News Nvidia's RTX 4070 to Launch at $599: Report

[Admin]

The Nvidia GeForce RTX 4070 appears to be primed for a $599 launch price, which some will see as a forgivable $100 increase on 2020’s RTX 3070 MSRP.

Nvidia's RTX 4070 to Launch at $599: Report : Read more

 

[BX4096]

The Nvidia GeForce RTX 4070 appears to be primed for a $599 launch price, which some will see as a forgivable $100 increase on 2020’s RTX 3070 MSRP.

That's only if you see it as a proper, good-faith successor to 3070. 4070 ti is already just 50-60% faster than the 3070 in raster (Tom's rasterization benchmarks place it at 30-72%, depending on the resolution), and a good part of that difference can largely be attributed to the 3070 cards 8GB VRAM limit. Considering how gimped 4070 appears to be compared to the so-called "ti" (a.k.a renamed "4080 12gb", another outrage), some will see it as a rebranded 4060 ti with a rather steep $200 increase from 3060 ti's $399 MRSP.

 

[helper800]

This GPU is at best a 3060 ti equivalent if you look at die sizes. The, "RTX 4070," has a 295mm² die while the 3060ti is at 392mm² and the 3060 has a 276mm² die size. I would not pay for a 600 dollar 3060 tier card. The 3060 was already overpriced, add another 200 dollars to that and you have an oligarchic Nvidia tax you are paying. That is my 2 cents, take it or leave it.

 

[ngilbert]

This GPU is at best a 3060 ti equivalent if you look at die sizes. The, "RTX 4070," has a 295mm² die while the 3060ti is at 392mm² and the 3060 has a 276mm² die size. I would not pay for a 600 dollar 3060 tier card. The 3060 was already overpriced, add another 200 dollars to that and you have an oligarchic Nvidia tax you are paying. That is my 2 cents, take it or leave it.

Because die size is such a great comparison when dealing with chips that are manufactured under very different processes. The RTX 30XX series were manufactured under 10nm process, while the 40XX are under a 5nm process - so the 295mm² 4070ti has 26% more transistors than the 628mm² 3090ti.

 

[Deleted member 2731765]

If you ask me, at $499 US price tag, this card would really have been even better and highly competitive against AMD's mid-range options that might launch soon.

Now if we take the new $599 USD into perspective, then it is worth noting that Nvidia's MSRP prices are minimum prices for their graphics cards, so expect many, or most, RTX 4070 GPU models from AIBs to sell at even higher prices.

 

[helper800]

Because die size is such a great comparison when dealing with chips that are manufactured under very different processes. The RTX 30XX series were manufactured under 10nm process, while the 40XX are under a 5nm process - so the 295mm² 4070ti has 26% more transistors than the 628mm² 3090ti.

Die size has been the main contributor to cost since the 90s and nodes have shrunk many times since then and similar sized die parts were almost always within the same price bracket based on die size. If price increased based on transistor count alone we would have billion dollar GPUs and CPUs depending on when you start scaling it. I dont buy all the excuses that the manufacturing is so much more expensive as an excuse for the price increases. Every single time there has been a new release of products they have been on teh most advanced node available to them which for the time was also the most expensive manufacturing process. Despite this, prices have increased hundreds of dollars multiple generations in a row for same named parts, let alone similar die sized parts. Again, these are some facts mixed with my opinions. I am fallible.

 

[Why_Me]

My thoughts ... first off this card will naturally be meant for gaming at 1440 but you are going to have n00bs complaining that this card doesn't do well at 4K. Secondly we can only assume for now until we see the reviews.

I'm of mind this card will average 15 - 20 FPS less than the 4070 Ti.

 

[Eximo]

Because die size is such a great comparison when dealing with chips that are manufactured under very different processes. The RTX 30XX series were manufactured under 10nm process, while the 40XX are under a 5nm process - so the 295mm² 4070ti has 26% more transistors than the 628mm² 3090ti.

Samsung 8nm actually.

And TSMC N4 (which is still part of their 7nm node family)

 

[InvalidError]

Die size has been the main contributor to cost since the 90s and nodes have shrunk many times since then and similar sized die parts were almost always within the same price bracket based on die size.

Up until 12-14nm, most costs were relatively constant and we had wafers at a steady ~$3000 for a very long time. With the new sub-10nm stuff though, we've seen per-wafer cost jump from ~$6000 to $17 000+. That 300sqmm silicon slab is getting legitimately expensive, albeit still not enough so to justify another $100 price hike on a 25% die+VRAM shrink.

 

[Eximo]

And that is why there is so much being invested in fabs right now. Too much demand for the best process nodes means that TSMC, in particular, can almost charge whatever they want. And when you have large customers like Apple and AMD (Playstation/Xbox is a huge one) on top of all three GPU designers also using them. Intel also using TSMC for CPU graphics, it really is adding up.

And they aren't likely to pass those fab savings on to us. They'll keep prices as high as they can as long a they can.

 

[InvalidError]

And that is why there is so much being invested in fabs right now.

Many companies have rolled back their fab expansion plans due to the global economic slowdown. No point in rushing to finish building fabs when your existing ones are no longer operating anywhere near capacity and you are starting to worry about excess inventory or even having to shut down some of your existing capacity.

 

[thisisaname]

It will be interesting to see if this is more than 20% faster than the over still over MSRP 3060.

 

[The Historical Fidelity]

Die size has been the main contributor to cost since the 90s and nodes have shrunk many times since then and similar sized die parts were almost always within the same price bracket based on die size. If price increased based on transistor count alone we would have billion dollar GPUs and CPUs depending on when you start scaling it. I dont buy all the excuses that the manufacturing is so much more expensive as an excuse for the price increases. Every single time there has been a new release of products they have been on teh most advanced node available to them which for the time was also the most expensive manufacturing process. Despite this, prices have increased hundreds of dollars multiple generations in a row for same named parts, let alone similar die sized parts. Again, these are some facts mixed with my opinions. I am fallible.

Well if you look at price increases gen on gen it actually does track with wafer cost price increases. For TSMC, 12 inch wafer prices have more or less doubled in price every generation: N12 = $3,984, N7 = $9,346, and N5 = $16,988. So taking the 4070’s 295mm^2 die size yields 185 complete dies per wafer, N5 has a current defect density of 0.09 defects per cm^2 so only 141 of these 295mm^2 dies are usable, so price per 295mm^2 die using the N5 defect rate for the following process nodes using their respective wafer prices gives N12 = $28.26, N7 = $66.28, N5 = $120.48 per die. Calculating the price of the RTX 3070’s 392nm^2 die on $5600 Samsung 8N process wafers with a 0.22 defect per cm^2 rate yields 58 good dies. However, since the 3070 die is a partially disabled die, nvidia also uses the defective dies that only have defects in the cuda cores. So let’s assume 33% of defective dies can be used (this is most likely much higher but wanted to be conservative), that gives 84 usable dies for the rtx 3070 at $66.67 per die. All these costs are assuming minimum viable distance between dies (to prevent die damage during wafer cutting step) and no wafer area used for internal testing structures for QA purposes which are always present on production wafers, and perfect square die dimensions to maximize the number of dies that can fit on the wafer.

So as you can see, using similar die sizes between generations is a misinformed way of judging product cost. Given the ~40% margin Nvidia puts on dies they sell to GPU manufacturers, the price is $93.38 per die for the 3070, and $168.68 for the 4070, or $75.30 more. Add on the additional cost of 4 additional gigabytes of GDDR6X at the advertised $12 per GB before GPU manufacturer margin is added and the total cost increase for the 4070 die+4GB additional memory vs 3070 is $123.30 before GPU manufacturer markup and not taking into account additional VRM and pcb mounted components.

Thus it seems that, finally with lovelace, Nvidia is forcing GPU makers to reduce their markup percentage compared to the higher tier 4000 series cards.

 

[The Historical Fidelity]

Samsung 8nm actually.

And TSMC N4 (which is still part of their 7nm node family)

N4 is in the N5 5nm node family, not N7 7nm.

 

[helper800]

Well if you look at price increases gen on gen it actually does track with wafer cost price increases. For TSMC, 12 inch wafer prices have more or less doubled in price every generation: N12 = $3,984, N7 = $9,346, and N5 = $16,988. So taking the 4070’s 295mm^2 die size yields 185 complete dies per wafer, N5 has a current defect density of 0.09 defects per cm^2 so only 141 of these 295mm^2 dies are usable, so price per 295mm^2 die using the N5 defect rate for the following process nodes using their respective wafer prices gives N12 = $28.26, N7 = $66.28, N5 = $120.48 per die. Calculating the price of the RTX 3070’s 392nm^2 die on $5600 Samsung 8N process wafers with a 0.22 defect per cm^2 rate yields 58 good dies. However, since the 3070 die is a partially disabled die, nvidia also uses the defective dies that only have defects in the cuda cores. So let’s assume 33% of defective dies can be used (this is most likely much higher but wanted to be conservative), that gives 84 usable dies for the rtx 3070 at $66.67 per die. All these costs are assuming minimum viable distance between dies (to prevent die damage during wafer cutting step) and no wafer area used for internal testing structures for QA purposes which are always present on production wafers, and perfect square die dimensions to maximize the number of dies that can fit on the wafer.

So as you can see, using similar die sizes between generations is a misinformed way of judging product cost. Given the ~40% margin Nvidia puts on dies they sell to GPU manufacturers, the price is $93.38 per die for the 3070, and $168.68 for the 4070, or $75.30 more. Add on the additional cost of 4 additional gigabytes of GDDR6X at the advertised $12 per GB before GPU manufacturer margin is added and the total cost increase for the 4070 die+4GB additional memory vs 3070 is $123.30 before GPU manufacturer markup and not taking into account additional VRM and pcb mounted components.

Thus it seems that, finally with lovelace, Nvidia is forcing GPU makers to reduce their markup percentage compared to the higher tier 4000 series cards.

Thanks for doing all of that math. I love seeing breakdowns like this. Given those numbers, calculate the cost of a completed graphics card and explain to me how 600 dollars is without Nvidia making gigantic profits above and beyond the 3000, 2000, and 1000 series as a, "tax," on consumers?

 

[Eximo]

N4 is in the N5 5nm node family, not N7 7nm.

Good to know, I did look around but couldn't find a decent source that spelled it out.

Wikichip seems to think that TSMC 5nm is only about 10% more dense than 7nm on average.

As AMD discussed a lot and implemented in their chiplets, and Intel is following suite, is that some features are already at a minimal size and even if the node is capable of smaller features it isn't practical.

 

[helper800]

Good to know, I did look around but couldn't find a decent source that spelled it out.

Wikichip seems to think that TSMC 5nm is only about 10% more dense than 7nm on average.

As AMD discussed a lot and implemented in their chiplets, and Intel is following suite, is that some features are already at a minimal size and even if the node is capable of smaller features it isn't practical.

It does not really matter anyways because none of the, "nm," talk even means anything anymore. Its basically all marketing so we can understand that, "5nm," is better than, "7nm," because smaller = better.

 

[Eximo]

Yep, none of them seem to want to move to a practical measurement like LMC (Logic/Memory/Cache sizes), which was proposed by their own engineers.

 

[InvalidError]

It does not really matter anyways because none of the, "nm," talk even means anything anymore. Its basically all marketing so we can understand that, "5nm," is better than, "7nm," because smaller = better.

Smaller has practically always been better in most ways. The only thing that has changed is the scaling gap between numerical process names and benefits.

There is only so much you can do to make numerical names make some degree of sense when you butt against practical limits on critical dimensions like channel length at ~50nm.

 

[helper800]

Smaller has practically always been better in most ways. The only thing that has changed is the scaling gap between numerical process names and benefits.

There is only so much you can do to make numerical names make some degree of sense when you butt against practical limits on critical dimensions like channel length at ~50nm.

The problem with the industry is being able to actually differentiate between the the different process nodes in a meaningful way in a couple words or less. They have stuck to what worked in the past which is the, "nm" measurement, but that has lost its meaning in its true sense. Coming up with a new easily comparable metric between process nodes that easily, at a glance, tells you, "this one is better/more advanced than that one," is nigh impossible with how complicated process nodes are now.

 

[The Historical Fidelity]

Thanks for doing all of that math. I love seeing breakdowns like this. Given those numbers, calculate the cost of a completed graphics card and explain to me how 600 dollars is without Nvidia making gigantic profits above and beyond the 3000, 2000, and 1000 series as a, "tax," on consumers?

I just did…I showed the cost between 3070 and 4070 dies goes up up by $75 dollars and $48 of additional ram on the 4070 more than justifies the $100 price increase meaning Nvidia is making less profit on the 4070 than the 3070.

To put it simply, when the cost of components goes up $123, then a $100 price increase to cover the additional cost leaves negative - $23 dollars for added profit. IE The 4070 at msrp has less margin than a 3070.

 

[helper800]

I just did…I showed the cost between 3070 and 4070 dies goes up up by $75 dollars and $48 of additional ram on the 4070 more than justifies the $100 price increase meaning Nvidia is making less profit on the 4070 than the 3070.

To put it simply, when the cost of components goes up $123, then a $100 price increase to cover the additional cost leaves negative - $23 dollars for added profit. IE The 4070 at msrp has less margin than a 3070.

Okay so, assuming the die of the GPU is as you say it is, 168.68, plus 12gb of GDDR6X at 12 dollars a gb (144$) we are looking at 312.68. Does the PCB, components on that PCB, and the HSF and fans add up to 600 dollars?

 

[InvalidError]

I just did…I showed the cost between 3070 and 4070 dies goes up up by $75 dollars and $48 of additional ram on the 4070 more than justifies the $100 price increase meaning Nvidia is making less profit on the 4070 than the 3070.

GDDR6X may have been $12/GB in 2021 but i doubt it is anywhere near that today where most DRAM prices have dropped 30-40% since.

 

[The Historical Fidelity]

GDDR6X may have been $12/GB in 2021 but i doubt it is anywhere near that today where most DRAM prices have dropped 30-40% since.

Regardless, the price of die and additional ram alone corroborates the $100 price increase over 3070, that was the argument.

 

[The Historical Fidelity]

Okay so, assuming the die of the GPU is as you say it is, 168.68, plus 12gb of GDDR6X at 12 dollars a gb (144$) we are looking at 312.68. Does the PCB, components on that PCB, and the HSF and fans add up to 600 dollars?

It doesn’t have to = $600, you are forgetting about gpu manufacturer margin which is ~20% and end seller margin of ~5-10%, so the real question is does the rest of the GPU + GPU manufacturer margin + end seller margin equal $600.