Understanding Nvidia

[Paranoir andando]

Understanding Nvidia (and its new 4060Ti)

 

I guess many people have already noticed it, but may be some have noticed there is something strange in Lovelace and simply they don't understand the cause.

 

Introduction

Nvidia has been following the same order for many years, dividing its GPU in 3 pieces: High, Mid, Low range, therefore, Full GPU (~600 mm²), 2/3 of full (66%), 1/3 of full (33%)

 

Usually:
High Range (< Full) = 100% - 67% CUDAs (TMU, RTcores, Tensor,...) with full ROP & Memory bus = 384 bit
Mid range (< 2/3) = 66% - 33% CUDAs (TMU, RTcores, Tensor,...) with 2/3 ROP & Memory bus = 256 bit
Low range (< 1/3)= 32% - 0% CUDAs (TMU, RTcores, Tensor,...) with 1/3 ROP & Memory bus = 128 bit

 

Near 100% cudas & 384 bit --> x90-x80Ti

 

(2/3) 66% cudas & 256 bit --> x80
. . . . (1/2) 50% cudas & 256 bit ---> x70

. . . . . . . . . 42~45% cudas & 192bit ---> x60Ti

 

(1/3) 33% cudas & 192 bit --> x60
. . . . (1/4) 25% cudas & 128 bit ---> x50
. . . . (1/6) 16,6% cudas & 128 bit -> Thrash, Low-end shits
Under this, the horror, & with less of 128bit

 

The x80 class changed a little in Ampere, it level up to 80%. (and Tesla2 had a 512b bus)

 

 

x50 class, fake x60 name, and x70 price.

 

In this new architecture, Lovelace, the engineers have done the same as always, but some marketing genius has thought, seeing the unfortunate situation of AMD, to reduce a step, or two, the whole Lovelace family, by putting fake names, to sell them for twice their usual price. It is not only in 4060Ti, the same goes for all other Lovelace.

 

Lovelace architecture has the double of cudas than Ampere (18432 vs 10752) thanks to being manufactured in 5nm instead of 8nm as Ampere, and higher frequency. A great advance that has been ruined by giving people less part of the cake than ever. Everything for me, as a glutton, must have thought Nvidia.

 

The Lovelace 4060Ti should have had the double of cudas(tmu,rt,tensor,etc) than the Ampere 3060Ti but it really has less cudas than previous Ampere 3060Ti. Sad but true. Because they are trying to sell you the x50 (24% of max Lovelace) renamed as x60Ti (should be 42% of max Lovelace, as is usual).
Nvidia put usually the name x60 Ti (42%-45% of max chip) to a slightly cutted version of x70 (1/2 of max chip), meanwhile the new marketing genius have put that name to a gpu of 1/4 of full Lovelace, which it should have been  the 4050, as always.

 

 

X50 series usually approaches or coincides with the performance of previous x60 and this small Lovelace 4050 (fake 4060Ti) with a huge factory “overclock” outperform previous x60 Ti (not always). Yes, Lovelace is a great advance. It perform very well for a x50 card, even with the PCIe 8X it has, which can be normal on these low-end cards.

 

Memory bandwidth is not the problem, the 128 bits are OK for this card, the issue is that a 24% of Lovelace can't give more. They’ve changed the name 4050 with the 4060Ti, as AMD did with the HD 7790 --> R7 260X with a sticker over the name, XD. There are people who say that Lovelace has a smaller bus & less bandwidth because it doesn't need so much bandwidth (and bus) as the previous architectures. This is absolutely wrong.

 

Lovelace has the same memory bus (& the same or higher bandwidth) as the previous architectures and Lovelace needs the same memory bus (& the same or higher bandwidth) as the previous architectures. The increase of cache is irrelevant for the comparison because is a characteristic of all Lovelace family, and I only compare with the full chip (inside the architecture):

 

Full Lovelace has and needs (same or) more bandwidth than previous architectures
1/2 of Lovelace has and needs (same or) more bandwidth than previous architectures
1/3 of Lovelace has and needs (same or) more bandwidth than previous architectures
1/4 of Lovelace has and needs (same or) more bandwidth than previous architectures
To say the opposite is to lie.

 

Price:
GTX 970 (54% of full Maxwell) cost $330 ,
GTX 1070 (50% of full Pascal) cost $379 ,
the expensive miner RTX 2070 (50% of full Touring) cost $499 & RTX 2070 Super (55% of full Touring) cost $499 and it's 445mm².
Even the very expensive miner RTX 3070 (55% of full Ampere) cost $499.
I hear from here Nvidia's laughs trying to sell this "a quarter-range GPU" (1/4 of full Lovelace), a 23,61% of Lovelace at the same price as previous x70. Nvidia don't know that plandemic & mining times have ended. All Lovelace has a horrible price.

Nvidia will cut prices soon (imho). The real4050 (fake4060Ti) --> $300. People will be happy. Paying twice as much money than previous generations for a x50 gpu, for a $150 gpu. It's a great sale strategy, to charge $300 for something of $150 and to be applauded.

 

Consumption:
Advantages: 4060Ti consumes less than 3060Ti, less than 2060, ...
Disadvantages: 4060Ti consumes less because is a 4050 with a new fake name, with a chip that is much smaller than it should (less cudas, less performance).

 

The entire Lovelace series have very high frequencies with huge consumption. You can see it in the largest 4090 with its 450W (it can easily increase up to 600W) that could have been 300W with a little less frequency, but also in this small chip, a 75W-100W x50 chip that consumes 160W:

GTX 1650 Die Size = 200 mm²
GTX 1650 Super . . = 284 mm²
RTX 3050 Die Size = 276 mm²
RTX 4060 Ti Die Size = 190 mm²
None of these cards consume as much as 4060Ti (real4050). The 4060Ti has a huge factory overclock that greatly increases the consumption of its ridiculous chip.

 

I hope you like it, especially the images. Sorry for the kicks to the english language.

[adm0n]

If you really want to understand the market situation, you'll also have to look at how Nvidia is making money. The consumer sector is great for volume, while the margins (and thus the profit), isn't that great. On the other hand, selling the same product to industry gives them by far greater profits.

 

Just yesterday, Nvidia told press, that they have beat their already impressive profit predictions by multiple billions of dollars. And the AI boom isn't over yet, so they can't produce enough chips to satiate the demand. So currently, they have little use for selling chips in volume to the consumer market. On top of that, AMDs competition has been lackluster, leaving them still as a pretty viable option.

 

All of this together means, that they can easily increase effective prices dramatically, while not really losing any profits, because they have a different market they are currently supplying. If these new prices stick, which they might, they can reap the profits. If they don't, they can go back a bit next generation and show dramatic improvements over the 4000 series, and everyone will be impressed.

 

They basically have their cake and eat it too. It is sad, but if it all works out for them, it was really a great business decision.

[Paul17]

21 hours ago, adm0n said:

If you really want to understand the market situation, you'll also have to look at how Nvidia is making money. The consumer sector is great for volume, while the margins (and thus the profit), isn't that great. On the other hand, selling the same product to industry gives them by far greater profits.

 

Just yesterday, Nvidia told press, that they have beat their already impressive profit predictions by multiple billions of dollars. And the AI boom isn't over yet, so they can't produce enough chips to satiate the demand. So currently, they have little use for selling chips in volume to the consumer market. On top of that, AMDs competition has been lackluster, leaving them still as a pretty viable option.

 

All of this together means, that they can easily increase effective prices dramatically, while not really losing any profits, because they have a different market they are currently supplying. If these new prices stick, which they might, they can reap the profits. If they don't, they can go back a bit next generation and show dramatic improvements over the 4000 series, and everyone will be impressed.

 

They basically have their cake and eat it too. It is sad, but if it all works out for them, it was really a great business decision.

Do you want to translate that to English?   I don't think the OP's first language is English either.

I read that Nvidia is heavily getting into AI.   That should scare a lot of ppl.    So, their investments in that is making them profits allowing them to 'discount' their current, crippled/gimped 40 series cards - everything 4070 Ti down, right?

With hardware tech reviewers (Youtube content creators) and ppl online repeating that Nvidia are greedy, shady con artists - the only ppl who will buy these cards are ppl who don't know any better - i.e. non-tech ppl who need a cheap video card.    They probably won't make a lot of money from that.   

[adm0n]

4 minutes ago, Paul17 said:

Do you want to translate that to English?

I was actually scared for a second, that I accidentally wrote in another language o__O

 

4 minutes ago, Paul17 said:

So, their investments in that is making them profits allowing them to 'discount' their current, crippled/gimped 40 series cards - everything 4070 Ti down, right?

My argument was the other way around. Right now, they don't need to sell GPUs to consumers, because the AI industry is buying up all the chips they can make. Together with the fact, that the chips Nvidia sells to the industry cost way more, they is an opportunity cost associated with them selling chips to consumers.

 

So they are not discounting cards for consumers with their AI earnings, they are increasing the price of the consumer GPUs, so they have more chips to sell to the AI industry.

 

7 minutes ago, Paul17 said:

They probably won't make a lot of money from that. 

Well apparently it works, because they somehow made 50% more money than predicted. And the prediction wasn't bad either.

[Paranoir andando]

On 5/26/2023 at 9:38 PM, adm0n said:

If you really want to understand the market situation, you'll also have to look at how Nvidia is making money. The consumer sector is great for volume, while the margins (and thus the profit), isn't that great. On the other hand, selling the same product to industry gives them by far greater profits.

 

Just yesterday, Nvidia told press, that they have beat their already impressive profit predictions by multiple billions of dollars. And the AI boom isn't over yet, so they can't produce enough chips to satiate the demand. So currently, they have little use for selling chips in volume to the consumer market. On top of that, AMDs competition has been lackluster, leaving them still as a pretty viable option.

 

All of this together means, that they can easily increase effective prices dramatically, while not really losing any profits, because they have a different market they are currently supplying. If these new prices stick, which they might, they can reap the profits. If they don't, they can go back a bit next generation and show dramatic improvements over the 4000 series, and everyone will be impressed.

 

They basically have their cake and eat it too. It is sad, but if it all works out for them, it was really a great business decision.

Yes, but...

 

Yes, but you should think in Nvidia as several different companies, with several CEOs. That one of them wins money does not imply that the other doesn't win it. Each CEO of each section wants to give the best results, and does not compete against the other but against itself, against its previous results of sales number, market share and benefits.

 

That Nvidia is doing well with machine learning chips doesn't mean it throws away its pc graphics card company. And this section is not going so well, Lovelace is selling very little cause the abusive prices it have,, and they know it. They know it so much that the 4060Ti on its first day has reached the market with discounts, in Germany it's seen for $378 and in Spain for $372 free shipping (+VAT).

...and Yes, they can produce enough chips to satiate the demand and more, too much more, TSMC and others foundries are producing at half capacity utilization because there is no demand, nothing is sold.

 

Jul_2022 - AMD, Apple, and Nvidia Reportedly Cutting Back on TSMC 5nm Orders

Oct_2022 - TSMC sees major clients continue order cuts, but remains confident

Dec_2022 - AMD, Intel, and Nvidia Reportedly Slash Orders with TSMC
Jan_2023 - TSMC's top 10 customers have reportedly cut their orders

...and no, Nvidia is not so important, its production is not so important. It is only a 6% of TSMC revenue:

May_2023 - TSMC’s top-10/20/30/40 customers; Who spends how much on TSMC? Apple revenue; Top-10 customer dynamics
May_2023 - TSMC's top ten customers are exposed!

 

 

(Yes I'm spanish and don't know english very well, but I try it. I can read well)

Spoiler

A bit humour

Better laugh than cry: RTX 4060 Ti and RX 7600 Generational Leaps  https://www.youtube.com/watch?v=unO5J_u4Yo4

 

 

[Paranoir andando]

To delete

[adm0n]

20 minutes ago, Paranoir andando said:

Yes, but you should think in Nvidia as several different companies, with several CEOs. That one of them wins money does not imply that the other doesn't win it. Each CEO of each section wants to give the best results, and does not compete against the other but against itself, against its previous results of sales number, market share and benefits.

 

That Nvidia is doing well with machine learning chips doesn't mean it throws away its pc graphics card company. And this section is not going so well, Lovelace is selling very little cause the abusive prices it have,, and they know it. They know it so much that the 4060Ti on its first day has reached the market with discounts, in Germany it's seen for $378 and in Spain for $372 free shipping (+VAT).

 

I think I made a pretty sound argument why what they are currently doing, can be very beneficial for Nvidias consumer department and with that Nvidia as a whole in the long run. They are resetting consumer expectations. That way, when the recession passes, and consumer spending increases again, they can have an competitive offering. As an added bonus, it will look especially impressive compared to the current lineup. Or if AMD isn't capable of reigning them in, the new precedence could just stick.

 

24 minutes ago, Paranoir andando said:

...and Yes, they can produce enough chips to satiate the demand and more, too much more, TSMC and others foundries are producing at half capacity utilization because there is no demand, nothing is sold.

They could start ramping up chip production now. That would require contracts with TSMC and all the other involved parties to go through. And then it will take time until TSMC can increase their output. So Nvidia is not capable of immediately reacting to that. Likewise, if the AI boom ends and that demand fades away, they don't want to be left with an oversupply of GPUs. Because once they make those contracts to produce the chips, it becomes pretty hard to get out of them again.

 

I also never attacked your main statement, that Nvidia is basically renaming their product stack. I just wanted to offer an explanation as to why it could be beneficial for them. It's also just a notion or speculation on my part.

[igormp]

What nvidia is doing is just leaving  scraps to their consumer lineup and shifting most - if not all - of their efforts into datacenter. Apart from the mining craze, their DC revenue has already outpaced the consumer market, and it'll only get worse with this AI hype.

[Paul17]

3 hours ago, adm0n said:

I was actually scared for a second, that I accidentally wrote in another language o__O

 

My argument was the other way around. Right now, they don't need to sell GPUs to consumers, because the AI industry is buying up all the chips they can make. Together with the fact, that the chips Nvidia sells to the industry cost way more, they is an opportunity cost associated with them selling chips to consumers.

 

So they are not discounting cards for consumers with their AI earnings, they are increasing the price of the consumer GPUs, so they have more chips to sell to the AI industry.

 

Well apparently it works, because they somehow made 50% more money than predicted. And the prediction wasn't bad either.

I think I'm saying something similar - they don't care that their video cards aren't selling - there's very minimal discounts to the more gimped cards - but, overall, there is no rush or urgency to sell cards (if we are talking about the gaming market) - their flagships 4080s/4090s - could sell more of those with discounts but they don't need to - they are making profits from AI.   That will allow them to weather 'poor sales' with their graphics cards.    In the mean time, if ppl do buy Nvidia cards, they will be at overinflated prices.

https://www.cnn.com/2023/05/25/tech/nvidia-earnings-ai/index.html

https://www.reuters.com/technology/why-are-nvidias-shares-soaring-what-is-its-role-ai-boom-2023-05-25/

They're concentrating on AI, now.   

[Paranoir andando]

You are right, Nvidia is not interested in selling, so the responsible director of graphics cards must be very happy to see this. Everyone is happy:

Nvidia don't want to sell and people don't want to buy. Everything is perfect!