of peaked there.
NightReaver
Supreme [H]ardness
- Joined
- Apr 20, 2017
- Messages
- 4,144
Fair enough. Enjoy
See ya over in the watercooling sub forum.That's one input, but Nvidia's margins say this is mostly on Nvidia. Frankly, that's fine. They're a publicly traded company, it's their job to charge what the market will bear. It's up to consumers to tell them the price it too high and so far we haven't.
MistaSparkul
2[H]4U
- Joined
- Jul 5, 2012
- Messages
- 4,025
1.56x is definitely something, but wasn't going from Samsung 8nm to TSMC 4nm something like 2.8x denser? That's a much bigger jump.
MistaSparkul
2[H]4U
- Joined
- Jul 5, 2012
- Messages
- 4,025
I think there's still some fuel left in the tank. The 4090 got a 0% increase in memory bandwidth over the 3090 Ti so who knows maybe GDDR7 with a huge uplift in memory bandwidth can do a good amount of the heavy lifting for performance gains next gen. Let's not forget that Nvidia also managed to pull some decent gains going from Kepler to Maxwell which was made on the same TSMC 28nm node. So if they could achieve what they achieved with Maxwell which had no node advantage over Kepler, then I'm sure going to 3nm they definitely have some room to work with.
If techpower up density estimate are right ( i am not sure about logic density, vs cache, etc...)
Ampere: 45.1M / mm²
TSMC 5 used in rdna 3: GCD 150.2M / mm²
TSMC 6 used in rdna 3: MCD Density: 54.64M / mm²
Lovelace: 125.3M / mm² (2.5 jump from ampere, but significantly lower than a RDNA 3 GCD)
GCD having such an high figure could be because of having an higher proportion of it that is "logic", so I am not sure you can have a pure density numbers they usually split by the nature-usage of them it seem, they tend to use a rough 50% logic, 30%sram, 20% analogue average.
Lovelace spend a giant budget on cache something that does not scale well with node progress, apparently: . The L2 cache blocks are at least 15% of the total die area, while the entire center portion of the die (L2 plus other logic) takes up 25% of the total.
https://www.tomshardware.com/news/why-nvidias-4080-4090-cost-so-damn-much
If they double memory bandwidth like early rumours seem to indicate maybe they scale the cache down and the transistor count will be in 2 direction, 1) better node, 2) more space dedicated to logic vs cache, so instead of a 1.56 you can end up closer to 1.7. And they can go bigger instead of going a bit smaller on the die size as well (628.4 down to 608 last time, they could afford it because how much better the node got but the 2080ti was 775mm, if they boost to 670mm, they could almost double the transistor count again).
They can make big die because the best one goes into some type of pro line, the good one goes to the high power gaming card and the lower bin one can go to the dual slot pro low power product line, the AD102 goes into
Perfect die: 18432 cores
L40/RTX6000: 18176 cores
4090: 16384 cores
RTX 5000: 12800 cores
you can score good yield with products that use only 70% of the chip and they sell them $4000 USD.... considering that model success I can see them go bigger die size for blackwell.
Ampere: 45.1M / mm²
TSMC 5 used in rdna 3: GCD 150.2M / mm²
TSMC 6 used in rdna 3: MCD Density: 54.64M / mm²
Lovelace: 125.3M / mm² (2.5 jump from ampere, but significantly lower than a RDNA 3 GCD)
GCD having such an high figure could be because of having an higher proportion of it that is "logic", so I am not sure you can have a pure density numbers they usually split by the nature-usage of them it seem, they tend to use a rough 50% logic, 30%sram, 20% analogue average.
Lovelace spend a giant budget on cache something that does not scale well with node progress, apparently: . The L2 cache blocks are at least 15% of the total die area, while the entire center portion of the die (L2 plus other logic) takes up 25% of the total.
https://www.tomshardware.com/news/why-nvidias-4080-4090-cost-so-damn-much
If they double memory bandwidth like early rumours seem to indicate maybe they scale the cache down and the transistor count will be in 2 direction, 1) better node, 2) more space dedicated to logic vs cache, so instead of a 1.56 you can end up closer to 1.7. And they can go bigger instead of going a bit smaller on the die size as well (628.4 down to 608 last time, they could afford it because how much better the node got but the 2080ti was 775mm, if they boost to 670mm, they could almost double the transistor count again).
They can make big die because the best one goes into some type of pro line, the good one goes to the high power gaming card and the lower bin one can go to the dual slot pro low power product line, the AD102 goes into
Perfect die: 18432 cores
L40/RTX6000: 18176 cores
4090: 16384 cores
RTX 5000: 12800 cores
you can score good yield with products that use only 70% of the chip and they sell them $4000 USD.... considering that model success I can see them go bigger die size for blackwell.
Last edited:
Navi 43 could just be navi 32 on a monolithic (4nm?) die
Navi 44 could be navi 33 on 4nm?
With move to monolithic dies & advanced processes, AMD should be able to boost the frequencies within the same power budget.
We could have:
5 8700 xt = 4070 ti for $500
5 8700 = 4070 for $400
5 8600xt = 4060 ti for $300
Hopefully all of the above in 1 year's time.
And by then the 7900xt could drop to $600 ?
& the 7900xtx to $750 ??
& the 7600 to $200 ???
& the 6400xt < $125
Navi 44 could be navi 33 on 4nm?
With move to monolithic dies & advanced processes, AMD should be able to boost the frequencies within the same power budget.
We could have:
Hopefully all of the above in 1 year's time.
And by then the 7900xt could drop to $600 ?
& the 7900xtx to $750 ??
& the 7600 to $200 ???
& the 6400xt < $125
Last edited:
Are those typo for 8700xt = 4070ti and so on ?
Will have to see how close the 7800xt for $500 will be to a 4070ti and where its msrp will be by rdna 4 release time to see how interesting this would be.
Last edited: