Nvidia GeForce RTX 4090 Review: Next-Gen GPU Monster

loki1944

Posts: 699   +523
3080 owners only have two real options, IMO. Either just say to hell with it, go all in and get a 4090 or wait for the 5080 in 2024, when the price/performance ratio will be back in sync for the 80 class cards.

..but the one thing they shouldn't do is go from their current 3080 to a 4080, like you say, terrible deal.

Not much point upgrading; even at 4K a 3070 is plenty.
 

Nintenboy01

Posts: 238   +183
You guys will have to benchmark/review this thing again after the Zen 4 3D chips are out. I have a feeling it'll still be CPU-bottlenecked at 1440p and maybe even 4K in some games with Raptor Lake.
 

Nintenboy01

Posts: 238   +183
Not much point upgrading; even at 4K a 3070 is plenty.
eh not really, unless you're ok with DLSS Balanced or lower in games that support it. It lacks the grunt and VRAM for many games at 4K to avoid stuttering/reduced texture quality and maintain at least 60FPS.
 

kiwigraeme

Posts: 1,399   +1,038
Steve - heading out - but I just realized there is a fundamental logical flaw .

Which is not applicable in this situation as buyers don't care about Cost per frame.

Cost per Frame - implies every frame is of equal value - when they are clearly not .
The first 30 are fundamental by definition.
The next 30 - are highly essential for smooth experience.

The next 84 ( 61-144 ) are highly desirable ,
Then those after are bonuses or completely un- necessary

In this sample for 1440p 201 frames is useless for many people
4K 144fps where it falls is perfect

Then is subjective to buyer

I think the way it's stated is fine - but needs qualification
 

waclark

Posts: 788   +488
I don't know. The 4080 16GB is nearly the size of the 3090 for core count and it has around a 50% frequency bump. It will be significantly faster than a 3090ti. Almost certainly in the 25-30% range. I'd expect the "4070" to be around 3090 level.
Very likely. I was being conservative in my estimates. The point being that the price points and performance level seem to be matched up. I'm holding judgement on the 4080 until I see performance. Still, even $800 is more than I want to spend. I'd like to be $650-700 and the 10G 3080 may get there but I like the RT improvements that 40 series looks to bring if the 4090 is any indication.
 

Mr Majestyk

Posts: 1,564   +1,469
Yes it's impressive, and you'd have been lying to yourself if you thought it wouldn't be, but let's face it, so many comments and barely anyone commenting will even be thinking of buying one.

AMD had better improves RT performance 150% minimum this gen or it will still look very bad. If they do however, insiders are saying it is much cheaper to manufacture (dies are much smaller than RDNA2) and AMD will easily be able to undercut Nvidia's looney pricing. I'm still saying a 7800XT will beat out a 4080 16GB, as it's moving to 320 bit bus, 7900XT is on 384 bit bus like 4090 and there is alo the 7950XT. I'll bet 7700XT will match 4080 12GB. BAck in the real world the 7800XT would be what I would buy, 4080 is garbage value and let's face it the 12GB card is a 4070 Ti in all but name.
 
Last edited:

EdmondRC

Posts: 447   +655
I don't know. The 4080 16GB is nearly the size of the 3090 for core count and it has around a 50% frequency bump. It will be significantly faster than a 3090ti. Almost certainly in the 25-30% range. I'd expect the "4070" to be around 3090 level.
The 4070 will be very close to the 4080 12 GB because the 4080 12 GB is essentially a 4070 Ti.
Here's the thing though, the 4090 accomplishes this kind of performance, but it has a massive increase to the number of cores as well. The 30 series scaled pretty linear with the number of cores, not perfectly, but surprisingly close. If this holds true for the 40 series, then you can expect the 4080 16 GB to be 40% slower than the 4090 and the 4080 12 GB to be more than 50% slower. The 4080 12 GB will likely be somewhere between the 3080 and 3090 in terms of performance.

The GPU to hold out for might be the 4080 Ti. It can't be more than $1600, but it also can't be less than $1200. That being said, with a rumored 14K cores, it will be closer in performance to the 4090 than the 4080, and it will probably have more overclocking overhead than the 4090. That is, assuming that the prices don't nosedive next year, which I fully expect them to.
 
"Quite incredibly, this has seen transistor count increase by 170% from 28.3 billion to an insane 76.3 billion."

Really? If you multiply 28.3 by 1.7 you get 76.3? Awesome math skills in this article.
Well actually, given that you START with 100%=28.3 billion and INCREASE BY 170%=48.1 Billion you should end up with 270%=76.4 billion, right? Or, going the other way, 76.3/28.3 = 2.7 to 1 decimal place. Not awesome, just correct.
 

rmcrys

Posts: 323   +265
You must live in Europe LOL. Noone cares about electricity cost for a PC in USA, granted even here prices have skyrocketed.

Don't come with cr.p talk. There are people in this planet that try to balance energy cost and quality of health and planet, not to deplete everything to be cheap at any cost.

"Oh EVs are also on expensive energy, oh coal is not that bad...": that is a talk only from someone who knows nothing about the matter.
 

yazooo

Posts: 7   +13
TechSpot Elite
I find the review score skewed too heavily towards the benchmarks alone, but not representing lack of improvement in power efficiency, size and worst of all - pricing! The margins for nvidia on this must be significant.
 

takaozo

Posts: 514   +812
And now I want to see the mITX and the mATX builds with this "brick".
This new cards dont fit even in a mATX board, it will cover all bottom connectors.
 

TheRealSCDC

Posts: 494   +853
Enviromentalists LOVE wasting electricity, look how much power an EV car takes to travel just a single mile, or how much chinese made garbage they consooom. Oil bad, magic electricity good.

The reason techspot whines about power usage is likely the same reason the rest of us do: excessive heat production is a royal PITA if you dont live in the arctic circle. (and techspot being run by environmental nuts? Huh? What on earth are you talking about, have you ever MET an ecomentalist?)


And such cards have not gone away. Look how small 3060s are. The 4200ti wasnt even the high end, that belonged to the 4600/4800, which were substantially longer, that was the era of IBM sized cards (and very loud annoying fans).


There's plenty of good games on PC from previous generations. From the current gen, Instruments of Destruction has been amazingly addictive. Enlisted has scratched that multiplayer itch, Destroy all humans 2 looks to be quite fun, We got Spiderman on PC this year too.

Gaming has always been full of detritus and greedy humans. There's still good games out there, and they're still fun to play. Just ignore the AAA trash.

In a free market we'd be free to pursue nuclear power instead of using natural gas. But hey, atoms bad, better to spend hundreds of millions on solar panels and legislate homeowners not being able to charge their new cars during the day.

Europe just made their entire grid reliant on fossil fuels from their neighbor and claimed it was "green", and were utterly SHOCKED to find out that this was a bad idea.

If only somebody had told them that....but hey, at least there are no more mean tweets, right?

Nailed it
 

neeyik

Posts: 2,420   +2,964
Staff member
not fare not to include radeon IC bandwidth
Nvidia hasn't disclosed the cache data path widths in Lovelace, so it's not possible to directly compare bandwidths. Both architectures have the same size L1 cache (128 kB) and for RDNA 2, the read data width is 128 bytes per clock and 32 bytes per clock for writes. In Nvidia's Ampere architecture, it's 128 bytes for both read/write and it's safe to assume that Lovelace will be the same (as nothing else has changed with the L1 cache).

A Navi 21 full chip has 4MB of 16-way L2 cache, with a data path width of 128 bytes per clock (read and write) per L2 partition. The 4090 has 72MB of L2 cache but an unknown path width; for Ampere, the partition data path is 128 bytes per clock, so again, Lovelace will be at least that.

And finally, RDNA 2 has the Infinity Cache, up to 128MB of 16-way L3, with a path width of 64 bytes per clock per partition. Lovelace obviously doesn't have any L3 cache.

The highest level of cache is particularly significant because that's what is paired with each memory controller. So for the Navi 21, each MC has 16MB of cache associated with it. In Ampere, it was just 0.5 GB, whereas in Lovelace, it's now 6MB for the 4090 and 8MB for the full AD102 chip.

AMD obviously has an advantage there, but the data paths to those partitions is relatively narrow at 64 bytes. If Lovelace has the same L2 cache structure as Ampere, then not only is clocked faster, the data path is also wider.

It will be interesting to see what AMD does with RDNA 3, where the general expectation is the L3 cache and MCs are on separate chiplets, surrounding the rest of the GPU core die.
 

neeyik

Posts: 2,420   +2,964
Staff member
I've noticed something else about the 40 series launch. Either TSMC's 4N(N) process has relatively awful yields or Nvidia is fobbing off consumers with some pretty rubbish dies, or it's a combination of both.

nv_new_arch_yield.png

Historically, the first 'best' model using their newest, biggest chip has been sporting the best dies from the wafer -- hardly anything was disabled due to defects. (But with Lovelace, there's quite a lot being nobbled, especially with the forthcoming 4080s. The 12GB version is essentially less than half a fully working AD102.

With Ampere, the absolutely worst GA102 dies ended up in the A4000 cards, but it was still 56 out 84 SMs used (67%), and L2 cache was full. With Turing, the duff TU102 dies ended up, eventually, in their now-useless CMP 50HX mining card with 56 out 72 SMs (78%).

The AD102 isn't as physically large as the GA102 and TU102, so wafer yields should be better. The lack of an announcement of any professional AD102-powered cards makes me think that the yields aren't all that good, and Nvidia's hand was somewhat forced to kit out the 4080s with some fairly miserable-looking dies.

Edit: Ignore the 4080 remarks. As @EdmondRC has kindly pointed out, they're on a different die.
 
Last edited:
"Quite incredibly, this has seen transistor count increase by 170% from 28.3 billion to an insane 76.3 billion."

Really? If you multiply 28.3 by 1.7 you get 76.3? Awesome math skills in this article.
Hi, percentages can be a tricky thing... when increasing a number by 100% then you are actually multiplying it by 2.

A 170% increase therefore means to actually multiply the original number by 2.7.

E.g. 100 increased by 50% = 150 (1.5x), but 100 increased by 100% = 200 (so multiplied by 2). Therefore 100 increased by 170% = 270 (or 2.7x) (y) (Y)


(Sorry, I see this post was already replied to and clarified.)
 
Last edited:

EdmondRC

Posts: 447   +655
Just for fun I tried to put a 4090 in my cart at Newegg and Best Buy, sold out nearly immediately. I never had a chance (I wasn't actually going to buy one, but I wanted to see if I could even get to checkout). Only a few SKUs were available. Appears to have been a paper launch with limited quantity. Already listings on eBay for over $3K.

 

Ohnooze

Posts: 445   +821
Well so far my 6800xt is doing really well at 4k so I think I'll wait it out this generation. It's nice to see the the jump in performance but the power demand and price is ridiculous. I'm very interested to see what AMD has in store.
 

EdmondRC

Posts: 447   +655
I've noticed something else about the 40 series launch. Either TSMC's 4N(N) process has relatively awful yields or Nvidia is fobbing off consumers with some pretty rubbish dies, or it's a combination of both.
Wait, I thought that the 4090, 4080 16GB, and 4080 12GB were all on separate dies? My understanding was that it was AD102, AD103, and AD104. Are you saying that the 4080 12GB is actually the same AD102 die just with more than half of the cores not activated?
 

EdmondRC

Posts: 447   +655
Well so far my 6800xt is doing really well at 4k so I think I'll wait it out this generation. It's nice to see the the jump in performance but the power demand and price is ridiculous. I'm very interested to see what AMD has in store.
The asking price of these cards is way too high. I'm definitely waiting for a significant price drop at the bare minimum, but most likely I'll hold on to my 3080 for at least a couple more years.
 

neeyik

Posts: 2,420   +2,964
Staff member
Wait, I thought that the 4090, 4080 16GB, and 4080 12GB were all on separate dies? My understanding was that it was AD102, AD103, and AD104. Are you saying that the 4080 12GB is actually the same AD102 die just with more than half of the cores not activated?
Ah, well, you see this is very much a technical case of "Nick is a halfwit and needs more coffee before thinking" :DAnyway, you're quite right - the two 4080s are on different dies altogether!

That said, the 4090 still isn't as 'full die' as previous top-end, first releases; 11% reduction in SMs and 25% reduction in L2 cache is pretty significant. Here's a corrected version of the table:

nv_new_arch_yield.png