Review Nvidia GeForce RTX 4070 Ti Super review: More VRAM and bandwidth, slightly higher performance

[AgentBirdnest]

I guess all RTX 4090 cards are "refuse" because they only have 128 out of 144 SMs enabled? And all H100 and A100 chips are also "refuse" because they're not fully enabled?

Also, for the analogy to work, Nabisco needs to be binning those few cookies that got the recipe 100% perfect. I can't wait for "Oreo Ti" cookies.
: P
When I first learned about binning, I was kinda pissed. It seemed so shady. "Wait, I have a 7700K in my PC but can't use it? This is BS!" (I had thought that every single CPU SKU was a different design.) But after learning a little more about the fabrication process, learning how much dies cost, and thinking about the logic, it made perfect sense.

 

[CmdrShepard]

Mashing them up and putting them back together isn't what happens, though, not even remotely.

Except I never claimed that's what happens with GPU dies.

I used a real-life example of a "package of whole biscuits" and a "package of ground biscuits" costing almost the same just because they have the same weight (same die type) even though one of them is made out of broken biscuits (faulty die) which couldn't be sold as a "package of whole biscuits" (fully operational die).

It's not my fault that you aren't capable of abstracting the process of die manufacturing and applying it to other manufacturing processes which use the same "optimization" trick.

And the downbinned chips are simply part of the design. You build in redundancies as well as the ability to fuse off parts of the chip so that you can actually make use of nearly all the dies that come from a wafer. To do otherwise would be wasteful and a poor approach to modern design.

I did say the same thing, so we agree on that. What I do not agree on is the ethics component of that process toward customers.

Also, some chips will "fail" because they just need more of a lower tier — so they don't actually fail. Parts of some chips might also fail due to the voltages needed for the highest clocks. So a chip can be fully functional, but if it needs 1.3V to hit the desired 2.5 GHz or whatever while other chips can do that with 1.2V, they can turn off the bits that need more voltage.

And it follows that I understand this as well, given how long I am building my own PCs.

Calling a chip "refuse" just because it's not fully enabled is, IMO, stupid. I guess all RTX 4090 cards are "refuse" because they only have 128 out of 144 SMs enabled? And all H100 and A100 chips are also "refuse" because they're not fully enabled?

Yes.

You don't buy a medium pizza and then complain that you didn't get a large pizza.

You also don't order a medium pizza, pay almost the same price as a large pizza (because they are made from the same wafer... sorry, baking tray), and then get a large pizza with a slice or two missing.

You seem to be justifying the current pricing scheme (too much cool aid from NVIDIA and AMD?), while I believe it's not based on actual die cost of those cut down products which are sold as whole new products.

That seems to be the real disagreement we have here.

 

[JarredWaltonGPU]

You also don't order a medium pizza, pay almost the same price as a large pizza (because they are made from the same wafer... sorry, baking tray), and then get a large pizza with a slice or two missing.

The prices aren't even remotely the same.

AD104 RTX 4070 Ti: $799 originally, now discontinued to around $700 while it lasts, uses the fully functional chip.
AD104 RTX 4070 Super: $599, uses a 93% enabled chip.
AD104 RTX 4070: $549 (now, formerly $599), uses a 77% enabled chip.

I do feel the 4070 could still drop lower, but what's the baseline here? If it's fully enabled chips, the 4070 Ti at $700 would be what we need to look at. So, you get about 20% less performance from the 4070 versus the 4070 Ti, and you pay 21% less money. That seems perfectly reasonable to me. Whether either one should cost what it does is just shouting at the sky for things not being how you might want them to be (you in general, not you specifically).

I've got a better analogy than pizza and biscuits, though! You can go to a restaurant and order the half rack of baby back ribs for $25, or the whole rack for $35. Either way, though, you're getting a 'binned' rack of ribs! Because if you've ever gone to buy a full rack of baby back ribs from somewhere like Costco, you'd know that the restaurant full rack is really half of a rack, and the half rack is a quarter of a rack. So no matter the price in the menu or the taste of the food, anything but the truly full rack of ribs is refuse because you didn't get everything the pig or cow provided!

I don't love the pricing, but Nvidia and AMD are doing what makes business sense to them. I look at prices of everything right now, and it's all clearly way more expensive than 5-10 years back. It certainly feels justifiable to me that a modern "high-end" GPU should cost more than what it did in 2014, or even 2019. I don't like that it costs more, just as I don't like how much my various bills have gone up, but that's the brakes.

 

[bit_user]

It's not my fault that you aren't capable of abstracting the process of die manufacturing ...

Woah, easy there. Given how quick you are to accuse others of ad hominem attacks, I expected better from you.

Jarred is taking his valuable time to talk with us, here in the forums. Be nice and don't take it for granted. The mods tend to take a pretty dim view of abuse towards the authors.

 

[CmdrShepard]

The prices aren't even remotely the same.

AD104 RTX 4070 Ti: $799 originally, now discontinued to around $700 while it lasts, uses the fully functional chip.
AD104 RTX 4070 Super: $599, uses a 93% enabled chip.
AD104 RTX 4070: $549 (now, formerly $599), uses a 77% enabled chip.

I do feel the 4070 could still drop lower, but what's the baseline here? If it's fully enabled chips, the 4070 Ti at $700 would be what we need to look at. So, you get about 20% less performance from the 4070 versus the 4070 Ti, and you pay 21% less money. That seems perfectly reasonable to me. Whether either one should cost what it does is just shouting at the sky for things not being how you might want them to be (you in general, not you specifically).

The problem is we don't know what's the manufacturing price of a single fully working die. It could be $150 for all we know and then the question is whether the $1,800 for a full product and $600 for the lowest possible one is really justified or just unchecked capitalist greed.

https://medium.com/@porinapew/estimating-gpu-die-costs-ccd65d4c1c70

Also, I wouldn't really say "93% enabled chip" just based on a CUDA core count. Are you factoring in reduced TMU, ROP, etc count and lower clocks into what's enabled?

 

[JarredWaltonGPU]

The problem is we don't know what's the manufacturing price of a single fully working die. It could be $150 for all we know and then the question is whether the $1,800 for a full product and $600 for the lowest possible one is really justified or just unchecked capitalist greed.

https://medium.com/@porinapew/estimating-gpu-die-costs-ccd65d4c1c70

Also, I wouldn't really say "93% enabled chip" just based on a CUDA core count. Are you factoring in reduced TMU, ROP, etc count and lower clocks into what's enabled?

4070 Super and 4070 Ti both have the full complement of ROPs: 80. TMU is tied to the SM count, so it's still 93%. And if you want to be pedantic about it, probably only half of the die is for all the SMs and related stuff, with the other half being memory controllers, video codec stuff, video outputs, etc.

In terms of raw performance, in real-world testing, the 4070 Ti is about ~8 percent faster than the 4070 Super. Or if you flip that, the 4070 Super is about ~7 percent slower. Either way, it's close enough to 93%.

What's the actual cost of the silicon? If we estimate the wafer price from TSMC at around $14,000 (give or take $1,000), each AD104 chip would cost $70~$75. But that's purely for the silicon. If you're only looking for perfect chips, the price goes much higher because yields will decrease. But again, only going after perfect chips is stupid on stuff the size of modern GPUs. Nvidia can get about 195~200 AD104 chips out of a wafer, and if it can use 99% of those via binning and harvesting, that brings the average price per chip down.

How much did R&D costs? How much for everything else like the memory, PCB, VRMs, etc.? And how much profit should a publicly traded company try to make? (Answer: As much as possible.) And everyone needs a cut: The AIB partners, the distributors, and retail.

So sure, call it unchecked capitalist greed. I call it getting paid for doing the hard work that's required to create new and interesting stuff. And when a company tries to charge more than you're willing to pay, find an alternative or just do without.

 

[Eximo]

The problem is we don't know what's the manufacturing price of a single fully working die. It could be $150 for all we know and then the question is whether the $1,800 for a full product and $600 for the lowest possible one is really justified or just unchecked capitalist greed.

https://medium.com/@porinapew/estimating-gpu-die-costs-ccd65d4c1c70

Also, I wouldn't really say "93% enabled chip" just based on a CUDA core count. Are you factoring in reduced TMU, ROP, etc count and lower clocks into what's enabled?

Essentially priced by the wafer and then you do the binning, and then calculate by yields. Then factor in all other business expenses including drivers, support, marketing, etc.

If the Quadro and Geforce use the same die, then why does one cost more than the other?

 

[JarredWaltonGPU]

FYI, in case anyone wants to see my calculations, this is the data I've put together on the various GPUs.

GPU	Size	Width	Height	Chips per Wafer	Wafer $	Chip $		Die Size Calc
AD102	608.5	23.2	26.2​	90​	$14,000​	$156​		607.84​
AD103	378.6	24.6	15.4​	150​	$14,000​	$93​		378.84​
AD104	294.5	19.5	15.1​	196​	$14,000​	$71​		294.45​
AD106	187.8	12.6	15.0​	302​	$14,000​	$46​		189​
AD107	158.7	12.5	12.7​	362​	$14,000​	$39​		158.75​
GA102	628.4	23.4	26.9​	88​	$5,000​	$57​		629.46​
GA104	392.5	19.9	19.7​	146​	$5,000​	$34​		392.03​
GA106	276	14.2	19.5​	206​	$5,000​	$24​		276.9​
GA107	200	14.6	13.7​	292​	$5,000​	$17​		200.02​
TU102	754	24.5	30.8​	71​	$4,000​	$56​		754.6​
TU104	545	24.4	22.3​	102​	$4,000​	$39​		544.12​
TU106	445	24.1	18.5​	126​	$4,000​	$32​		445.85​
TU116	284	22.3	12.7​	200​	$4,000​	$20​		283.21​
TU117	200	14.3	14.0​	291​	$4,000​	$14​		200.2​
Navi 31	300	24.5	12.3​	188​	$14,000​	$74​		301.35​
Navi 32	200	12.3	16.2​	280​	$14,000​	$50​		199.26​
Navi 33	204	13.0	15.7​	281​	$7,000​	$25​		204.1​
Navi 21	520	28.6	18.2​	107​	$7,000​	$65​		520.52​
Navi 22	335	18.3	18.3​	169​	$7,000​	$41​		334.89​
Navi 23	237	16.6	14.3​	245​	$7,000​	$29​		237.38​
Navi 24	107	13.2	8.1​	558​	$7,000​	$13​		106.92​
Navi 10	251	17.8	14.1​	228​	$7,000​	$31​		250.98​
Navi 14	158	13.6	11.6​	374​	$7,000​	$19​		157.76​
H100	814	24.8	32.9​	64​	$14,000​	$219​		815.92​

The chip dimensions are based off of chip images, and should be accurate to within a few percent. Wafer costs are basically an educated guess and you can put in your own values. Chips per wafer comes from this calculator, using the default scribe sizes and the dimensions listed in the table, with a 300mm wafer size.

The TLDR is that the actual cost of the silicon for a chip is relatively negligible compared to the price of the full graphics card.

Most of that is for R&D (basically anyone working for a GPU company needs to be paid), is my assumption. And there's probably a minimum cost of something like $50 just to do the PCB, cooler, and other stuff associated with a graphics card — at least for the typical dual-slot desktop cards that need one or more external power connectors. Bigger, more powerful cards obviously need better coolers and such, but even on something like RTX 4090 or H100, that's probably going to be $250 at most.

 

[bit_user]

Most of that is for R&D (basically anyone working for a GPU company needs to be paid), is my assumption.

Jensen has previously claimed that most of their engineering employees are software developers. That was several years ago, so I'm sure it's now even more true.

People expect their expensive GPUs to be supported for many years. They want game day drivers and optimized DLSS, even for previous-generation hardware. That requires work and costs $$$! As consumers don't like paying subscription fees (i.e. support contracts) for this stuff, all of that must be funded by up-front hardware pricing.

 

[Eximo]

As consumers don't like paying subscription fees (i.e. support contracts) for this stuff, all of that must be funded by up-front hardware pricing.

Don't give them any ideas...