News Nvidia Benchmarks Show 4080 12GB Up to 30% Slower Than 16GB Model

[BILL1957]

So far this generation feels like a really bad joke. The 4080 16GB is under performing by a good 9-12%. The 4080 12GB isn't even an 80 class card, its a 70 class card. Which means that every sku below it got bumped up too. 50->60 class, 60->70 class and as discussed 70->80 class.

Insanesly Nvidia made the RTX 4090 the best price to performance this gen. Nvidia is leaving a bad taste in my mouth. I might just suffer AMD drivers just to make a point this gen depending on RDNA 3 pricing/performance. Otherwise if this becomes the new norm in PC gaming...I might not be a PC gamer much longer. I'd hate to see tech greed kill PC gaming. But it might happen at this rate...

You have to remember that even the 16g 4080 is not on equal footing to the prior gen 80 tier card when compared to the 90 cards.
Last gen the 3080 used a slightly cut back version of the exact same chip that was used in the 3090 cards.

This 40 series the 4080 is NOT using the same chip as what is used for the 4090 but the 4080 was nerfed to permit more of a performance separation to the 4090 to more justify the 4090's much higher price point and to make the 4090 the top tier card and the 4080 16g a clear 2nd tier offering.

We all know that the 4080 12g was originally supposed to be the 4070 but with the overstock of the 30 series cards there was no way they could sell those while releasing the 40 series cards at their original intended tier and price points.
I think it is possible that we will be seeing the change to a point that a product stack will be based entirely on performance level and vram amounts and not the components that the card is configured off of.

There may not have been as much pushback if the 4080 12g had of been released more at a $750 msrp rather than at a ridiculous $900 and the 4080 16g at around $1000.

 

[helper800]

Those die sizes are correct, but the same size basically costs double what it did before. So from that perspective, plus the massive transistor density increase, bumping things up a tier for roughly equivalent size makes sense. Nvidia should get 120~142 potential GA106 chips per wafer (for 3070), versus 170~193 AD104 chips per wafer (RTX 4080 12GB and probably 4070 as well). Assuming TSMC 4N wafers cost twice as much (possibly more), that works out to 50% more per AD104 chip. And Nvidia is NOT going to eat that difference if the new part is faster; it will raise prices.

What's interesting is that, even if we assume a price of $15,000 per 4N wafer, that's still only about $78~$88 per chip. And GA106 chips on a $4000 wafer are potentially just $28~$33. But of course there are other costs (packaging, wire bonding, etc.) that those prices probably don't include. Still, the actual silicon costs likely isn't all that high. Hell, even on AD102, at 80~89 chips per wafer (some may be bad), that would be $169~$188 per chip. 😯

Essentially this proves the point that the cost increases do not justify the increased pricing on a tier for tier basis compared to last generation. The only reason pricing for 4000 series is where it is compared to tiers from previous gen is pure greed. I would not be surprised if Nvidia's profit margin on the 4080 12gb isn't at least 60%. If Nvidia makes the barrier to entry to 4000 series 900 dollars for what is essentially a XX60 tier part I would hope they lose customers faster than they make money with the increased margins.

I predict next gen the 5000 series will be even worse. a XX60 tier card will be 1200 dollars because they cannot sell all their 4000 series junk. They mind as well start making 10,000 cards a generation and put them on ebay and let the highest 10,000 bidders get the cards. This kind of price increase precedent is a race to the bottom. I wish Nvidia would run out of idiots to sell card to or we get reasonable competition.

 

[JarredWaltonGPU]

Essentially this proves the point that the cost increases do not justify the increased pricing on a tier for tier basis compared to last generation. The only reason pricing for 4000 series is where it is at comparative tiers from previous gen is pure greed. I would not be surprised if Nvidia's profit margin on the 4080 12gb isn't at least 60%. If Nvidia makes the barrier to entry to 4000 series 900 dollars for what is essentially a XX60 tier part I would hope they lose customers faster than they make money with the increased margins.

I predict next gen the 5000 series will be even worse. a XX60 tier card will be 1200 dollars because they cannot sell all their 4000 series junk. They mind as well start making 10,000 cards a generation and put them on ebay and let the highest 10,000 bidders get the cards. This kind of price increase precedent is a race to the bottom. I wish Nvidia would run out of idiots to sell card to or we get reasonable competition.

As I've said elsewhere, there are other important considerations. Nvidia needs to pay thousands of employees, and the R&D costs are not trivial for developing a new architecture. But as for the future RTX 50-series, I feel 40-series in many ways echoes the 20-series launch. Prices were increased because prior to the launch there was a GPU shortage, thanks to crypto mining. Then there were rumors Nvidia delayed the 20-series launch, with it arriving in the early fall instead of the spring, in order to clear out some existing 10-series inventory. The 30-series launch pricing was a correction from the 20-series pricing, and you all know what happened.

The other issue is that scalpers are and are like to remain a problem for major GPU launches. Everyone was pretty sure the 4090 cards would immediately sell out, even at $1,600+. They did, and naturally some of that was due to scalpers who then listed the cards on eBay for $2,500+. Without mining, that markup absolutely is not sustainable, but for the first month or two (depending on supply), there will always be people that attempt to profiteer via scalping. Nvidia will surely see this yet again, and now it has to be thinking, "You know, we could have marked up the cards to $2,000 at launch and for the first few months we would have still sold everything!"

But that's a special case of a new, significantly faster GPU arriving. I'm not at all convinced scalpers will be able to do the same thing with the RTX 4080 16GB, and especially the 4080 12GB. I already think the 12GB model is going to be overpriced for what it provides. We'll see what the test results look like, but even if you can buy one at $900 at launch, you'd need some very stupid and impatient people to be able to flip that for $1,200+ on eBay. But there always seems to be an abundance of stupid and/or impatient people soooo... who knows what will happen? I'd love to see the scalpers take a loss on 4080 cards.

 

[-Fran-]

As I've said elsewhere, there are other important considerations. Nvidia needs to pay thousands of employees, and the R&D costs are not trivial for developing a new architecture. But as for the future RTX 50-series, I feel 40-series in many ways echoes the 20-series launch. Prices were increased because prior to the launch there was a GPU shortage, thanks to crypto mining. Then there were rumors Nvidia delayed the 20-series launch, with it arriving in the early fall instead of the spring, in order to clear out some existing 10-series inventory. The 30-series launch pricing was a correction from the 20-series pricing, and you all know what happened.

The other issue is that scalpers are and are like to remain a problem for major GPU launches. Everyone was pretty sure the 4090 cards would immediately sell out, even at $1,600+. They did, and naturally some of that was due to scalpers who then listed the cards on eBay for $2,500+. Without mining, that markup absolutely is not sustainable, but for the first month or two (depending on supply), there will always be people that attempt to profiteer via scalping. Nvidia will surely see this yet again, and now it has to be thinking, "You know, we could have marked up the cards to $2,000 at launch and for the first few months we would have still sold everything!"

But that's a special case of a new, significantly faster GPU arriving. I'm not at all convinced scalpers will be able to do the same thing with the RTX 4080 16GB, and especially the 4080 12GB. I already think the 12GB model is going to be overpriced for what it provides. We'll see what the test results look like, but even if you can buy one at $900 at launch, you'd need some very stupid and impatient people to be able to flip that for $1,200+ on eBay. But there always seems to be an abundance of stupid and/or impatient people soooo... who knows what will happen? I'd love to see the scalpers take a loss on 4080 cards.

I'm sorry Jarred, but when you look at the margins nVidia is making as a whole, your whole argument loses a lot of strength.

I'm not saying you're flat out wrong from the practical and literal sense, but this is just "greed" at work with no second readings. This is nVidia wanting to become Apple. Nothing more, nothing less. Will nVidia get annoyed they could have gotten away with $2000 instead of $1600? Probably, but I'm 150% sure they've already weighted that in. At least Jensen has. They've also been very open about the used market as well and how they want to control it somehow (ugh, I can't remember the exact quote and when, so I could be wrong). Scalpers will always exist when there's scarcity and people willing to pay for the higher prices. I'm sure nVidia is balancing that somehow. Maybe they know the price is in fact a hard pill to swallow, so they just scammed scalpers with the 4090-only release in low-ish volume and then flood with 4080's close to AMD's launch (kind of like with the 3060, I'd like to say?). They could even drop prices sacrificing margins, but that's never the case with them, I think? I mean, the 3090ti was a $2K card and now is going for half or less than that. And I think, nVidia at least, is still making a profit with those or breaking even somehow; accounting shenanigans always work as you can write a lot of those in marketing expenses and other such things.

So, all in all, it's not about how that money is being used by nVidia (cost structure or BOM), but just know that they have absurdly good margins (reportedly even?) on anything they sell and they want to keep it that way I'm sure. I'd love to be proven wrong, but I doubt nVidia would be clearcut on their costs and we can only get approximations from AIBs, earning reports or leaks. Maybe this would be something interesting to investigate for Tom's Hardware? You may make nVidia angry for not adhering to their "guidelines", though

Regards.

 

[atomicWAR]

You have to remember that even the 16g 4080 is not on equal footing to the prior gen 80 tier card when compared to the 90 cards.
Last gen the 3080 used a slightly cut back version of the exact same chip that was used in the 3090 cards.

Yeah I am all to aware. This is the gtx 680 all over again. Use a 103 die which is typically a 70 series as 80 series because the performance jump is so big compared to previous gen that old gen stock won't sell (or that is their worry).

Regardless its a lame move by Ngreedia. It has me truly bent as a PC gamer of over 30 years and Nvidia user since Tnt 2... Not to say I didn't use ATI/AMD as well on some budget builds.

 

[JarredWaltonGPU]

Yeah I am all to aware. This is the gtx 680 all over again. Use a 103 die which is typically a 70 series as 80 series because the performance jump is so big compared to previous gen that old gen stock won't sell (or that is their worry).

Regardless its a lame move by Ngreedia. It has me truly bent as a PC gamer of over 30 years and Nvidia user since Tnt 2... Not to say I didn't use ATI/AMD as well on some budget builds.

Wow! Nvidia actually heard the unhappiness!

Unlaunching The 12GB 4080

16GB GeForce RTX 4080 on track to delight gamers everywhere November 16th.

www.nvidia.com

Article coming shortly... but, what people really want is a lower 4080 16GB price of $899, and then the formerly 4080 12GB can launch as a 4070 at $599.

 

[helper800]

Wow! Nvidia actually heard the unhappiness!

Unlaunching The 12GB 4080

16GB GeForce RTX 4080 on track to delight gamers everywhere November 16th.

www.nvidia.com

Article coming shortly... but, what people really want is a lower 4080 16GB price of $899, and then the formerly 4080 12GB can launch as a 4070 at $599.

That would be a great way to secure at least some good will with me. Thanks for the update, I look forward to the article!

 

[Marv_Ellis]

The 80 series cards have historically "5 years ago" been either a 256 or 384 bus width rate depending if it was a non TI or TI version, they have now pushed the 384 bus width to the 90's series, 256 bus to the 80 series, 192 series to the 70's series and 128 to the 60's.

that lower end 3080 12 gig card has a 192 bus width which essentially makes it a 70 series card, with boosted frequency equal to the 80 series which should make it a TI version of the 70's series

this all started to change with 80's TI's cards some years ago with the 384 bus width.

Essentially im sure as others have stated a exec who has no clue how it works decided to get greedy, there is obvious reasons the 16 gig version is around 30% slower because it has 33% less bus-width.

 

[JarredWaltonGPU]

The 80 series cards have historically "5 years ago" been either a 256 or 384 bus width rate depending if it was a non TI or TI version, they have now pushed the 384 bus width to the 90's series, 256 bus to the 80 series, 192 series to the 70's series and 128 to the 60's.

that lower end 3080 12 gig card has a 192 bus width which essentially makes it a 70 series card, with boosted frequency equal to the 80 series which should make it a TI version of the 70's series

this all started to change with 80's TI's cards some years ago with the 384 bus width.

Essentially im sure as others have stated a exec who has no clue how it works decided to get greedy, there is obvious reasons the 16 gig version is around 30% slower because it has 33% less bus-width.

Bus width likely won't be nearly as critical with the addition of a much larger L2 cache, at least for a lot of workloads.

 

[Awev]

As AMD first showed, and just recently Intel, if you use ReBAR/Smart Memory Access you are able to overcome a lot of - not all of - the performance penalties of a smaller bus width.

 

[Giroro]

If true, the 4090 was built as a halo product to prove a point! In the lower tiers, AMD may obliterate the 4080 class cards...

All halo products exist to prove a point, and that point is always to trick customers into believing that an overpriced "middle" product is a "good value".
Humans always consider options on a relative scale, and we always land in the middle.
Olive garden knows they won't often sell their $140 bottle of wine nor their $45 bottle of house red. They do it to convince people its normal to pay $54 for a $15 bottle of Apothic or Zin. It's cold, cynical, sociopathic mindgames.
And it's worked shockingly well. We have people begging for an $800 luxary-priced high end, high margin 4080. It's partly because Nvidia has oversaturated the "indie tech" media with guerrilla marketing and LTT style astroturf shills.. But also because there is real perception that their still-severely overpriced product would be "discounted", and by extension a good deal. Nvidia's Halo products really do convince some people that it's reasonable to spend $800 on a gaming GPU. Entertainment, not necessity. A gpu that plays the same games as the last gen, or a console. It's not like there's some exclusive content, groundbreaking technology, or killer app.
Only a couple years ago, your average PC gamer completely balked at the idea of spending over $300 on a gpu.

Nvidia has successfully convinced some people that a "middle" GPU should cost more than every other part of your high end PC, combined. They make the market, and all other options have been eliminated from their product line. They've totally changed the narrative.

It's gross. This is why Monopolies are bad. We are dealing with multiple layers of monopoly between Nvidia's IP and TSMC's stranglehold on silicon production.

 

[helper800]

All halo products exist to prove a point, and that point is always to trick customers into believing that an overpriced "middle" product is a "good value".
Humans always consider options on a relative scale, and we always land in the middle.
Olive garden knows they won't often sell their $140 bottle of wine nor their $45 bottle of house red. They do it to convince people its normal to pay $54 for a $15 bottle of Apothic or Zin. It's cold, cynical, sociopathic mindgames.
And it's worked shockingly well. We have people begging for an $800 luxary-priced high end, high margin 4080. It's partly because Nvidia has oversaturated the "indie tech" media with guerrilla marketing and LTT style astroturf shills.. But also because there is real perception that their still-severely overpriced product would be "discounted", and by extension a good deal. Nvidia's Halo products really do convince some people that it's reasonable to spend $800 on a gaming GPU. Entertainment, not necessity. A gpu that plays the same games as the last gen, or a console. It's not like there's some exclusive content, groundbreaking technology, or killer app.
Only a couple years ago, your average PC gamer completely balked at the idea of spending over $300 on a gpu.

Nvidia has successfully convinced some people that a "middle" GPU should cost more than every other part of your high end PC, combined. They make the market, and all other options have been eliminated from their product line. They've totally changed the narrative.

It's gross. This is why Monopolies are bad. We are dealing with multiple layers of monopoly between Nvidia's IP and TSMC's stranglehold on silicon production.

This is essentially what I have been saying since the 2080 ti came out and all of a sudden a 70 class card was 500 dollars...

 

[JarredWaltonGPU]

As AMD first showed, and just recently Intel, if you use ReBAR/Smart Memory Access you are able to overcome a lot of - not all of - the performance penalties of a smaller bus width.

I don't know that ReBAR has much do do with overcoming a narrower bus on Intel, considering Intel's memory bus isn't particularly narrow at 256-bits for the A770/A750 (and soon A580). It's all about the Infinity Cache for AMD's RDNA2 cards.