A 4070 SUPER is better than a 3090? Wow. Nvidia showcase SUPER GPUs

[thorhammerz]

6 hours ago, Deadpool2onBlu-Ray said:

I don't know why people are so optimistic that prices will go down next gen

They remember a time when a certain tier of products were sold at another (in this case, lower) price point. Some however, believe they are entitled to the products at the old price point. Look no further than the poetic waxings on how the vendors/manufacturers are all immoral for changing the price against their favor.

[starsmine]

1 hour ago, thorhammerz said:

They remember a time when a certain tier of products were sold at another (in this case, lower) price point. Some however, believe they are entitled to the products at the old price point. Look no further than the poetic waxings on how the vendors/manufacturers are all immoral for changing the price against their favor.

yea but next generation has two major problems. 3nm costs more per mm^2. 3nm has ZERO SRAM scaling. So if nvidia wants to increase cache and be monolithic, there is no way to do that but cut into logic die space making it a smaller uplift in core count then otherwise could have been at similar die size.

There is no chance of Blackwell being cheaper then ada at a given tier. the best you can hope for is stagnation from pascal pricing, as in Blackwell 104 full die being 700 even though buy all metrics it will cost more to manufacture (even inflation adjusted) then the GTX 1080. 
I do want @IkeaGnometo explain why he thinks what I said is wrong though. 

[RejZoR]

12 hours ago, Deadpool2onBlu-Ray said:

5080 will not be $799. Those ships sailed long ago. You're only upgrade path for south of $799 will likely be the "5070". I don't know why people are so optimistic that prices will go down next gen

If people weren't so enthusiastically be buying them, they would. But hey, charge away at those 4090's at 1600€ and more...

[IkeaGnome]

4 hours ago, starsmine said:

I do want @IkeaGnometo explain why he thinks what I said is wrong though. 

I what? I had to double check. I haven’t posted in this thread before now. 

[thorhammerz]

9 hours ago, starsmine said:

yea but next generation has two major problems. 3nm costs more per mm^2. 3nm has ZERO SRAM scaling. So if nvidia wants to increase cache and be monolithic, there is no way to do that but cut into logic die space making it a smaller uplift in core count then otherwise could have been at similar die size.

There is no chance of Blackwell being cheaper then ada at a given tier. the best you can hope for is stagnation from pascal pricing, as in Blackwell 104 full die being 700 even though buy all metrics it will cost more to manufacture (even inflation adjusted) then the GTX 1080. 

I'd imagine die-size is important moreso for how it affects yields (rather than raw production cost out of the fab, which is low relative to the sticker price), as that directly impacts the amount of supply available for any given configuration / product tier. Though, I suppose that also depends on how wide the full wafer of silicon is.

[Senzelian]

Cool. Prices still suck.
I will keep that in mind when I will upgrade to another AMD GPU. That will make the purchase even more enjoyable.

 

 

5 hours ago, IkeaGnome said:

I what? I had to double check. I haven’t posted in this thread before now. 

Doesn't matter. It's all your fault. 

[porina]

1 hour ago, thorhammerz said:

I'd imagine die-size is important moreso for how it affects yields (rather than raw production cost out of the fab, which is low relative to the sticker price), as that directly impacts the amount of supply available for any given configuration / product tier. Though, I suppose that also depends on how wide the full wafer of silicon is.

Wafer sizes are standard on the leading nodes at 12"/300mm. Cost per wafer is speculated to be 20k for N3, 16k for N5, 10k for N7. We may never know the real numbers and each major customer is likely to be able to negotiate their own special rates rather than go off a price list. There is a 25% increase in cost per area going from N5 to N3, meaning for the same die size expect that component cost to go up correspondingly. Alternatively you could go to a smaller die to offset cost. Given the increase in density is greater than increase in area cost, it makes the potential transistor count per cost increase by 28%. Nvidia uses custom 4N based off N5 so presumably costs are similar. If this is remotely correct, the move to N3 from N5 class seems a no-brainer.

 

I'm not in the right mind state to do it right now, but it should be possible to do a "cost per transistor" of GPUs through recent generations and see if there is a trend there.

[starsmine]

17 minutes ago, porina said:

Wafer sizes are standard on the leading nodes at 12"/300mm. Cost per wafer is speculated to be 20k for N3, 16k for N5, 10k for N7. We may never know the real numbers and each major customer is likely to be able to negotiate their own special rates rather than go off a price list. There is a 25% increase in cost per area going from N5 to N3, meaning for the same die size expect that component cost to go up correspondingly. Alternatively you could go to a smaller die to offset cost. Given the increase in density is greater than increase in area cost, it makes the potential transistor count per cost increase by 28%. Nvidia uses custom 4N based off N5 so presumably costs are similar. If this is remotely correct, the move to N3 from N5 class seems a no-brainer.

 

I'm not in the right mind state to do it right now, but it should be possible to do a "cost per transistor" of GPUs through recent generations and see if there is a trend there.

only logic will go up in density is a major issue, Memory controllers (analog) and cache are the same density, so there is no way for transistor density to go up 28% with this next node. 
 

  

5 hours ago, IkeaGnome said:

I what? I had to double check. I haven’t posted in this thread before now. 

you reacted but didnt expand on my post earlier. 

[porina]

1 hour ago, starsmine said:

only logic will go up in density is a major issue, Memory controllers (analog) and cache are the same density, so there is no way for transistor density to go up 28% with this next node. 

I recall seeing the lower SRAM scaling earlier, and a reference here suggests only 5% shrink going from N5 to N3. I don't recall seeing similar about analog. Based on this, GA102 L2 cache is only 7%. I haven't found similar data for Ada yet. So even if it isn't the best case for logic, the gains in density should still help out offsetting silicon costs.

[IkeaGnome]

6 hours ago, starsmine said:

you reacted but didnt expand on my post earlier. 

Which one? 
 

 

6 hours ago, Senzelian said:

Doesn't matter. It's all your fault. 

No. It’s even better. I’m 99% sure I just got pulled into a new thread over something I did at least a week ago. I went back 10 pages in “my reputation” which shows reactions given and received and I didn’t see where I’ve interacted with that user. 
 

10 pages put me into October..

 

 

Edit: Found it. Not sure that was me disagreeing. 

[filpo]

Can't add new vids to my post with reviews (for some reason) so making this post for the 4080 SUPER review, and it's not looking too bad

TLDR: Essentially a 7900 xtx in rasterization (slightly behind) but GN shows that it's value is quite good

NVIDIA GeForce RTX 4080 Super Founders Edition Review - Helping you Save $200 | TechPowerUpMSI GeForce RTX 4080 SUPER Expert 16 GB GPU Review - The Mercedes of GPUs (wccftech.com)

Nvidia RTX 4080 Super review: $999 is the main feature - The Verge

Nvidia GeForce RTX 4080 Super review: second only to the RTX 4090, and finally worth buying | TechRadar

Nvidia GeForce RTX 4080 Super Founders Edition review | PC Gamer

Nvidia GeForce RTX 4080 Super Review - IGN

Nvidia GeForce RTX 4080 Super review: Slightly faster than the 4080, but $200 cheaper | Tom's Hardware (tomshardware.com)

[porina]

Looking at TechPowerUp's written review, my take away points are:

4080S is 1% faster than 4080, despite having more cores and faster VRAM. Might be power limited?

7900XTX slightly ahead of 4080S on average by 1% to 4% outside of RT, where it is 20% slower.

4080S is available at MSRP in UK (£959), in stock as I write. 7900XTX cheapest in stock at moment is £870 from ebuyer, then £895 from a place I never heard of before and going up from there.

 

Edit: that didn't last long. MSRP models are gone. More expensive ones remain.

[thechinchinsong]

1 hour ago, porina said:

Looking at TechPowerUp's written review, my take away points are:

4080S is 1% faster than 4080, despite having more cores and faster VRAM. Might be power limited?

7900XTX slightly ahead of 4080S on average by 1% to 4% outside of RT, where it is 20% slower.

4080S is available at MSRP in UK (£959), in stock as I write. 7900XTX cheapest in stock at moment is £870 from ebuyer, then £895 from a place I never heard of before and going up from there.

 

Edit: that didn't last long. MSRP models are gone. More expensive ones remain.

The FE models were available for a couple hours this morning on Nvidia's website and Best Buy for MSRP but they're sold out too now. Same story with some models 40-100$ over MSRP being available. This is in the continental US.

[Likwid]

On 1/12/2024 at 7:48 PM, filpo said:

additional models within the range off 40 series

If I'm not mistaken most of the performance uplift of the 4070 Ti SUPER is going to 16GB of vram

It's a bit like they did with the 20 series, a mild increase. Not so big it's a Ti but not just an addition of VRAM (in the 4070 Ti SUPER's case)

better yet a quattro is actually an add on to their cars (four wheel drive)

 

but tbh this could go for all car companies. For example you've got BMW with their 318i, 320i, 330i, 330e, M340i and then M3

Too bad there are Torsen (you want this one) and Haldex but good luck knowing that
And I really don't like newer car engine names 30D is 2.0 tdi  and 50D is 3.0tdi

[Stahlmann]

On 1/31/2024 at 6:19 PM, porina said:

Edit: that didn't last long. MSRP models are gone. More expensive ones remain.

There you have it. That's the reason these cards are priced like they are. People are buying them.

[starsmine]

48 minutes ago, Stahlmann said:

There you have it. That's the reason these cards are priced like they are. People are buying them.

They also are largely not stocked.

Personal computing Hardware sales have been down hard post pandemic. So production runs are MUCH lower then historically expected. 
It means more that Nvidia projection on sales numbers were solid more then people are buying them. Sales are still down. They just shipped out the correct supply to match the downed sales.

[porina]

1 hour ago, Stahlmann said:

There you have it. That's the reason these cards are priced like they are. People are buying them.

You can say that about any product. It isn't viable unless people buy it.

 

Looking right now, the cheapest in stock 4080S I can find in UK is 2% above MSRP. Close enough.

 

That the cheaper models sell faster may be in part a sign that many people don't value additional build quality or other physical features, if it is there at all.

 

Been looking at Steam Hardware Survey. January results seem to be a normal month. That is, no unusual shifts in the numbers. 4080 comes in a 0.75%, behind the 0.93% of the 4090. Most popular current gen desktop GPU is 4070 at 1.55%, with 4060 Laptop most popular in the mobile space at 2.61%. AMD still don't have anything making the cut apart from the 7900XTX at 0.34%. Intel who? I guess the question now is, how long will it take for the Supers to make the list? Presuming the outgoing models will decline in share as new sales drop off.

[thorhammerz]

10 hours ago, porina said:

You can say that about any product. It isn't viable unless people buy it.

Markets / auctions are beautifully simple in that respect. You emplace an offer. Someone responds with a bid. If bid matches offer, product changes hands.

 

Management of inventory (such that supply never quite drowns out demand at any point in time) is just a prudent move on NVDA's part to keep prices in their favor (or at least, stable) when there are no competitors trying to flood that particular product space with their own supply.

[Monkey Dust]

6 hours ago, porina said:

That the cheaper models sell faster may be in part a sign that many people don't value additional build quality or other physical features, if it is there at all.

 

Apart from avoiding a random brand you've never heard of, it's rarely worth stepping up from the cheapest version, IMO.

 

The performance improvement for the more expensive versions is only noticeable if you really look for it, and have a stock clocked card to compare too. The more expensive versions might run a little cooler and quieter, but are you really going to notice when it's in your case? The cheaper card is more likely to have a less absurdly large cooler, too.