Nvidia is lying to you

[Zodiark1593]

4 minutes ago, Vishera said:

Most PC gamers don't use ray tracing anyway.

Currently anyway. Though what will happen in the near future, who’s to say. 

[Sir Beregond]

2 hours ago, Zodiark1593 said:

Perhaps reality doesn’t exactly conform to what we want?

Nvidia and AMD have to sell and move inventory. They can't not do that, especially considering they are allocated x number of wafers from TSMC. Their options if cards don't sell are to drop prices to where they will sell, or try to weasel out from what they've agreed to buy from TSMC and in doing so risk losing preferred status with TSMC for future allocation.

 

If we aren't buying said inventory and stuff rots on shelves, something has to change. I know it looks like it can't but that's just economic reality that they will both have to face at some point. Right now they are still trying to ride high on market conditions that no longer exist to keep padding those high margins while making "Moore's law is dead, COVID, supply chain, costs, inflation, etc." excuses that don't line up with reality of their current sky high margins.

 

Everyone is sitting here like we can't do anything. We absolutely can. Don't buy these cards at current prices. Simple as that.

 

EDIT: That said, I seriously feel like if we continue to see cards rot on shelves and nothing changes pricing wise in the coming month or two, then Nvidia and AMD should be investigated for price fixing collusion and market manipulation to try to maintain artificially high prices as if it were still 2021. Remember, after all:

 

[JCHelios]

Quote

Unfortunately for Nvidia, the vast majority of our viewers have never once enabled ray tracing. ... So, at least for you guys, it might not be the trump card that they were hoping it would be.

"Never" in this poll should have been separated out into "Never, because I don't want to" and "Never, because my GPU doesn't support ray tracing." (@Cardoso there gets it!)

 

It's important because those two answers point in opposite directions: People who don't care about RT features can focus purely on rasterization, and would be just as happy with AMD cards. But to the extent that people are still using old pre-RT cards... which, as you pointed out by citing the Steam Hardware Survey, is substantially the case... then they could be champing at the bit to turn on RT just as soon as they upgrade, and would want a card that can do that well (i.e., Nvidia).

With that having been said, I certainly l agree with the larger point, which I take to be along the lines of "Hey, remember when the GTX 970 was $329? Wasn't that great?!" Heck yeah it was. Requiescat in pace, attractively-priced 70-class cards.

[Sir Beregond]

On 1/4/2023 at 8:15 PM, LinusTech said:

They won't. Gaming is STILL among the best bang for the buck forms of entertainment on the planet. 

 

Playing badminton costs $25/hr for court rental plus on average another 25 bucks an hour of shuttles. Even split 4 ways that's really expensive and it's nothing compared to many hobbies... 

 

Gaming on the other hand is pay once (or never with F2P) and enjoy forever. 

 

The ceiling may be higher yet... 

Apologies ahead of time for the wall of text.

 

One thing I'd like to see the mainstream tech-tubers cover more these days isn't just the branding comparisons (i.e. this gen's 70 card to last gen's 70 card). With Kepler something shifted and that was what die class was what cards. It was a massive price hike of traditional mid-range silicon that was $250 in Fermi that got a name change to "80-class" branding and with it, the 80-class pricing. In one gen, mid-range dies go from being $250 products to $500 products. Stuff shifted again with Turing, Ampere, and Lovelace. That is totally a point of consideration because if you take away the "Titan-class" marketing associated with 90-class cards, its clear that they are to the 30 and 40-series generations what the GTX 580 for $500 was to the 500-series generation. And we can pretty reliably look at the die code-names and see this trend:

 

I'll just use an example, let's take a look at 70-branded cards. (For those reading, I otherwise assume Linus knows how this works) If you aren't aware of how Nvidia die codenames work, let me break it down for you real quick:

 

GA102 (GA is the generation for Ampere. I like to think of it as "Geforce Ampere". Ada Lovelace is "AD", Pascal was "GP" etc. etc.) When it comes to the numbers the first two numbers relate to the generation and if there is a revision. For example the 400-series and 500-series were both Fermi products where the GTX 480 was GF100 and its refresh card the GTX 580 was GF110. The real number to pay attention to is the very last number, be it the 0, 2, 3, 4, 6, etc. The lower the number, the higher end the chip, for example the 4090 is AD102 and the next card down, the 4080 is AD103.

 

Here's a breakdown of how you got less for more over time in this class, similar to the x80 tier.
 

• GTX 470 (2010): 1.28GB GDDR5 320-bit bus, GF100 chip (cut down highest end Fermi die) - $349
• GTX 570 (2010): 1.28GB GDDR5 320-bit bus, GF110 chip (cut down highest end Fermi refresh die) - $349
• GTX 670 (2012): 2GB GDDR5 256-bit bus, GK104 chip (cut down mid-range Kepler die) - $400
• GTX 770 (2013): 2GB GDDR5 256-bit bus, GK104 chip (rebrand of the GTX 680) - $399
• GTX 970 (2014): 3.5+0.5GB GDDR5 224+32-bit bus, GM204 chip (cut down mid-range Maxwell die) - $329
• GTX 1070 (2016): 8GB GDDR5 256-bit bus, GP204 chip (cut down mid-range Pascal die) - $449 (Founder's Edition), $379+ (AIB models)
• GTX 1070 Ti (2017): 8GB GDDR5 256-bit bus, GP104 chip (functionally same mid-range Pascal die with only one less SM than the GTX 1080 vs 5 on the 1070 - Released to beat the Vega 56) - $449
• RTX 2070 (2018): 8GB GDDR6 256-bit bus, TU106 chip (Not even the mid-range die anymore, now shares same chip as xx60 class) - $599 (Founder's Edition), $499+ (AIB models)
• RTX 2070 Super (2019): 8GB GDDR6 256-bit bus, TU104 chip (Notice, this version has same mid-range Turing die as 2080 and 2080 Super now, only cut down vs those chips. Released to combat 5700XT) - $499+
• RTX 3070 (2020): 8GB GDDR6 256-bit bus, GA104 chip (mid-range Ampere die) - $499 (Founder Edition), $499+ (AIB models)
• RTX 3070 Ti (2021) 8GB GDDR6X 256-bit bus, GA104 chip (mid-range Ampere die) - $599
• RTX 4070 Ti (2023) 12GB GDDR6X 192-bit bus, AD104 chip (3rd tier Ada Lovelace die) - $799

So let's analyze. We see in the Fermi generation that the 70-class utilized the same die as the 80-class, only cut-down. Since in that generation, the 80-class card was the top flagship card (there were no 80 Ti's and Titans), we see you got the second best card on the top end die for around $350.

 

Then something changed with Kepler. The GTX 670 still shared the same die class as the GTX 680, but noticed it went from a 100 die to a 104. They dropped a die class for the 70 (and 80) card, die size (meaning cheaper to make, larger yields per wafer, higher margin), and yet increased the price to $400. And if you were around back then, you know Nvidia pulled one over on everyone because the mainstream tech press loved these cards even though they were barely 30% better than Fermi (and at the time that was considered a very poor generational performance upgrade). Its clear what happened here. The 7970 was beatable by Nvidia with their mid-range silicon, so they rebranded, kept the high-end pricing and we didn't see any actual high-end Kepler until GK110 products released. Now let's talk about those.

 

Once Nvidia was ready to show GK110 products we first got the GTX Titan. It was an at the time ludicrous $1000 and not at all worth the cost. And believe me, at that time it was marketed for gaming. Yet even at $1000, it was not the fully unlocked chip. But notice how in one generation the flagship went from $500 to $1000. The GTX 780 then released utilizing the same die but it was heavily cut down and the 70-class card was simply a rebranded GTX 680 still on GK104. Nvidia then got shocked and upset by the $550 R9 290X and then released full GK110 in the 780 Ti.

 

Then with both Maxwell and Pascal we got a similar story, the 80-class and 70-class cards were on the same mid-range silicon while the x80 Ti cards were on traditional larger/flagship silicon for more money and Titans did the same still.

 

Then let's talk about Turing. Minuscule generational on generational performance increases from Pascal, yet came with a huge price increase and for the first time we see the 70-class card drop down to the third tier chip (TU106), the same silicon for 60-class cards since the Kepler generation. Yet it was also the at the time, highest price for a 70-class card ever. It was corrected with the 2070 Super.

 

Ampere sort of resets some things but also introduced some other shenanigans. The 80-class card could not be on 104 silicon because Samsung's 8nm process was too old and Nvidia was tapping it for all it could give, so it went back to the 3070 (and later 3070 Ti). But higher up in the stack we see something strange and that's the introduction of 90-class cards. Linus himself said it during the 3090 review: When you look at the performance vs the Titan RTX, it's clear the 3090 is not actually a Titan. In fact I'd posit that it was a rebrand and replacement for the 2080 Ti with a $300 price hike. Everyone would consider the 2080 Ti the flagship gaming card of the Turing generation. Absolutely nobody would call the 3080 Ti the flagship gaming card of the Ampere generation. I rest my case.

 

And now the Ada Lovelace generation. We once again see a Turing like drop in die-class for the 70-class product, yet we also see a massive price hike.

 

I say all this to provide a bit of a history lesson and to show Nvidia has been screwing us gamers for a long time and we've been buying them up and letting them get away with it. It's good to see many taking a stand now, but I fear we've let the cat out of the bag 11 years ago, and what we've seen is the natural evolution of that. Absolutely no one in mainstream tech press called out Nvidia for what they did with Kepler in the GTX 670/680 and in fact gave them glowing reviews. And thus we started the march towards Nvidia charging more and giving less.

 

Moral of the story, you give Nvidia an inch, they will gladly take a mile and then some over time.

 

Thanks for listening to my TedTalk.

[LinusTech]

1 hour ago, Sir Beregond said:

Apologies ahead of time for the wall of text.

 

One thing I'd like to see the mainstream tech-tubers cover more these days isn't just the branding comparisons (i.e. this gen's 70 card to last gen's 70 card). With Kepler something shifted and that was what die class was what cards. It was a massive price hike of traditional mid-range silicon that was $250 in Fermi that got a name change to "80-class" and with it, the 80-class branding. Stuff shifted again with Turing, Ampere, and Lovelace. That is totally a point of consideration because if you take away the "Titan-class" marketing associated with 90-class cards, its clear that they are to the 30 and 40-series generations what the GTX 580 for $500 was to the 500-series generation. And we can pretty reliably look at the die code-names and see this trend:

 

I'll just use an example, let's take a look at 70-branded cards. If you aren't aware of how Nvidia die codenames work, let me break it down for you real quick:

 

GA102 (GA is the generation for Ampere. I like to think of it as "Geforce Ampere". Ada Lovelace is "AD", Pascal was "GP" etc. etc.) When it comes to the numbers the first two numbers relate to the generation and if there is a revision. For example the 400-series and 500-series were both Fermi products where the GTX 480 was GF100 and its refresh card the GTX 580 was GF110. The real number to pay attention to is the very last number, be it the 0, 2, 3, 4, 6, etc. The lower the number, the higher end the chip, for example the 4090 is AD102 and the next card down, the 4080 is AD103.

 

Here's a breakdown of how you got less for more over time in this class, similar to the x80 tier.
 

• GTX 470 (2010): 1.28GB GDDR5 320-bit bus, GF100 chip (cut down highest end Fermi die) - $349
• GTX 570 (2010): 1.28GB GDDR5 320-bit bus, GF110 chip (cut down highest end Fermi refresh die) - $349
• GTX 670 (2012): 2GB GDDR5 256-bit bus, GK104 chip (cut down mid-range Kepler die) - $400
• GTX 770 (2013): 2GB GDDR5 256-bit bus, GK104 chip (rebrand of the GTX 680) - $399
• GTX 970 (2014): 3.5+0.5GB GDDR5 224+32-bit bus, GM204 chip (cut down mid-range Maxwell die) - $329
• GTX 1070 (2016): 8GB GDDR5 256-bit bus, GP204 chip (cut down mid-range Pascal die) - $449 (Founder's Edition), $379+ (AIB models)
• GTX 1070 Ti (2017): 8GB GDDR5 256-bit bus, GP104 chip (functionally same mid-range Pascal die with only one less SM than the GTX 1080 vs 5 on the 1070 - Released to beat the Vega 56) - $449
• RTX 2070 (2018): 8GB GDDR6 256-bit bus, TU106 chip (Not even the mid-range die anymore, now shares same chip as xx60 class) - $599 (Founder's Edition), $499+ (AIB models)
• RTX 2070 Super (2019): 8GB GDDR6 256-bit bus, TU104 chip (Notice, this version has same mid-range Turing die as 2080 and 2080 Super now, only cut down vs those chips. Released to combat 5700XT) - $499+
• RTX 3070 (2020): 8GB GDDR6 256-bit bus, GA104 chip (mid-range Ampere die) - $499 (Founder Edition), $499+ (AIB models)
• RTX 3070 Ti (2021) 8GB GDDR6X 256-bit bus, GA104 chip (mid-range Ampere die) - $599
• RTX 4070 Ti (2023) 12GB GDDR6X 192-bit bus, AD104 chip (3rd tier Ada Lovelace die) - $799

So let's analyze. We see in the Fermi generation that the 70-class utilized the same die as the 80-class, only cut-down. Since in that generation, the 80-class card was the top flagship card (there were no 80 Ti's and Titans), we see you got the second best card on the top end die for around $350.

 

Then something changed with Kepler. The GTX 670 still shared the same die class as the GTX 680, but noticed it went from a 100 die to a 104. They dropped a die class for the 70 (and 80) card, die size (meaning cheaper to make, larger yields per wafer, higher margin), and yet increased the price to $400. And if you were around back then, you know Nvidia pulled one over on everyone because the mainstream tech press loved these cards even though they were barely 30% better than Fermi (and at the time that was considered a very poor generational performance upgrade). Its clear what happened here. The 7970 was beatable by Nvidia with their mid-range silicon, so they rebranded, kept the high-end pricing and we didn't see any actual high-end Kepler until GK110 products released. Now let's talk about those.

 

Once Nvidia was ready to show GK110 products we first got the GTX Titan. It was an at the time ludicrous $1000 and not at all worth the cost. And believe me, at that time it was marketed for gaming. Yet even at $1000, it was not the fully unlocked chip. But notice how in one generation the flagship went from $500 to $1000. The GTX 780 then released utilizing the same die but it was heavily cut down and the 70-class card was simply a rebranded GTX 680 still on GK104. Nvidia then got shocked and upset by the $550 R9 290X and then released full GK110 in the 780 Ti.

 

Then with both Maxwell and Pascal we got a similar story, the 80-class and 70-class cards were on the same mid-range silicon while the x80 Ti cards were on traditional larger/flagship silicon for more money and Titans did the same still.

 

Then let's talk about Turing. Minuscule generational on generational performance increases from Pascal, yet came with a huge price increase and for the first time we see the 70-class card drop down to the third tier chip (TU106), the same silicon for 60-class cards since the Kepler generation. Yet it was also the at the time, highest price for a 70-class card ever. It was corrected with the 2070 Super.

 

Ampere sort of resets some things but also introduced some other shenanigans. The 80-class card could not be on 104 silicon because Samsung's 8nm process was too old and Nvidia was tapping it for all it could give, so it went back to the 3070 (and later 3070 Ti). But higher up in the stack we see something strange and that's the introduction of 90-class cards. Linus himself said it during the 3090 review: When you look at the performance vs the Titan RTX, it's clear the 3090 is not actually a Titan. In fact I'd posit that it was a rebrand and replacement for the 2080 Ti with a $300 price hike. Everyone would consider the 2080 Ti the flagship gaming card of the Turing generation. Absolutely nobody would call the 3080 Ti the flagship gaming card of the Ampere generation. I rest my case.

 

And now the Ada Lovelace generation. We once again see a Turing like drop in die-class for the 70-class product, yet we also see a massive price hike.

 

I say all this to provide a bit of a history lesson and to show Nvidia has been screwing us gamers for a long time and we've been buying them up and letting them get away with it. It's good to see many taking a stand now, but I fear we've let the cat out of the bag 11 years ago, and what we've seen is the natural evolution of that. Absolutely no one in mainstream tech press called out Nvidia for what they did with Kepler in the GTX 670/680 and in fact gave them glowing reviews. And thus we started the march towards Nvidia charging more and giving less.

 

Moral of the story, you give Nvidia an inch, they will gladly take a mile and then some over time.

 

Thanks for listening to my TedTalk.

We actually talk about this a fair bit on WAN show, but I recently realized we've never really done a dedicated LTT on the subject. I actually asked for this from the labs 2 weeks ago

 

 

[Bitter]

34 minutes ago, LinusTech said:

We actually talk about this a fair bit on WAN show, but I recently realized we've never really done a dedicated LTT on the subject. I actually asked for this from the labs 2 weeks ago

 

 

Go old school, triangles per dollar! This will be an enlightening video when it's made.

[Sir Beregond]

1 hour ago, LinusTech said:

We actually talk about this a fair bit on WAN show, but I recently realized we've never really done a dedicated LTT on the subject. I actually asked for this from the labs 2 weeks ago

 

 

Very exciting! I look forward to that content!

 

I have done a lot of thought about this very thing over the last several years. I'm glad its really starting to be something people are paying attention to and that more are shining a light on it.

 

49 minutes ago, Bitter said:

Go old school, triangles per dollar! This will be an enlightening video when it's made.

Hahaha, that takes me back. We can talk about shader and vertex pipes instead of CUDA cores.