NVIDIA GeForce RTX 5080 Founders Edition

[Bwaze]

It's funny how the reviews here tried to predict all possible price points for RTX 5080 cards, we have entry points for "discounted" $800, one for MSRP at $1000, then the ones for inflated prices - $1200 and $1400.

Should there be more realistic price points added, like $1600, $1800, $2000 and over?




I know some of the scalped prices shown in price comparisons are from individual sellers - they can of course place the price as they see fit. But here all the cards in proper computer stores arrived at already scalper prices - lowest I have seen is 1600 EUR, $1660! Some are straight 2x MSRP!

There were some high prices at RTX 4080 release, but nothing similar to this.

But the good thing is, if you're a wealthy gamer, you have dozens of cards to choose from, everything is still in stock.

 

[Legacy-ZA]

It's funny how the reviews here tried to predict all possible price points for RTX 5080 cards, we have entry points for "discounted" $800, one for MSRP at $1000, then the ones for inflated prices - $1200 and $1400.

Should there be more realistic price points added, like $1600, $1800, $2000 and over?




I know some of the scalped prices shown in price comparisons are from individual sellers - they can of course place the price as they see fit. But here all the cards in proper computer stores arrived at already scalper prices - lowest I have seen is 1600 EUR, $1660! Some are straight 2x MSRP!

There were some high prices at RTX 4080 release, but nothing similar to this.

But the good thing is, if you're a wealthy gamer, you have dozens of cards to choose from, everything is still in stock.

Yes, these ludicrous prices are from online stores, not actual scalpers, these stores are the scalpers now. Absolutely disgusting.

 

[Chrispy_]

Dont buy at launch, wait till mid gen?

But I do think its amateur hour that retailers cant reserve products in a basket. People should not have to speedrun a checkout.

nvidia.com should have been better prepared too, it's not like they're short of cash or inexperienced at this point:

11:59am "coming soon"
12:00pm <server error> isup.me say down for everyone.
12:13pm "out of stock"

Yes, these ludicrous prices are from online stores, not actual scalpers, these stores are the scalpers now. Absolutely disgusting.

Dont forget the AIBs. Bone-stock cards with a $1200 MSRP instead of $1000. What does that extra $200 get you? I sure as heck can't see anything of value. IMO the bone-stock reference cards from AIBs should be cheaper - they're using plastic shrouds and simple copper baseplates with heatpipes rather than vapor chambers. They're a significantly lower quality than the $999 5080FE.

 

[AnotherReader]

nvidia.com should have been better prepared too, it's not like they're short of cash or inexperienced at this point:

11:59am "coming soon"
12:00pm <server error> isup.me say down for everyone.
12:13pm "out of stock"


Dont forget the AIBs. Bone-stock cards with a $1200 MSRP instead of $1000. What does that extra $200 get you? I sure as heck can't see anything of value. IMO the bone-stock reference cards from AIBs should be cheaper - they're using plastic shrouds and simple copper baseplates with heatpipes rather than vapor chambers. They're a significantly lower quality than the $999 5080FE.

This suggests that EVGA was right and Nvidia charges AIBs so much for the higher end chips that it's infeasible for them to offer most SKUs at the founder edition's MSRP.

Paul is wrong on at least one point; There ARE bad GPUs.

However he makes a ton of great points, including a few that I've been taking about. This is the first time I can think of that a new gen of cards didn't surpass the previous gen flagship and yet the price didn't scale appropriately. Paul's points about NVidia's business model are seemingly spot on and it's a piss-poor, pathetically predatory thing.


Do you know what those numbers mean? .10% defects per square centimeter is equal to a 37% yield, give or take.. That is a terrible yield!

Defect distribution follows the Murphy Yield model. Common die yield calculators, e.g. the one from iSine, use that model. You can plug in the numbers and see what the yields for a 5080 would be. A defect density of 0.1 per square centimeter results in 69% yields while a more likely 0.06 increases it to 80%

 

[HD64G]

Yes, these ludicrous prices are from online stores, not actual scalpers, these stores are the scalpers now. Absolutely disgusting.

And its true value as a 70 class tier GPU is between $600-700 mind you. Nvidia is the first scalper and retailers the second ones. The last ones are sadly, the only ones called such.

 

[BigMack70]

Yes, these ludicrous prices are from online stores, not actual scalpers, these stores are the scalpers now. Absolutely disgusting.

I have a hard time believing the scalped prices for the 5080 will stick for more than a couple months. It's not much better than the 4080 and the 4080 / 4080S did not sell well when it was only available at $1200+.

The only world $1400+ prices land for the 5080 and remain is if second hand 4090 supply dries up, the 5090 remains vaporware, and RDNA 4 is uncompetitive. I think the current 5080 pricing is driven mostly by scarcity of high end cards, not real demand.

 

[10tothemin9volts]

Well it was a paper launch or bots bought everything. Good luck to scalpers selling 5080, somebody didnt tell them it's not mining times anymore.

Yes, trying to sell me a 16GB VRAM, and (almost) same performance and power efficiency, 1000 bucks, GPU that late in the game, when 3GB capacity GDDR7 chips are around the corner (see the table, same 378 mm² chip: 16GB VRAM, 24GB VRAM (=3GB GDDR7 chips)). For the rest of the lineup, the 3GB GDDR7 chips may come in next year's Refresh (or in GeForce 60-series).

 

[Chrispy_]

This suggests that EVGA was right and Nvidia charges AIBs so much for the higher end chips that it's infeasible for them to offer most SKUs at the founder edition's MSRP.


Defect distribution follows the Murphy Yield model. Common die yield calculators, e.g. the one from iSine, use that model. You can plug in the numbers and see what the yields for a 5080 would be. A defect density of 0.1 per square centimeter results in 69% yields while a more likely 0.06 increases it to 80%

Last I heard TSMC was at 0.08 defect on that node, and that's 5-6 month old info.

 

[bitsandboots]

Explain the RTX 4090, which absolutely nobody needs, but yet has outsold anything AMD has produced in the last several years. That tells me that $1,000 isn’t too much. Price is only part of the value equation.

I said "$1000 for any computer part is too much, so anyone who buys such a thing must convince themselves it's worth it somehow."
For a 4090, people could convince themselves it was worth buying: it was significantly faster than everything else.
It's also not true that "absolutely nobody" needs a 4090. At the high end of consumer parts you start to get people who use the hardware for jobs of some sort. Cheaper alternatives to actual professional hardware. It's just a cheaper L40 / RTX 6000 for people who don't need double the VRAM.

 

[AnotherReader]

Last I heard TSMC was at 0.08 defect on that node, and that's 5-6 month old info.

Is that for N4 or N3? If it's the former, then it leads to a yield of 74.5% for perfect dies or 104 RTX 5080 candidates per wafer. A good portion of the 36 defective dies could be repurposed as 5070 Ti so the 5080 is nowhere near as costly to make as a naive analysis may indicate.

 

[lexluthermiester]

Huh. HUB and I actually agree on something? That is some weird stuff there.. It's not wrong. The 5080 is what was expected from the 5070. The 5080 was supposed to be about 50% the difference in performance between the 4090 and 5090 and it's not. I know they've been working on something that sits in that performance profile range, so the 5080 is a head scratcher..

Last I heard TSMC was at 0.08 defect on that node, and that's 5-6 month old info.

That number means .08 defects per CM squared. That is an average, but also lot of defects.

Defect distribution follows the Murphy Yield model. Common die yield calculators, e.g. the one from iSine, use that model. You can plug in the numbers and see what the yields for a 5080 would be. A defect density of 0.1 per square centimeter results in 69% yields while a more likely 0.06 increases it to 80%

Not everyone calculates that set of numbers the same way.

 

[Chrispy_]

Is that for N4 or N3? If it's the former, then it leads to a yield of 74.5% for perfect dies or 104 RTX 5080 candidates per wafer. A good portion of the 36 defective dies could be repurposed as 5070 Ti so the 5080 is nowhere near as costly to make as a naive analysis may indicate.

That figure was from a 4090 and H100 discussion, so 4NP I think - which is what 50-series is using, too.

 

[AnotherReader]

Huh. HUB and I actually agree on something? That is some weird stuff there.. It's not wrong. The 5080 is what was expected from the 5070. The 5080 was supposed to be about 50% the difference in performance between the 4090 and 5090 and it's not. I know they've been working on something the sits in that performance profile range, so the 5080 is a head scratcher..


That number means .08 defects per CM squared. That is an average, but also lot of defects.

0.08 per square centimeter is pretty good. AMD managed to ship out the Radeon 5870 when TSMC's 40 nm had a defect rate just below 0.4 per square cm; of course, that required an excellent understanding of why TSMC's 40 nm was a failure at first. A defect rate of 0.08 means that 74.5 % of candidates 5080 dies would be perfect. A big portion of the remaining 25% could be sold as 5070 Ti instead. N4 and N5 aren't Intel 10 or TSMC's early 40 nm; they are mature nodes with excellent parametric and functional yields.

 

[lexluthermiester]

0.08 per square centimeter is pretty good. AMD managed to ship out the Radeon 5870 when TSMC's 40 nm had a defect rate just below 0.4 per square cm; of course, that required an excellent understanding of why TSMC's 40 nm was a failure at first. A defect rate of 0.08 means that 74.5 % of candidates 5080 dies would be perfect. A big portion of the remaining 25% could be sold as 5070 Ti instead. N4 and N5 aren't Intel 10 or TSMC's early 40 nm; they are mature nodes with excellent parametric and functional yields.

Again, not everyone calculates that set of numbers the same way. But I digress...

 

[AnotherReader]

Again, not everyone calculates that set of numbers the same way. But I digress...

No, you can not use a naive model for the calculation. The standard is Murphy's yield model.

Defect clustering allows less defects over large areas of the wafer than if the defects are randomly and uniformly distributed.

 

[lexluthermiester]

No, you can not use a naive model for the calculation. The standard is Murphy's yield model.

Again, not everyone uses Murphy's yield calculations. Why? Because they don't always apply well to the modern methods for IC production. Also, your link article is under review for accuracy.

 

[AnotherReader]

Again, not everyone uses Murphy's yield calculations. Why? Because they don't always apply well to the modern methods for IC production. Also, your link article is under review for accuracy.

Thanks for noting that, but I believe we have to disagree about the applicability of Murphy's yield; iSine and others use it and they are semiconductor engineers. That's fine; reasonable people can disagree and I believe you're one of the most reasonable posters here.

Edit: the review is due to attribution of the various equations, not their accuracy:

However, there is one obscure error. You refer to Gordon Moore's yield model (Y = e-?(AD)) as "the Seeds Model" and don't give Seeds credit for the model that is his, Y = 1/(1+AD), which you call the exponential yield model.
....

R.A. Seeds should also be credited with what is now called the Bose-Einstein yield model, Y = 1/(1+AD)^k, where k is a layer-dependent factor.

 

[Scattergrunt]

Not all supers were good 2080 super and 4080 super was let down, hardly much of an improvement and I think the 5080 super will be the same with maybe 3-5% improvements

I liked them aesthetically a lot, ignoring performance. 4080 Super was nice for the price cut, the 2080 super was the epitome of a product refresh. Would agree with you on the potential 5080 super thing though..

 

[lexluthermiester]

Thanks for noting that, but I believe we have to disagree about the applicability of Murphy's yield; iSine and others use it and they are semiconductor engineers. That's fine; reasonable people can disagree and I believe you're one of the most reasonable posters here.

Ok, here's why Moore's method isn't used anymore: Binning. In the 1970's when Gordon Moore created the equation he did so based on monolithic designs that had to be perfect to be functional, IE, CPU's that could not withstand any defect and remain functional. If a defect was present, that die was a write-off and was discarded. This continued into the late 80's and even the early 90's. However, after that binning was being utilized in new ways and defects became less of a concern. By the middle of the 2000's, Moore's method was only a "best guess" calculation. From the 10's on, it's useless from a practical application stand point. So when basing a calculation of yields where defects are concerned Moore's method is not valid for modern IC wafers and has to be calculated based on actual die test-through numbers.

Saying a wafer renders a 90% yield, is that perfect dies? Is it usable incomplete dies? It's not an exact science anymore and hasn't been for decades. Moore's applied method does work in the past for old IC manufacturing. It doesn't work for the modern era of wafer manufacturing.

 

[AnotherReader]

Ok, he3re's why Moore's method isn't used anymore: Binning. In the 1970's when Gordon Moore created the equation he did so based on monolithic designs that had to be perfect to be functional, IE, CPU's that could not withstand any defect and remain functional. If a defect was present, that die was a write-off and was discarded. This continued into the late 80's and even the early 90's. However, after that binning was being utilized in new ways and defects became less of a concern. By the middle of the 2000's, Moore's method was only a "best guess" calculation. From the 10's on, it's useless from a practical application stand point. So when basing a calculation of yields where defects are concerned Moore's method is not valid for modern IC wafers and has to be calculated based on actual die test-through numbers.

Saying a wafer renders a 90% yield, is that perfect dies? Is it usable incomplete dies? It's not an exact science anymore and hasn't been for decades. Moore's applied method does work in the past for old IC manufacturing. It doesn't work for the modern era of wafer manufacturing.

That's a good point, but even today, it tells you the number of perfectly functional dies. Of course, with the dies having multiple independent cores, actual yields are better as even defective dies can be repurposed. Parametric yield, of course, is different and I have never seen any mention of it from any of the foundries.

 

[xorbe]

Is the FE edition 5080/5090 issue with pcie5 a real thing? I haven't seen anyone discussing the couple of YT videos about it.

 

[lexluthermiester]

Is the FE edition 5080/5090 issue with pcie5 a real thing?

What issue?

 

[Palladium]

I went straight to the 1440p overall FPS and power draw graphs, and I LMAO'ed.

BTW it's pretty fascinating how the bad the GPU market is, while at the same time EA/Ubisoft/Sony etc is crumbling because pushing graphical fidelity is no longer a surefire way to print $$$.

 

[purpleduggy]

no stock. not paying scalpers. Vapourware. Vapour frames. Paper launch. Will wait for RTX6090 with frame gen x8 in 2026

 

[lexluthermiester]

Jay just did some overclocking. Seems to have had a good go of it. @W1zzard What about your examples? Did you have a chance to push the limits a bit more? Or did Jay really have a "golden sample"?