nvidia Rtx 4090 is a monster (Official Benchmarks)

[bowrilla]

11 minutes ago, porina said:

I'd also have to assign a perf value to each GPU.

A German media outlet has a performance index derived from all their benchmarks for both native resolution and raytracing (and you can also select by resolution if you want). It's not a full list (lowest cards are 1070 and Vega 56).

[igormp]

19 minutes ago, porina said:

Looking only at nvidia+AMD GPUs only, it was already 29% in the February 2022 survey. I know that intentionally excludes Apple and Intel so would skew it.

Got the data from last month, seems like you're right:

 

That's including both Intel and Apple. Here's the list of ray-tracing enabled GPUs on steam in case anyone is curious:

[igormp]

22 minutes ago, porina said:

I'd really like to transcribe the latest data but I couldn't find a sane way to import the data without doing it manually. The webpage formatting I just couldn't get working how I wanted when dumping into a spreadsheet. If anyone can find a way to automate that, I'd love to play with the data, even historic to show better trends.

I simply copy pasted the website text and turned that into a csv in less than 30 seconds lol

steamdata.csv

[leadeater]

2 hours ago, Shzzit said:

So for watercooling an FE would do great?

FE cards have always been a good option for water cooling.

[leadeater]

37 minutes ago, Shimmy Gummi said:

BUT - 13th gen launch coming up will probably mean the reviewers will now use the 4090 - meaning hopefully Zen4 chips will get a re-test. We knew that the 6950XT/3090ti were the limiting factors in a LOT of the 1080p data so it didn't really give us good comparisons between Intel and AMD architectures. Very interested to see this retested.

13th Gen won't change that much I'm afraid. Anyway why would you get a 4090 for that resolutions in the first place, it'll be CPU limited even by 14th Gen. It's all about balance.

[tim0901]

2 hours ago, bowrilla said:

While I do understand your point here, I don't fully agree but that's ok.

 

What I do think though however: I wonder how much sense RTX 50 will make in 2-3 years time when the 4090 is already being CPU limited at 1440p and even high end games with massive resource hunger run easily at 4k ultra. The "problem" with these kind of performance leaps is: developers can't really aim for that kind of hardware unless they intentionally leave out the majority of gamers with even a little bit older hardware. It's hard to top it all at this point and it makes me wonder what's coming up next. Personally I do expect the performance level of the 4090 to stay in the top 10 for probably at least 3 if not even 4 generations of GPUs. This generation might actually be what the GTX 10 series was many years ago and the 4090 being the Titan equivalent.

RT usage is just going to go up and up. RT as it's used today barely scrapes the surface of its capabilities as is pretty much only used for lighting effects and reflections at a rather low quality (only 1-2 rays per pixel - more rays = better image).

 

Real-time fully path traced rendering is kinda the "holy grail" of computer graphics, but there's a reason it's only used in very simple games like Minecraft or Quake 2. Even the 3090 only reaches ~30fps in Quake 2 RTX at 4K, let alone in a game that has more than 20 triangles. Modern AAA titles will need GPUS to have orders of magnitudes more RT capability to even think about going to fully path traced rendering.

 

As such, in future generations I imagine we'll see more and more development going towards RT capabilities. We'll still get more/better general purpose compute cores, but those will be added primarily for their use in compute applications, rather than for rasterization.

 

But that's the key of it all. GPUs are becoming less and less "graphics" cards and more and more "massively parallel compute" cards. Gamers aren't the target audience of these architectures and haven't been since at least Turing. We'll get more RT cores, but that's primarily because render farms and workstations are demanding the additional performance for their productions, not for RTX. Gaming is just a lucrative market that Nvidia can milk for a lot of money with scraps from their compute business.

[porina]

18 minutes ago, igormp said:

I simply copy pasted the website text and turned that into a csv in less than 30 seconds lol

That was the first thing I tried in February and it came up with all sorts of empty cells everywhere. Hadn't tried it again since, so maybe something changed. Still that gives something to play with.

[Vibora]

How would the 12900 run with the 4090? Will the system be CPU bottlenecked?

[tim0901]

11 minutes ago, leadeater said:

FE cards have always been a good option for water cooling.

Yep, they're by far the easiest to find waterblocks for. You'll be lucky to find a single model of waterblock for a third party board, let alone choices.

[leadeater]

4 hours ago, yolosnail said:

Is it just me or do the reviews seem to be getting crap views compared to other launches?

 

3 hours ago, tim0901 said:

The timing of the launch just wasn't a great time for peak viewership hours. Most of Europe is only just now getting home from work, while in the US it's still the middle of the day. Give it another 12 hours and I imagine those numbers will look quite different.

Also if YouTube detects huge jump in views on a new video the view counter gets frozen, it's a anti-viral video protection for videos that shouldn't be getting the views.

[Mister Woof]

11 minutes ago, leadeater said:

13th Gen won't change that much I'm afraid. Anyway why would you get a 4090 for that resolutions in the first place, it'll be CPU limited even by 14th Gen. It's all about balance.

It will more likely force them to retest, is what I meant.

[igormp]

5 minutes ago, porina said:

That was the first thing I tried in February and it came up with all sorts of empty cells everywhere. Hadn't tried it again since, so maybe something changed. Still that gives something to play with.

I pasted it into a text editor and not excel, fwiw.

[tim0901]

12 minutes ago, Vibora said:

How would the 12900 run with the 4090?

Probably very much like this:

Given that GN currently use a 12700KF for their GPU bench.

 

12 minutes ago, Vibora said:

Will the system be CPU bottlenecked?

Will one of the fastest gaming CPUs around be CPU bottlenecked?

 

In a handful of easy-to-run titles at "lower" (non-4K) resolutions, maybe. But it's just as possible that they're seeing engine limitations there. In most games at 4k? No.

[Brooksie359]

5 hours ago, Sir Beregond said:

 

The 8800 GTX would like a word. At least 130% rasterization generational improvement over the 7800 GTX. If I recall, the 6800 Ultra was something like 110-120% over the FX 5800/5900 Ultra too. Again, rasterization generational improvement.

 

Sure if we are counting RT and DLSS the 4090 improvements are huge, but raster is not 130% faster than the 3090 Ti. They made this same claim with Ampere too and that was definitely NOT "the biggest generational leap ever". More like biggest leaps of past decade?

 

Not trying to cast shade on the performance, the 4090 looks like a beast. Just calling out marketing BS.

I'm technically you would be wrong as they never specified percentage they just said leap so if you looked at the difference in teraflops and other metrics it would still be bigger than between 8800 gtx and 7800 gtx. 

[Shzzit]

Looks like most FE cards overclock to around 3ghz at about 68-70c.  So exiting, even at that most cards don’t go over 450 watts.

 

Womdwr if an AIB card could do 600 watts would it let higher over clocks?

 

Are cards like the strix thats 2000 have binned chips for better overclockes

 

Trying to figure out why a strix is 400 over msrp.

[bowrilla]

7 minutes ago, tim0901 said:

Yep, they're by far the easiest to find waterblocks for. You'll be lucky to find a single model of waterblock for a third party board, let alone choices.

Well ... most reference card models are easy to find blocks for. EK, Watercool and Alphacool have usually lots of options for the most wide spread models. Custom PCB cards are tricky though.

[Sir Beregond]

11 minutes ago, Brooksie359 said:

I'm technically you would be wrong as they never specified percentage they just said leap so if you looked at the difference in teraflops and other metrics it would still be bigger than between 8800 gtx and 7800 gtx. 

So, marketing BS?

[Sir Beregond]

7 minutes ago, bowrilla said:

Well ... most reference card models are easy to find blocks for. EK, Watercool and Alphacool have usually lots of options for the most wide spread models. Custom PCB cards are tricky though.

Reminder that FE is not reference, so got figure out what cards are actually "reference".

[Brooksie359]

3 minutes ago, Sir Beregond said:

So, marketing BS?

I mean it's true so how is it bs? I mean this is an incredible leap in rasterization and raytracing performance so to go and look all the way back to the 8800gtx and say oh the percentage leap was higher is sorta reaching tbh. 

[tim0901]

1 minute ago, bowrilla said:

Well ... most reference card models are easy to find blocks for. EK, Watercool and Alphacool have usually lots of options for the most wide spread models. Custom PCB cards are tricky though.

I may have forgotten that the reference board still exists, it's just not the exact one used by the FE card these days.

 

But IMO they're still a bad choice, as you're spending more than you would spend on an FE card to let an AIB design a fancy cooler that you're only going to rip off and replace with a waterblock anyway. You're not gaining anything over the FE in terms of power regulation etc that would unlock more performance.

[Kisai]

1 hour ago, porina said:

Where do you draw the line?

 

The 3090 is the biggest card, physically, I can fit in my mid-tower case

The last time I replaced the chassis, it was to fit the 1080, because the case before it (which was a large tower case) wouldn't fit it, the card was longer and hit the drive cage.

 

These "3 slot" cards should not exist. I'd actually draw the line at "1.9 slot", because motherboards aren't being designed to accommodate 3-slot cards. m.2 SSD's end up underneath the coolers for the GPU's to make use of that otherwise wasted MB space. However you have to pull the GPU to replace the SSD then. I doubt we will see a revision of the ATX "motherboard" standard revised any time soon to bring back the full 7-slots, and as a consequence MB's are getting more and more proprietary while being ATX-in-name-only.

 

Hell, this is what I would do:

1. Sell the mainstream 450w "3 slot" parts with a x16 extension and mounting bracket to allow it to be used on chassis that have a dedicated "GPU" space. Below: Bequiet Silent Base 802

2. Sell a 2-slot model that has only a cold plate (eg for after-market coolers) with a bios switch for "Unlocked" (liquid cooled)/"450w"/"300w" mode

3. Every tier beneath has a standard cold plate mount, but are shipped with the OEM cooler. So all AIB coolers would also fit just as well as aftermarket ones.

 

[Sir Beregond]

20 minutes ago, Brooksie359 said:

I mean it's true so how is it bs? I mean this is an incredible leap in rasterization and raytracing performance so to go and look all the way back to the 8800gtx and say oh the percentage leap was higher is sorta reaching tbh. 

If I am measuring performance, not TFLOPs, then yeah "biggest leap" is a stretch and factually untrue for new cards.

 

How is my example a stretch. Those are literally the biggest uplifts "in history". Not Ampere. Not Lovelace. 2x to 4x uplifts are not rasterization with Ampere or Lovelace. Its highly dependent on DLSS3 and RT to achieve those, so in games where raster matters...yep its pretty far off from 130%.

 

So no, it is in fact, not the biggest generational uplift in history. Of the past decade is probably more accurate but to be fair I haven't sat down and checked the flagship to flagship differences each gen recently.

[tim0901]

26 minutes ago, Brooksie359 said:

I mean it's true so how is it bs? I mean this is an incredible leap in rasterization and raytracing performance so to go and look all the way back to the 8800gtx and say oh the percentage leap was higher is sorta reaching tbh. 

1 minute ago, Sir Beregond said:

If I am measuring performance, not TFLOPs, then yeah "biggest leap" is a stretch and factually untrue for new cards.

 

How is my example a stretch. Those are literally the biggest uplifts "in history". Not Ampere. Not Lovelace.

You're both wrong.

 

The biggest leap obviously came with the NV1, because it was infinitely better than all Nvidia GPUs released before it. Of which there were none.

[Sir Beregond]

38 minutes ago, tim0901 said:

You're both wrong.

 

The biggest leap obviously came with the NV1, because it was infinitely better than all Nvidia GPUs released before it. Of which there were none.

It was definitely a first of its kind, but definitely not the first graphics chip/card. It succeeded the Nvidia RIVA TNT2.

 

EDIT: Nope I am an idiot. You said NV1 not NV10. Oops 

[ewitte]

I use a view 71 but I removed all of the front 3.5in drive bays.  I've got three behind the motherboard though and three NVME drives.  I think you can get one of the two 3.5in*4 bays in the front without the radiator and not hit the 4090.  I have something like 20 inch clearance without bays but with 3-4inches of radiator and fans.  The biggest 4090 is 14-14.5 inches.