$2500 5090 I’m out

[Fluxmaven]

They do the same thing every generation and somehow people still get excited for it... People were panic selling selling 2080Ti's when the "just as good" 3070 was coming out for $500. Same for 3090s and I'm sure there will be people that sell 4090s for cheap. We're a long way out from 5070's being readily available. I'm sure once they actually are out and proper benchmarks are available, it will seem like a fairly average card to most people. The scenarios where it's "Just as good" as a 4090 will likely be few and far between.

At $2500 the 5090 was a pass for me. At $2000, I might be convinced to pick one up. Especially if the scalpers that buy them opening day can't unload them and they end up in the open box section at Microcenter for a little discount.

 

[Sir Beregond]

I guess the one thing I noticed was the caveat of pushing the performance....with AI. Will be interested to see real world performance reviews outside very specific scenarios Nvidia used.

One thing of note especially is that some of Ada's huge improvements over Ampere came from the move to TSMC 5nm from Samsung 8nm. Given Blackwell is still on TSMC 5nm, it seems more akin to the move from Kepler to Maxwell (both on 28nm). All the improvements to traditional raster is going to come from architecture improvements, larger power draw, and definitely in the case of the 5090, larger die. As for RT and the rest, well presumably all this AI / DLSS4 so they will get a larger improvement, and I think that was kind of expected in the lead up to this gen.

EDIT: Forgot to mention move from G6X to G7 should be a big uplift as well to the overall memory bandwidth.

I am happy they didn't raise prices again. I fully expected $1199 5080. As for me...I'm going to wait and see. I still would prefer a 24GB card for my 4k display as an upgrade to my 12GB card. But...my test bench could use a modern GPU, so will see how the 5070 or 5070 Ti looks, or maybe a used 40-series.

 

[WilyKit]

Barely at all, unless you're starting from under 60fps. It kicks butt!

Under 60 is where you need extra frames most though.

 

[Sickb0y]

I can’t even imagine what the scalper price is gonna be lol

Same as a down payment on a house lol

 

[WilyKit]

4070 base clock = 1.9 GHz
5070 base clock = 2.2 GHz

4070 bandwidth = 504 GB/s
5070 bandwidth = 672 GB/s

That's a 32.8% increase in cores over the 4090.

Which is a low uptick gen over gen. A 4090 has 56% more CUDA cores than a 3090 and 52% more than a 3090Ti

 

[spine]

One thing of note especially is that some of Ada's huge improvements over Ampere came from the move to TSMC 5nm from Samsung 8nm. Given Blackwell is still on TSMC 5nm, it seems more akin to the move from Kepler to Maxwell (both on 28nm). All the improvements to traditional raster is going to come from architecture improvements, larger power draw, and definitely in the case of the 5090, larger die.

This. Nvidia themselves haven't mentioned anything about improved efficiences in the traditional render path. So all we really have here is that the power usage has gone up, along with increased memory speeds at presumably low wattage to allow more for the GPU additionally. I don't see any reason to believe 10,000 5th Gen cores will significantly outperform 10,000 4th gen cores running at the same speed. Lower wattage, sure, but who cares.

 

[killroy67]

Staying put with my 4080 is sounding much more appealing than spending over $1000 for a 5080, especially with the same 16gb of memory I have now. Currently I can play Stalker2 at 4k and most of the settings cranked to epic using DLSS3.5 and get 100ish FPS. With the support of DLSS4 that will improve even more, Stalker2 is one of the games to get DLSS4 support. That said, worst case I could always give up on 4k gaming, and hook up my Dell 2k monitor if I found my 4080 to be really struggling, So who knows, I might be able to get another several years or more of high end gaming with the 4080, which might get me to being 60 years old, at which point I may just throw in the towel on PC gaming after more than 30 years.

 

[The Cobra]

 

[Zorachus]

I think the only reason to get a RTX-5090 would be if you run a 4k monitor and you want a constant 144 fps and higher with all settings Ultra maxed out.

Or that upcoming 5k Ultrawide monitor from LG, yeah just remembered that thing is coming this year. Ok that monitor probably would need a 5090 to run games smoothly.

But man, that will run you like $4,500 - $5,000 for the card and monitor and tax. Ouch too much for my blood, but would be sweet. But 5 grand for just a video card and new monitor is big bucks.

I recall just like 3 or 4 years ago, buying a brand new RTX-3080 few months after release at Microcenter for $850, and the best OLED monitor available at the time, the AW3423DW for $1199, so both together were like $2 grand or so.

 

[Sir Beregond]

 

[undertaker2k8]

Anyhow, selling my 4090 for 1570 and getting an OB 4080s for 780ish net was still a good move. I think the 5070 Ti is roughly going to equal a 4080s.

 

[Zarathustra[H]]

Same bullshit different day.

Previous Gen: $1,599
Actual (after AIB premiums and horrendous markups and scalping) $2,299

Rumor $2,500.
Response: OMG, Shock and horror. Gaming is over. The end is coming. No one will be able to afford PC gaming anymore. I'm calling it quits.

Actual next gen: $2,000
Response: Oh, that's not as bad as we feared. OK then.

Maybe not as bad as we feared, but it is still a massive 25% MSRP increase in only 2 years. it IS pretty bad.

And what it actually going to wind up costing when the board partners, retailers and scalpers are done with it? $2,500 was probably pretty accurate TBH.

I feel like this is a purposeful strategy to let people get out all of their rage on a straw man of what could be, and minimize their rage at the real bad thing that actually happened. And it happens just about every release.

 

[Zarathustra[H]]

I think the 5070 Ti is roughly going to equal a 4080s.

I'm still not convinced that their 4090 comparison is just based on DLSS4 AI fakery/lies just like they did when the original DLSS launched and all of their comparisons were old cards without DLSS vs new cards with DLSS.

I just don't think there is enough of a perf/watt improvement in the process (TSMC 4N vs. TSMC 4NP) to support these claims otherwise. AI trickery is the only viable explanation.

4080 - as you suggest - might actually be feasible.

Though I'll happily be proven wrong come review time.

Show me head to head comparisons with all DLSS2/3/4 turned off. That's the only metric that matters.

 

[Sir Beregond]

Same bullshit different day.

Previous Gen: $1,599
Actual (after AIB premiums and horrendous markups and scalping) $2,299

Rumor $2,500.
Response: OMG, Shock and horror. Gaming is over. The end is coming. No one will be able to afford PC gaming anymore. I'm calling it quits.

Actual next gen: $2,000
Response: Oh, that's not as bad as we feared. OK then.

Maybe not as bad as we feared, but it is still a massive 25% MSRP increase in only 2 years. it IS pretty bad.

And what it actually going to wind up costing when the board partners, retailers and scalpers are done with it? $2,500 was probably pretty accurate TBH.

I feel like this is a purposeful strategy to let people get out all of their rage on a straw man of what could be, and minimize their rage at the real bad thing that actually happened. And it happens just about every release.

Not to play devil's advocate here, but I think it was to be expected though if Nvidia minimally wanted to maintain their same margins. 4090 -> 5090 = biggest die they've ever done in a GeForce card period. And move from 24GB G6X to 32GB G7. Minimally that was going to raise the manufacturing cost of the 5090 over the 4090. I'd suspect $1999 just maintains the same margin they had with the $1599 4090.

EDIT: I am wrong, TU102 was ever so slightly larger. But the difference is it wasn't on a cutting edge new node and was instead a half-stepping of TSMC 16nm at the time.

 

[Zarathustra[H]]

Not to play devil's advocate here, but I think it was to be expected though if Nvidia minimally wanted to maintain their same margins. 4090 -> 5090 = biggest die they've ever done in a GeForce card period. And move from 24GB G6X to 32GB G7. Minimally that was going to raise the manufacturing cost of the 5090 over the 4090. I'd suspect $1999 just maintains the same margin they had with the $1599 4090.

EDIT: I am wrong, TU102 was ever so slightly larger. But the difference is it wasn't on a cutting edge new node and was instead a half-stepping of TSMC 16nm at the time.

Hadn't had a chance to look at the details of die size yet, but I guess that makes at least a little bit of sense. Its not a full die though is it? It's still some sort of binned part with disabled sections based on the full die that goes into eye-wateringly expensive AI/Enterprise compute products? Because then, cost is a bit relative. We are getting the disregarded refuse not good enough for AI customers, and are being charged thousands for the privilege.

One of the only ways they were ever going to hit perf/watt targets was to go with a bigger die, clocked lower at more optimal voltages tunes for perf/watt than the type of high voltage, high core clock screamers we have seen in the past.

In a way this reminds me of my 6GB Titan from 2013 which ran at a base clock of 837Mhz, an average boost of 876Mhz and a max boost of 993Mhz.

Its one of the many strategies in chip design to make the most out of a current gen when you are riding up against process node limits (and die size/cost is not an object)

 

[Zorachus]

I like how people are saying $2500? No way I'm out. Now oh it's "only" $1999 I'm totally in?

WTF

That still way too expensive for a video card, when historically they have always been under $1000 for the top tier until just recently, now they 100% higher, and "inflation" hasn't made costs go up 100%, maybe 25% at most.

The 780 Ti was $699
The 980 Ti was $649
The 1080 Ti was $699
The 2080 Ti was $999
The 3090 was $1499 ( why such a massive price increase ? )
The 4090 was $1599
The 5090 is $2000 why not $1699?

So in around ten years video cards have TRIPLED in price. Other high-end products have not tripled whatsoever in that time, cars have gotten more expensive but certainly not triple the cost from 10 years ago, or OLED TV's have not tripled in price, or CPU's have not tripled in price. Or smartphones have not tripled in price. Only video cars have tripled.

 

[Brackle]

This was taken from Nvidia's own website........

Yes you get better frames with frame gen and lowering your image quality, but lets not beat around the bush that the 5090 is going to be what 35-45% faster overall?

 

[Sir Beregond]

Hadn't had a chance to look at the details of die size yet, but I guess that makes at least a little bit of sense. Its not a full die though is it? It's still some sort of binned part with disabled sections based on the full die that goes into eye-wateringly expensive AI/Enterprise compute products? Because then, cost is a bit relative. We are getting the disregarded refuse not good enough for AI customers, and are being charged thousands for the privilege.

One of the only ways they were ever going to hit perf/watt targets was to go with a bigger die, clocked lower at more optimal voltages tunes for perf/watt than the type of high voltage, high core clock screamers we have seen in the past.

In a way this reminds me of my 6GB Titan from 2013 which ran at a base clock of 837Mhz, an average boost of 876Mhz and a max boost of 993Mhz.

Its one of the many strategies in chip design to make the most out of a current gen when you are riding up against process node limits (and die size/cost is not an object)

True. It appears to be a cut-down of GB202.

If TPU GPU DB is correct, full GB202 is 24576, while 5090 is 21760. At least currently, there doesn't appear to be any other GB202 products besides the 5090, yet.

 

[Gideon]

This was taken from Nvidia's own website........

Yes you get better frames with frame gen and lowering your image quality, but lets not beat around the bush that the 5090 is going to be what 35-45% faster overall?

View attachment 702316

Will see when the reviews come out, but I feel your estimate is on the high side. Also that frame rate still sucks for a 2,000 dollar card, I mean 27 vs 20 is still just bad.

 

[Brackle]

Will see when the reviews come out, but I feel your estimate is on the high side. Also that frame rate still sucks for a 2,000 dollar card, I mean 27 vs 20 is still just bad.

Considering they were using a 14900k which isnt the top CPU, there is room for improvement? Maybe another 2-3 fps with a 9800x3d? lol

 

[Zarathustra[H]]

This was taken from Nvidia's own website........

Yes you get better frames with frame gen and lowering your image quality, but lets not beat around the bush that the 5090 is going to be what 35-45% faster overall?

View attachment 702316

I mean, that's a solid 35%.

If it holds true come review time (vendors haven't exactly been great at picking representative comparisons for their own products in the past, because, - you know all marketing is really just professional lying) that's nothing to scoff at. Sure beats the 2%-5% generation over generation improvements Intel was giving us for a decade+ there.

That said, is it worth plunking down $2k for? Eeeh? Not sure yet.

Though if you are a high end graphics enthusiast like myself, it's not like anyone else is providing this level of performance right now, what with AMD completely having dropped out of the high end (at least this generation) and Intel promising they will launch a product in that space, but not having anything materialize yet.

I made this decision for myself when I embraced 4k back in 2015. I knew then that I would be chasing acceptable framerates with even the highest end GPU's for years, just like I had done the last time I was an early adopter of a new high resolution (2560x1600 in 2010). Maybe it was a mistake, but the experience has been amazing.

So after much protesting, chances are good I'll still go for a 5090, but one has to vent a little first, ya know?

 

[Zarathustra[H]]

Will see when the reviews come out, but I feel your estimate is on the high side. Also that frame rate still sucks for a 2,000 dollar card, I mean 27 vs 20 is still just bad.

Oh but just turn on DLSS2 at the "performance" setting and render at 1080p scaled up to 4k. Then turn on DLSS4 and turn every frame into 4 frames.

You won't notice the difference. Promise, bro.


All of that said, I can't tell what that is they are rendering in those screenshots. With that level of performance, my guess would have to be Cyberpunk 2077 in Path Tracing mode.

Path Tracing brings anything to its knees. It is really more of a technology demonstration tool than a playable feature at this point.

 

[Brackle]

You claimed you only run 4k native without dlss in the other thread where we were discussing things and you bailed .

I never claimed that. You said I did without me saying I did not run dlss.

Please stop spreading fud lol.

 

[Kardonxt]

I like how people are saying $2500? No way I'm out. Now oh it's "only" $1999 I'm totally in?

WTF

That still way too expensive for a video card, when historically they have always been under $1000 for the top tier until just recently, now they 100% higher, and "inflation" hasn't made costs go up 100%, maybe 25% at most.

The 780 Ti was $699
The 980 Ti was $649
The 1080 Ti was $699
The 2080 Ti was $999
The 3090 was $1499 ( why such a massive price increase ? )
The 4090 was $1599
The 5090 is $2000 why not $1699?

So in around ten years video cards have TRIPLED in price. Other high-end products have not tripled whatsoever in that time, cars have gotten more expensive but certainly not triple the cost from 10 years ago, or OLED TV's have not tripled in price, or CPU's have not tripled in price. Or smartphones have not tripled in price. Only video cars have tripled.

I agree pricing sucks but in all fairness the 20XX series was the last time NVidia released a Titan branded GPU. Your comparison would be more accurate continuing with the 80 series GPUs as the 90s essentially replaced Titans. (In hindsight the rebrand was pretty clever. Gamers were quick to write of Titans as rich power user cards, but happy to buy the same cards when the model number indicated top of the gaming product stack. )

This is why I jumped on the 3080 FE. $699 was an unexpected return to normalcy after 20XX series pricing.

 

[KickAssCop]

Why can’t we pre order this shit in 2025. All this AI bullshit and we can’t even have consumer buy products in a civil fashion especially across the world. . AI should be able to detect bots and crap.

Even Sony and Microsoft sorted out their pre orders for release of XSX and PS5 and that was during covid or something.

 

[criccio]

Hot take: Some of you don't want to hear it, but the quieter and quieter group that in 2025 still only cares about native raster performance is slowly going to go away. Either get on board or I guess just moan on forums until you give up gaming.

 

[Zorachus]

Why can’t we pre order this shit in 2025. All this AI bullshit and we can’t even have consumer buy products in a civil fashion especially across the world. . AI should be able to detect bots and crap.

Even Sony and Microsoft sorted out their pre orders for release of XSX and PS5 and that was during covid or something.

Good point, you can pre-order a Google Pixel or Apple iPhone the day they announce it, and then releases a few weeks later, and ships out. Why can't nVidia do the same with a video card?

 

[Furious_Styles]

Oh but just turn on DLSS2 at the "performance" setting and render at 1080p scaled up to 4k. Then turn on DLSS4 and turn every frame into 4 frames.

You won't notice the difference. Promise, bro.


All of that said, I can't tell what that is they are rendering in those screenshots. With that level of performance, my guess would have to be Cyberpunk 2077 in Path Tracing mode.

Path Tracing brings anything to its knees. It is really more of a technology demonstration tool than a playable feature at this point.

I'd say it's playable for some. For me I tried it on Indy with PT turned on (was typically around 40-60fps) and the extra lag it added didn't make it worth using. In the end I had just turned PT and it off so I could get a much better FPS. But I've always been one to prefer a smoother, higher FPS over the best visuals.

 

[Kardonxt]

This was taken from Nvidia's own website........

Yes you get better frames with frame gen and lowering your image quality, but lets not beat around the bush that the 5090 is going to be what 35-45% faster overall?

View attachment 702316

These numbers are also with Full RT. For all we know the uplift could be entirely from improved RT performance.

I'm definitely interested to see what real benchmarks show.

 

[Zarathustra[H]]

You claimed you only run 4k native without dlss in the other thread where we were discussing things and you bailed .

If he is anything like me, he strongly prefers to not use anything DLSS, but when push comes to shove reluctantly turns it on, in as limited a capacity as possible.

We are often forced to make choices we don't want to make.

In my experience DLSS2 does actually look fairly decent at the "quality" setting. Is it as good as the real thing? No. Would I run it if I didn't have to to exceed minimum frame rates of 60fps? Also no. But when push comes to shove, I'll experiment with it. I'll drop a few "ultra" settings down to "high" and run them native res, and then compare that to Ultra with DLSS2-performance on and then select the one I think looks best, all while cursing under my breath.

I obviously have not had a chance to try DLSS4 yet, but as far as DLSS3 goes, the experience runs the gambit in my book.

Sometimes, in some titles it looks gorgeous and smooth. In other titles it can introduce weird artifacts and stutters.

In general, even when it looks smooth, unless you already have a reasonably decent frame rate (at least a minimum of 60fps+) before turning DLSS3 on, as soon as you grab that mouse, you feel it. Input lag is is going to be just as bad as it was at your native DLSS3 frame rate.

To be fair to Nvidia, this IS better than the old school frame interpolation tech TV's used to (still do?) have, as it would introduce a stupid amount of input lag, as it waited for the next frame, then processed interpolation between the two frames, and only then released the frames to the screen. This resulted in smooth motion, with way worse input lag than the native frame rate.

Through AI magic DLSS3 has seemingly done away with the need to wait for the next frame, and instead only suffers the same input lag as at the native frame rate.

While this is a pretty impressive improvement, it still means that the only time DLSS3 is useful is when you already have an acceptable frame rate anyway.

I already had an acceptable frame rate in Dying Light 2 when the patch was released to enable DLSS3. I turned it on just for testing, and it was actually not bad. Looked very smooth, and input lag did not suffer at all. (input lag probably was not as good as if the frame rate had been native, but at least this was pretty much "free" visual smoothness.

On the flip-side, when Stalker2 launched, I downloaded it and tested the prologue. My Threadripper 3960x was just completely unable to keep up with the game. I was CPU limited at like 52-53 FPS. I tried enabling DLSS2 in Quality mode and it didn't help at all, confirming I was CPU limited, not GPU limited,

Then I turned on DLSS3 frame generation and what appeared on the screen really did look very smooth, but mouse feel was still that of a game at 50fps.

So yeah, frame gen - even the fancy modern AI frame gen that doesn't have to wait for a future frame in order to interpolate it - is one of those tricky features in that it only helps you when you don't need it anyway.

If you already have acceptable frame rates, I guess it doesn't hurt to turn it on and benefit from slightly smoother visuals, but if you have unacceptable frame rates to begin with, it does nothing but take a marginally playable game, and make it a smoother, prettier marginally playable game. And that's when it works. Sometimes the artifacts are just too much to bear.

And that is why it is frustrating when Frame Gen technology is pushed as a solution to CPU and GPU limitations in games. It doesn't work that way. If you have too low frame rates to enjoy a game, frame gen won't save you. And I am willing to bet the farm on the fact that DLSS4 makes no improvements what so ever in this regard compared to DLSS3. DLSS4 just seems like "More DLSS3, but with less VRAM use", and doesn't appear to fix any of the underlying DLSS3 issues at all. Probably because they can't be fixed.

So we keep being jerked around with non-solutions to real problems, and they try to sell them to us like they are the greatest thing ever. What else is new.

 

[Sir Beregond]

CUDA core count comparison as a percentage change gen over gen:

4090 (16,384) -> 5090 (21,760) = 32.81% increase
4080 Super (10,240) -> 5080 (10,752) = 5% increase
4070 Ti Super (8,448) -> 5070 Ti (8,960) = 6.06% increase
4070 Super (7,168) -> 5070 (6,144) = 14.29% decrease
4070 (5,888) -> 5070 (6,144) = 4.35% increase

 

[Armenius]

I like how people are saying $2500? No way I'm out. Now oh it's "only" $1999 I'm totally in?

WTF

That still way too expensive for a video card, when historically they have always been under $1000 for the top tier until just recently, now they 100% higher, and "inflation" hasn't made costs go up 100%, maybe 25% at most.

The 780 Ti was $699
The 980 Ti was $649
The 1080 Ti was $699
The 2080 Ti was $999
The 3090 was $1499 ( why such a massive price increase ? )
The 4090 was $1599
The 5090 is $2000 why not $1699?

So in around ten years video cards have TRIPLED in price. Other high-end products have not tripled whatsoever in that time, cars have gotten more expensive but certainly not triple the cost from 10 years ago, or OLED TV's have not tripled in price, or CPU's have not tripled in price. Or smartphones have not tripled in price. Only video cars have tripled.

The xx90 card is the Titan replacement, not the xx80 Ti replacement. There was a 3080 Ti and it was $1,199.

GTX Titan = $999
GTX Titan X = $999
Titan X = $1,199
Titan RTX = $2,499

The increase in price for the 5090 is because of the die size. There was no reduction in wafer cost and density has little improvement over N5 while the number of transistors increased by 20.5%

 

[Zarathustra[H]]

But I've always been one to prefer a smoother, higher FPS over the best visuals.

I'm a little bi-polar in this regard, depending on the framerate I am getting.

I target getting minimum frame rates (0.1% minimums or whatever the reviewers call them these days) of at least 60fps. This usually winds up resulting in an average frame rate of ~90FPS, but I judge it by the minimums, not by the average.

If I can't get minimums of at least 60fps, I will tweak the settings until I do. There is no accepting any less.

Once I hit that "minimums above 60" threshold - however - I will turn the eye candy up as high as it will go without dropping me below it again.

That said, I almost exclusively play single player titles these days. If I still played competitive twitchy games, I might be convinced to go for higher frame rate. Who knows.

 

[TheHig]

Because the hype and launch day hysteria would be replaced with “listen we are trying to make more but preorders have sold the cards out for the next 3 months.” You can’t cash in on FOMO without people actually missing out. The mayhem, scalping and all plays onto that gotta have now it and I’ll pay anything mindset.

Why can’t we pre order this shit in 2025. All this AI bullshit and we can’t even have consumer buy products in a civil fashion especially across the world. . AI should be able to detect bots and crap.

Even Sony and Microsoft sorted out their pre orders for release of XSX and PS5 and that was during covid or something.

 

[Zarathustra[H]]

CUDA core count comparison as a percentage change gen over gen:

4090 (16,384) -> 5090 (21,760) = 32.81% increase
4080 Super (10,240) -> 5080 (10,752) = 5% increase
4070 Ti Super (8,448) -> 5070 Ti (8,960) = 6.06% increase
4070 Super (7,168) -> 5070 (6,144) = 14.29% decrease
4070 (5,888) -> 5070 (6,144) = 4.35% increase

That comparison isn't very useful, as the architecture of those cores change from generation to generation.

It's better to look at a combination of total die size and total power envelope.

 

[Zarathustra[H]]

The xx90 card is the Titan replacement, not the xx80 Ti replacement. There was a 3080 Ti and it was $1,199.

GTX Titan = $999
GTX Titan X = $999
Titan X = $1,199
Titan RTX = $2,499

The increase in price for the 5090 is because of the die size. There was no reduction in wafer cost and density has little improvement over N5 while the number of transistors increased by 20.5%

Pretty sure Ada Lovelace was TSMC 4N, not 5N.

 

[Sir Beregond]

The xx90 card is the Titan replacement, not the xx80 Ti replacement. There was a 3080 Ti and it was $1,199.

I would argue xx90 replaced both. Turing and before, the xx80 Ti was the flagship gaming card of it's generation while the Titan was a Titan...remember back then Titan had Titan drivers which ungimped their performance for certain workloads that GeForce gaming drivers gimped. Sort of a halfway to Quadro drivers. In fact when the 3090 released, it did worse than a Titan RTX at certain tasks because of this. Hardly a Titan replacement in that sense at the time in reality despite the "Titan class" marketing when Ampere launched. These days' that's not true anymore because we don't have the same product split, and you now have a choice of drivers.

The 3080 Ti was not really a flagship card of that generation like the 2080 Ti or 1080 Ti was to theirs. Ampere had Samsung 8nm (10nm) which was less advanced than TSMC's 7nm, and much cheaper. Ampere is kind of the odd ball out of recent gens in terms of how many products they released based on GA102 (5 total: 3080, 3080 12GB, 3080 Ti, 3090, 3090 Ti).

But I'd make my case that since then, 40-series never got a xx80 Ti. I doubt 50-series will get any further cut down GB202 based xx80 Ti card for similar reasons as 40-series.

So goes back to what are the gaming and productivity flagships since the 40-series? Funny...it's the same card and its the xx90.

That comparison isn't very useful, as the architecture of those cores change from generation to generation.

It's better to look at a combination of total die size and total power envelope.

Fair...but I still find it interesting that the 5090 got the only really big increase here, so it will be interesting to see how that plays out. You are correct we also seem to be getting a bigger die on the 5090, but similar size dies on everything else to their predecessors. TDP figures up.

And yes, well aware of architectural differences gen to gen as well as power figures. Given both gens are on derivatives of TSMC 5nm, it will be interesting to see the difference here, and I still find the singular model (5090) increasing core count massively compared to the single digit percentage difference on the rest of the announced stack to be interesting and may tell a story when it comes to traditional render comparisons.

 

[Armenius]

Pretty sure Ada Lovelace was TSMC 4N, not 5N.

Looks like it was, and Blackwell is on N4P. The statement about density and cost still hold true. They're all derivatives of TSMC's 5nm process.

 

[Zarathustra[H]]

Why can’t we pre order this shit in 2025. All this AI bullshit and we can’t even have consumer buy products in a civil fashion especially across the world. . AI should be able to detect bots and crap.

Even Sony and Microsoft sorted out their pre orders for release of XSX and PS5 and that was during covid or something.

I would argue that pre-ordering hardware is even more foolish than pre-ordering games.

That said, I do wish they got a hold of the launch shortages and scalping somehow. But that is not in their best interest.

They do the best when there is hype and shortages.

Not to mention that production capacity is limited. it takes like a decade and is prohibitively expensive to increase production capacity just to pump up launch quantities.

And if you instead wait, and build launch quantities before launching, the longer you wait, the less impressive your shiny new product is.

No, their incentive is to launch as soon as they can. They will build a minimal launch quantity so that it is not accused of being a paper launch, but after that, they are happy to brag to investors that they sold out, while letting angry consumers who can't find the products anywhere blame everyone from scalpers, to miners to AI companies (all of which are true, btw.) for the shortages, but never the manufacturer....

Whenever you wonder why something is a certain way, just follow the incentive.

Some of this will hopefully get a little better after some of the many new fabs under construction come online, and there is greater manufacturing capacity in the industry, so that everyone (AMD, Apple, Nvidia, Qualcomm and now even Intel) go shopping for manufacturing capacity at the same two fabs at TSMC and Samsung, but the truth is, demand is probably insatiable at this point. New fabs will come online, and heir capacity will likely be gobbled up faster than they can take orders.

There is sadly no going back the early 2000's when we were a niche market that wasn't competing as much with other markets, and thus weren't always fighting for capacity. PC Gaming is HUGE now, and its fighting for production capacity with any number of other industries, many of which have deeper pockets than we do, and it brings us to the situation we are in.

Can't blame Nvidia for prioritizing $20k AI products over even crazy $2K graphics cards when allocating the share of TSMC's manufacturing capacity that wasn't taken by Apple and others. Anyone would do the same in their situation. That's just business. Prioritize your biggest margin profit centers.

 

[Legendary Gamer]

I would argue xx90 replaced both. Turing and before, the xx80 Ti was the flagship gaming card of it's generation while the Titan was a Titan...remember back then Titan had Titan drivers which ungimped their performance for certain workloads that GeForce gaming drivers gimped. Sort of a halfway to Quadro drivers. In fact when the 3090 released, it did worse than a Titan RTX at certain tasks because of this. Hardly a Titan replacement in that sense at the time in reality despite the "Titan class" marketing when Ampere launched. These days' that's not true anymore because we don't have the same product split, and you now have a choice of drivers.

The 3080 Ti was not really a flagship card of that generation like the 2080 Ti or 1080 Ti was to theirs. Ampere had Samsung 8nm (10nm) which was less advanced than TSMC's 7nm, and much cheaper. Ampere is kind of the odd ball out of recent gens in terms of how many products they released based on GA102.

But I'd make my case that since then, 40-series never got a xx80 Ti. I doubt 50-series will get any further cut down GB202 based xx80 Ti card.

So goes back to what is the gaming and productivity flagship since the 40-series? Funny...it's the same card and its the xx90.


Fair...but I still find it interesting that the 5090 got the only really big increase here, so it will be interesting to see how that plays out. You are correct we also seem to be getting a bigger die on the 5090, but similar size dies on everything else to their predecessors. TDP figures up.

We might be seeing very marginal gains from the 4000 series to the 5000 series on pure raster with moderate RT gains. Nvidia looks to be betting everything on Frame Generation and that's where these cards are getting their performance from. Just look at the hardware specs of everything less than the 5090. It's not impressive and the memory bandwidth is not where it could be. The memory subsystems are 128 - 256 bits. Banking on GDDR7 as well to make up for the lower bandwidth. They're betting on their better memory utilization / new compression schemes for the lower RAM on these cards.

The 5090 looks more like the jump from the 1080Ti to the 2080Ti, it was an increase of 25% or so. These are all datacenter cards repurposed for gaming. Even AMD is going that direction, re-integrating compute with their main lineup. Technically, Intel is also there too. Datacenter designed cards that are released to function as gaming ones. Though they didn't realize they had a hit until it flew off the shelves (and Pat G is on record saying that there is no future in GPUs for Intel, lack of vision and can't read the room or the market...).