When will gpu prices return to normal.

[ARF]

The "Ti" versions are stop-gaps normally launched much later - RTX 3080 Ti in May 2021, while original RTX 3080 much earlier in 2020.

This lineup looks very strange, the performance estimates simply do not look right, and the lineup is heavily rearranged / rebalanced - the RTX 4090 will pull forward and the gap with the rest of the lineup will be huge.

RTX 4090 = RTX 3090 + 56% more shaders.
RTX 4080 = RTX 3080 +18% more shaders.
RTX 4070 = RTX 3070 + 21% more shaders.

While RTX 3090 was only ~10-15% faster on average than the RTX 3080.
RTX 3080 was 30% faster than RTX 3070.

If the shaders do not give higher IPC, then the RTX 4070 will remain about the same performance or slower than the September 2020 10 GB RTX 3080.

 

[N/A]

have no doubts that 4070 /80 are at least 50% faster. shaders IPC and clocks are not the same for ampere and ada, but that remains to be seen.
take 1070 / 970 for example. 15% more shaders, 40% clock speed. 61% faster, it's actually too perfect. but somehow the story repeats itself.

 

[ARF]

have no doubts that 4070 /80 are at least 50% faster. shaders IPC and clocks are not the same for ampere and ada, but that remains to be seen.
take 1070 / 970 for example. 15% more shaders, 40% clock speed. 61% faster, it's actually too perfect. but somehow the story repeats itself.

Yeah, the N4 process should be 50-60% faster at the same wattage (than Samsung 8N).
Also, the clocks must be much higher. 2.5 GHz? 2.8 GHz AD103?

The question is - are the shaders smaller with lower IPC or the same size but slightly faster?

 

[ModEl4]

The "Ti" versions are stop-gaps normally launched much later - RTX 3080 Ti in May 2021, while original RTX 3080 much earlier in 2020.

This lineup looks very strange, the performance estimates simply do not look right, and the lineup is heavily rearranged / rebalanced - the RTX 4090 will pull forward and the gap with the rest of the lineup will be huge.

RTX 4090 = RTX 3090 + 56% more shaders.
RTX 4080 = RTX 3080 +18% more shaders.
RTX 4070 = RTX 3070 + 21% more shaders.

While RTX 3090 was only ~10-15% faster on average than the RTX 3080.
RTX 3080 was 30% faster than RTX 3070.

If the shaders do not give higher IPC, then the RTX 4070 will remain about the same performance or slower than the September 2020 10 GB RTX 3080.

Usually yes, regarding Ti versions launch, but there are exceptions, for example 3060Ti launched 2 months and 1 week after 3090, 2080Ti launched at the same time with 2080 etc.
And after all, the model numbering could be different from the table you quoted.(We could have for example a RTX 4090 Ultra for a Full AD102 version, who knows?
Irrespective from model numbering which can change, there are many examples (but not always) when the most cut down part of a GPU die comes at an earlier stage of a GPU lifecycle (concurrent or with small difference like 3060Ti) and as the yields improve over the lifetime of the product, either we have a refresh or if not the manufacturer can limit just a little bit the availability of the most cut down part if it make sense based on yields.
Another reason for AD103 not to be excessively stressed out is potentially the SM count, full AD102 (192SM) will be served with 384bit bus & 24Gbps GDDR6X (possibly limited availability in order to be able to support many SKUs) and to have the same bandwidth per/SM with 21Gbps GDDR6X memory on a 256bit bus, AD103 must be 112SM instead of the more orthodoxal 128SM, if Ada design is bandwidth limited (I think it will be, just like Pascal) there is no reason to stress AD103 frequency too much since the gains will be limited anyway, but this is just my speculation, we will see.
On top of specs, additional info can occure if you see what historically Nvidia did in the past when it launched a new lineup and how they brought their top last gen performance level at what (lower) price points.
Try to imagine Nvidia's CEO on stage announcing the $500 Ada part, then correlate with corresponding ada model/specs based on your optic and what performance that part must at least reach in order to generate the minimum buzz and then extrapolate from there.
Without even taking account the specs, how likely is it according to your perception, the $499 part (whatever the name) to be less than 3080? (imo logically it will be at least matching 3080 12GB and at that point Navi 33 how much faster it will be if any, since everybody saying that it will be slower at 4K than 6900XT due to 128bit bus (probably it will have around 6800XT 4K performance?)

Yeah, the N4 process should be 50-60% faster at the same wattage (than Samsung 8N).
Also, the clocks must be much higher. 2.5 GHz? 2.8 GHz AD103?

The question is - are the shaders smaller with lower IPC or the same size but slightly faster?

N4 has 63% clock speed increase potential for logic vs N16 according to TSMC and N16 can hit very similar frequencies to Samsung's 8nm.
Logically there will be eventually OC ada models close to 3GHz if ada architecture is designed for high clocks throughput.

 

[SpittinFax]

The mid-high tier RTX 3060 and RTX 3060 Ti cards in Australia have stubbornly refused to drop in price since early May, unfortunately. Only the MSI RTX 3060 Ti Ventus 2X has seen a substantial price drop of AU$100, the rest have hardly moved.

I took some screenshots on 12th May of prices to keep track of them. Here's pricing for those cards on 12th May:

Spoiler: RTX 3060 - 12 May 2022

Spoiler: RTX 3060 Ti - 12 May 2022

Now the current pricing for comparison:

Spoiler: RTX 3060 - 26 June 2022

Spoiler: RTX 3060 Ti - 26 June 2022

Granted there's more stock, but most cards haven't seen any change and most prices still suck.

 

[ARF]

I don't know what decision the Nvidia's managers will make but I sure know that this doesn't seem to be the right one...

These know that the users have no choice - it is 50-50 AMD or nvidia, so no matter what they do or how bad their action is, the sales will go on, anyways. They have nothing to lose, so they have the freedom not to think in the right direction.

 

[ModEl4]

These know that the users have no choice - it is 50-50 AMD or nvidia, so no matter what they do or how bad their action is, the sales will go on, anyways. They have nothing to lose, so they have the freedom not to think in the right direction.

Maybe, we'll see.
The only reason to do it imo (still wrong decision) is if they saw according to their simulations that full Navi32 is ahead of AD103 and thought to compensate with a +5% performance boost pushing frequency up (and power consumption through the roof).
They could just take the 5% beating (like in 3070Ti/6800 case, since at 350W the performance/W comparison would be fine and the memory size the same unlike what we have now for 3070Ti/6800), inform partners to balance more their production towards OC models and selling just a little bit above SRP these OC models while entry models are just at SRP in order to save face lol (but it doesn't matter if the entry models don't sell much since partners will have already arrange their production accordingly) and maybe add a game bundle at a later date if they still need to push further. (But logically there will not be a need to push further or push any at all lol, the performance/W difference will be less than 10% and the memory size the same, right now 6800 is 35% more efficient vs 3070Ti and has double the memory and still Mindfactory starts at 699€ for a 3070Ti (with a KFA² entry or an INNO3D 3X lol) while 6800 starts at -50€ (649€) for a much more high quality Asrock Phantom Gaming model, so no need at all for Nvidia to worry for 5% performance loss...
I wonder have many people from the TPU audience that buys 4080 level GPUs, would be deterred from the increased TDP. (the below chart doesn't say much it's too generic)

 

[ARF]

AMD Radeon RX 6400 Grafikkarte Preisvergleich | Günstig bei idealo kaufen
AMD Radeon RX 6400 Specs | TechPowerUp GPU Database

Actual retail prices:
Radeon RX 6400 - from 171 euros (MSRP 159 dollars)
Radeon RX 6500 XT - from 170 euros (MSRP 199 dollars)
Radeon RX 6600 - from 298 euros (MSRP 329 dollars)
Radeon RX 6600 XT - from 398 euros (MSRP 379 dollars)
Radeon RX 6650 XT - from 405 euros (MSRP 399 dollars)
Radeon RX 6700 XT - from 509 euros (MSRP 479 dollars)
Radeon RX 6750 XT - from 600 euros (MSRP 549 dollars)
Radeon RX 6800 - from 680 euros (MSRP 579 dollars)
Radeon RX 6800 XT - from 800 euros (MSRP 649 dollars)
Radeon RX 6900 XT - from 900 euros (MSRP 999 dollars)
Radeon RX 6950 XT - from 1208 euros (MSRP 1099 dollars)

 

[ARF]

Very interesting: This is how much graphics cards should cost based on their performance alone - VideoCardz.com

 

[ModEl4]

Very interesting: This is how much graphics cards should cost based on their performance alone - VideoCardz.com

View attachment 253134

View attachment 253135

It doesn't take account the memory size at all, so it ends being a little unfair depending the comparison.
You could argue that the PCB components or the memory type/speed used is affecting performance so taking account the performance metric you are somehow including them also in a "fair" comparison regarding price/performance metric, but in some games and resolutions, memory size doesn't bring more performance, it just covers you for future games so you can keep your VGA more years, that's an important extra value relatively easier to quantify (in relation with other values like media engine, s/w etc) that's not taken into account at all (in their experiment with the logic they used a 16GB 6800XT/3070 should have cost the same as a 8GB 6800XT/3070.
Also while RTX 3050 is compared based on 4K difference, Navi24 models are compared based on 1080p difference which is unfair since both are 1080p level cards, the reason they doing it, is because memory size & bandwidth and infinity cache size absolutely kill the 4K performance of Navi24 models.
At least they could have used 1080p difference for Navi24/Navi23 (which is the more indicative difference based on what it targets) and QHD difference for RTX3050/3060 if not FHD, QHD it wouldn't change much for 3050 but at least it would "seem" a little bit more fair.

 

[Count von Schwalbe]

It doesn't take account the memory size at all, so it ends being a little unfair depending the comparison.
You could argue that the PCB components or the memory type/speed used is affecting performance so taking account the performance metric you are somehow including them also in a "fair" comparison regarding price/performance metric, but in some games and resolutions, memory size doesn't bring more performance, it just covers you for future games so you can keep your VGA more years, that's an important extra value relatively easier to quantify (in relation with other values like media engine, s/w etc) that's not taken into account at all (in their experiment with the logic they used a 16GB 6800XT/3070 should have cost the same as a 8GB 6800XT/3070.
Also while RTX 3050 is compared based on 4K difference, Navi24 models are compared based on 1080p difference which is unfair since both are 1080p level cards, the reason they doing it, is because memory size & bandwidth and infinity cache size absolutely kill the 4K performance of Navi24 models.
At least they could have used 1080p difference for Navi24/Navi23 (which is the more indicative difference based on what it targets) and QHD difference for RTX3050/3060 if not FHD, QHD it wouldn't change much for 3050 but at least it would "seem" a little bit more fair.

It would be more logical to use QHD across the board, as that is where AMD/Nvidia reach parity for equal performing parts. Using 4k seems strongly Nvidia biased to me, especially as 1080p is still the most common gaming resolution.

 

[ModEl4]

It would be more logical to use QHD across the board, as that is where AMD/Nvidia reach parity for equal performing parts. Using 4k seems strongly Nvidia biased to me, especially as 1080p is still the most common gaming resolution.

I don't know much about 3DCenter but they don't seem to me Nvidia biased, in their performance charts AMD's models enjoy slightly better % positioning in relation to Nvidia's "similar" models than in TPU % performance charts (for example RX6400 is above GTX1650 in FHD in their charts) i guess they just used 4K in order to be more "fair" to the higher end models?
It's the right thing to do also unless a model suffers from excessive performance loss in 4K like Navi24 in most games (or excessive performance loss in only a handful of games (1?) like GTX 1660 series strangely lol)
If you check performance deltas in older titles at max settings (or esport titles) or newer titles at minimum/medium settings and compare them with new titles at max settings, you will see that as you increase the effects implemented/texture size etc in an engine/game, so the performance delta increasing between cards.
So if you want to see what the performance differences will end up between 2 same gen/manufacturer models after 2-3 or more years, the current 4K indication is possibly a better example because on average due to the more effects implemented in future titles the performance delta will likely increase also (after all we are at cross-gen phase right now, so even more potential for the performance delta to be bigger in future titles imo) of course it's not accurate and can't give exactly what the performance delta will be after 2-3 years but closer in order to be preferable (i mean current 4K deltas will be closer to future QHD deltas than current QHD deltas) and anyway the current QHD deltas isn't dramatically different from current 4K deltas anyway, unless we are talking about specific models, usually Navi 23 and below due to bandwidth limitations & infinity cache size used or other models due to memory size etc, for those models maybe a specific exception, although it isn't out of the realm of possibility in 2 years from now, an unreal engine 5 game (or some other game that pushes the visual boundaries less well optimized than unreal engine's 5) can show performance problems even in QHD for Navi23 due to infinity cache size/bandwidth limitations.

 

[Count von Schwalbe]

I don't know much about 3DCenter but they don't seem to me Nvidia biased, in their performance charts AMD's models enjoy slightly better % positioning in relation to Nvidia's "similar" models than in TPU % performance charts (for example RX6400 is above GTX1650 in FHD in their charts) i guess they just used 4K in order to be more "fair" to the higher end models?
It's the right thing to do also unless a model suffers from excessive performance loss in 4K like Navi24 in most games (or excessive performance loss in only a handful of games (1?) like GTX 1660 series strangely lol)
If you check performance deltas in older titles at max settings (or esport titles) or newer titles at minimum/medium settings and compare them with new titles at max settings, you will see that as you increase the effects implemented/texture size etc in an engine/game, so the performance delta increasing between cards.
So if you want to see what the performance differences will end up between 2 same gen/manufacturer models after 2-3 or more years, the current 4K indication is possibly a better example because on average due to the more effects implemented in future titles the performance delta will likely increase also (after all we are at cross-gen phase right now, so even more potential for the performance delta to be bigger in future titles imo) of course it's not accurate and can't give exactly what the performance delta will be after 2-3 years but closer in order to be preferable (i mean current 4K deltas will be closer to future QHD deltas than current QHD deltas) and anyway the current QHD deltas isn't dramatically different from current 4K deltas anyway, unless we are talking about specific models, usually Navi 23 and below due to bandwidth limitations & infinity cache size used, for those models maybe a specific exception, although it isn't out of the realm of possibility in 2 years from now, an unreal engine 5 game (or some other game that pushes the visual boundaries less well optimized than unreal engine's 5) can show performance problems even in QHD for Navi23 due to infinity cache size/bandwidth limitations.

True, I can see your point. I was just referring to the fact that RDNA2 in general performs superior to Ampere at 1080p, roughly equivalent at QHD, and inferior at 4K.

 

[ARF]

Memory size?
If they do that, they should include everything else - die size, memory bus width, memory type, etc...

True, I can see your point. I was just referring to the fact that RDNA2 in general performs superior to Ampere at 1080p, roughly equivalent at QHD, and inferior at 4K.

Yeah, it's inferior at UHD because of older memory standard and lower memory throughput.
It is the most annoying "feature" of the RDNA 2-based cards - they lose competitiveness with higher resolutions.

I hope RDNA 3 will fix this and will do the opposite - become stronger at UHD, and ray-tracing.

 

[DemonicRyzen666]

Online retailers are still over priced even for used/refurbished, they're double the price of e-bay sellers. Sometimes you're better off going ino the brick and motar to find sales.

 

[ModEl4]

Memory size?
If they do that, they should include everything else - die size, memory bus width, memory type, etc...



Yeah, it's inferior at UHD because of older memory standard and lower memory throughput.
It is the most annoying "feature" of the RDNA 2-based cards - they lose competitiveness with higher resolutions.

I hope RDNA 3 will fix this and will do the opposite - become stronger at UHD, and ray-tracing.

Nothing bad about 3Dcenter's approach after all the part «based on their performance alone» says it all, it's clear what they did, they didn't misslead anyone, it's just that since memory is easier to quantify in relation to other differences like media engine s/w etc they could have gone a step further (die size-number of transistors preferably since the process difference will lead to even more virtual differences at what transistors alone will give, memory bus width, memory type are all affecting the final performance of each card anyway so taking account the performance in a degree includes the effect of all those metrics in the price/performance placement. In some cases depending the gpu/game/resolution, extra memory doesn't add anything in the performance, it just can give extra value to your card and you can keep it more years, that's why it would have been interesting to included it somehow in their conclusions.
AMD isn't inferior to Nvidia at 4K due to Nvidia's use of GDDR6X (newer memory standard), since GDDR6X gives little performance benefit for raster, just check 3070/3070Ti (although 3070 regarding raster is the most "memory bandwidth starved" model by far in Ampere lineup with 96Rops & 448GB/s, it's only 7% slower at 4K with -4% less shaders/TMUs etc. vs 3070Ti, for other Ampere models jumping to GDDR6X will give even less difference for raster specifically since ampere raster performance is not memory bandwidth limited like Pascal was for example) the reasons are mostly (like you mentioned) memory bandwidth (24Gbps GDDR6 would help a lot) and also infinity cache size related, RDNA3 according to leaks will fix at least the size of Infinity cache so it will certainly help there (a big cache can help with raytracing also but depends on the raytracing implementation, it must be designed to take advantage of the cache system)

 

[ppn]

The narrow bus of RTX 4080 and RX 7800 XT renders them equally worse at 4K than prev gen. As for 4070 160bit and 7700 128bit watch them jump off a hill.

 

[ARF]

The narrow bus of RTX 4080 and RX 7800 XT renders them equally worse at 4K than prev gen. As for 4070 160bit and 7700 128bit watch them jump off a hill.

That will be the second generation product lineup with limited memory throughput, so it is highly likely the issue has been addressed and resolved somehow.
We'll see soon, hopefully.

 

[gffermari]

I'm thinking of replacing my 2080Ti to 3080Ti/3090. The prices in the second hand market are tempting.
I believe the 4080 will be 5-10% faster than 3080Ti at best, will be available -practically- at Christmas the soonest and cost about 1000 euros, also practically.

So, a 3080Ti at that price or lower, doesn't seem a bad idea to me.

Any suggestions?

 

[ORLY]

do you guys think that gpu prices will be better by then?

The market works in a simple way pretty much: a seller increases prices up to the maximum level that a buyer is willing to pay.
It's like that for any product that is aimed to be bought by millions of people.

A manufacturer spends a specific amount of money to produce a product. Of course they want to make as much profit as possible.
But a regular buyer of regular hardware is not rich, so e.g. if NVIDIA spends let's say 150$ to manufacture one RTX 3060 videocard and then tries to sold it for 1500$ - 99,9% of such videocards will not be bought at all and all that money (150$ multiplied by the number of produced videocards) will be wasted because who needs that peformance for THAT price?!
Of course there are some additional "features" of that but the main idea is the same - a manufacturer produces goods to sell, and if they fail to sell - they lose money that they spent on production. Because NVIDIA, AMD, Intel don't need to keep any of the stuff they produced, they need to make money from selling it.

You know what I mean - the prices are being crazy only up to a point that there's no enough buyers willing to pay them. And who is willing to overpay now? That's correct - the miners.
While there's so many miners overpaying for videocards - the prices will be crazy. And they're willing to overpay until it's profitable.

 

[oxrufiioxo]

I'm thinking of replacing my 2080Ti to 3080Ti/3090. The prices in the second hand market are tempting.
I believe the 4080 will be 5-10% faster than 3080Ti at best, will be available -practically- at Christmas the soonest and cost about 1000 euros, also practically.

So, a 3080Ti at that price or lower, doesn't seem a bad idea to me.

Any suggestions?

I'd wait pretty sure whatever costs 800-1000 usd will be 40-60% faster and even if you decide to grab a 3080 ti betting there will be a fire sale on ebay whenever the 4000 series is announced... I have a 2080ti and a 3080ti the increase is nice but really only at 4k does it feel like a generational improvement.

I'd also want to see RDNA 3 before making a decision.... Ampere is old at this point a 3080 12G makes more sense if you can get it 200-ish cheaper than the ti

 

[80251]

I'm thinking of replacing my 2080Ti to 3080Ti/3090. The prices in the second hand market are tempting.
I believe the 4080 will be 5-10% faster than 3080Ti at best, will be available -practically- at Christmas the soonest and cost about 1000 euros, also practically.

So, a 3080Ti at that price or lower, doesn't seem a bad idea to me.

Any suggestions?

But will any of the 4080's actually be available? Or will the miners and scalpers be hoovering them up like dust on a dirty carpet?

 

[gffermari]

That's the problem. Even if the 4080 is a bargain(msrp/performance) against the 3080/3080Ti, it would be difficult to get one, either because of the availability or the scalping prices.
So, the good scenario would be a meh price for a 4080, about now...next year.

 

[P4-630]

When will gpu prices return to normal?

 

[ARF]

I'm thinking of replacing my 2080Ti to 3080Ti/3090. The prices in the second hand market are tempting.
I believe the 4080 will be 5-10% faster than 3080Ti at best, will be available -practically- at Christmas the soonest and cost about 1000 euros, also practically.

So, a 3080Ti at that price or lower, doesn't seem a bad idea to me.

Any suggestions?

No, there is simply no way the 4080 to be so slow.
Look at the general performance gaps - they are tiny.

NVIDIA GeForce RTX 3080 Specs | TechPowerUp GPU Database

I'm thinking of replacing my 2080Ti to 3080Ti/3090. The prices in the second hand market are tempting.
I believe the 4080 will be 5-10% faster than 3080Ti at best, will be available -practically- at Christmas the soonest and cost about 1000 euros, also practically.

So, a 3080Ti at that price or lower, doesn't seem a bad idea to me.

Any suggestions?

Suggestion - wait the next Black Friday.