NVIDIA GeForce RTX 50 Technical Deep Dive

[oxrufiioxo]

I wonder if SteamOS for desktop will support Nvidia or if it will still be AMD system locked as it is now designed. I wouldn't mind upgrading to a 5080, but I also want to dual boot Windows 11 and SteamOS in the future. Hmm.

It would be odd to ignore 80%+ of the dgpu market. Unless Valve thinks the amount if Nvidia dgpu users that would switch to using steamOS is insignificant. I don't know a single person locally me and 10 of my closest friends who rock Nvidia are overly excited about SteamOS though.

 

[freeagent]

Please when reviewing the fe edition check any variance for potential throttling in a closed case real world scenarios.

FWIW my system runs cooler with the panels on

 

[Space Lynx]

It would be odd to ignore 80%+ of the dgpu market. Unless Valve thinks the amount if Nvidia dgpu users that would switch to using steamOS is insignificant. I don't know a single person locally me and 10 of my closest friends who rock Nvidia are overly excited about SteamOS though.

I don't think Valve wants to leave Nvidia out, its just the open source drivers AMD provides for Linux was the deciding factor. SteamOS makes a lot of sense for me, combine that with Heroic Launcher and like 80% of my games can be played on SteamOS.

 

[oxrufiioxo]

FWIW my system runs cooler with the panels on

Get a box fan and point it at an open test bench then compare lol....

I don't think Valve wants to leave Nvidia out, its just the open source drivers AMD provides for Linux was the deciding factor. SteamOS makes a lot of sense for me, combine that with Heroic Launcher and like 80% of my games can be played on SteamOS.

I understand it's appeal it makes pc more like a console somthing amd is better at than desktop as it is.

 

[Garrus]

NVidia marketing pretending to be an article though.

Noticed my youtube is full of it today also.

What people actually want to see, still can't see.

 

[oxrufiioxo]

NVidia marketing pretending to be an article though.

Noticed my youtube is full of it today also.

What people actually want to see, still can't see.

That's not true we had one apples to apples benchmarks and one semi apples to apples...

Now we have 3 apples to apples and 1 semi lol.... #Progress....

5090 review isn't far out if it's dissappointing the whole stack is in trouble.

 

[wNotyarD]

5090 review isn't far out if it's dissappointing the whole stack is in trouble.

The stack may be in trouble even if the 5090 itself is not. The 90-tier is the only one with noticeable core/shader increase. Everything else will rely on VRAM bandwidth mostly.

 

[Jtuck9]

I don't expect the 5070 to age well but would love to be wrong.

I was deliberating this myself compared to the sell on value of something like the 9070xt

 

[freeagent]

Get a box fan and point it at an open test bench then compare lol....

I dont have a box fan

 

[oxrufiioxo]

The stack may be in trouble even if the 5090 itself is not. The 90-tier is the only one with noticeable core/shader increase. Everything else will rely on VRAM bandwidth mostly.

I'm not sure how much trouble it'll actually be in when the fastest card from the competition is likely to be a 7900XTX or a 9070XT which might be a discounted 4070ti people could have purchased for 2 years now so will live or die based in it's msrp....

40 series buyers like me are going to be pretty disappointed and maybe even 30 series owners though but it looks like Nvidia only cares about content creators and 20 series and older gpu users this generation lol.

I dont have a box fan

View attachment 380100

The problem with case testing is they are all drastically different the only actual fair and accurate comparison is an open test bench.

People also have vastly different ambient etc.

 

[freeagent]

I honestly don't like open air benches

They are ok if you swap all kinds of hardware around, but I don't like it as a daily.

18c in my space right now, a bit chilly.

 

[evernessince]

Just watched this interview as this was posted on TPU:

I have to wonder if the use of cables connecting the slot PCB and IO to the primary PCB will impact longevity. Aside from the extreme signal integrity issues that come with carrying a PCIe 5.0 signal, I feel like a cable is just less reliable.

A lot of great engineering went into this card but I'd almost be inclined to get a AIB model simply to avoid potential cable issues down the line. Nvidia's history with cutting edge connectors and cables doesn't necessarily instill confidence either.

 

[oxrufiioxo]

I honestly don't like open air benches

They are ok if you swap all kinds of hardware around, but I don't like it as a daily.

18c in my space right now, a bit chilly.

Me either but it's still the only apples to apples testing we can get on thermals after all somebody buying a terrible case with crap case fans is on them not the maker of the gpu or cpu thermal solution.

HUB sticks their gpus in a case though for reviewes if people are into that.... A middling one last I checked but a 20C ambient which is probably cooler than most people's ambient people hate ac for some reason... lol

 

[adilazimdegilx]

The most interesting part of the 5000 series is sadly the the new cooler on 5090. I kinda like that they essentially changing pcb design we stuck with since the first expansion cards and improving cooling capacity without increasing cooler size. We'll see how it actually performs but I'm optimistic about it.
That's pretty much it though. Even the performance numbers by Nvidia aren't showing a substantial increase and it seems that with the addition of new game data to their slides, they actually lowered it a little more. 5090 is about 30% over 4090 while others are between 15-20% over their 4000 counterparts (base models, not supers). Which I believe will make this gen the worst gen-gen uplift since.. idk but before Pascal for sure.

 

[oxrufiioxo]

The most interesting part of the 5000 series is sadly the the new cooler on 5090. I kinda like that they essentially changing pcb design we stuck with since the first expansion cards and improving cooling capacity without increasing cooler size. We'll see how it actually performs but I'm optimistic about it.
That's pretty much it though. Even the performance numbers by Nvidia aren't showing a substantial increase and it seems that with the addition of new game data to their slides, they actually lowered it a little more. 5090 is about 30% over 4090 while others are between 15-20% over their 4000 counterparts (base models, not supers)

Yeah I'm most interested to see how it performs going back to two slot more normal sized cards is appealing to me although I doubt AIB can replicate it at anywhere near the same cost.

 

[RedelZaVedno]

5090 is about 30% over 4090 while others are between 15-20% over their 4000 counterparts (base models, not supers)

Yeah, it's a total shitshow especially when compared to Supers. Nvidia si basically saying buy 5090 or F off. It makes sense since Jensen believes "true Nvidia gamers" rock $10K gaming monsters at home.

 

[codex5600x]

so I was “pushed” over the edge to buy a ps5pro because my games look better…

Will the 5090 make my games look better than a 7900xtx. ? or a 5080? It is faster, but do I really need the speed of the 5090?, when I can pay half for a 5080 and get everything…

100% don't need but if for you money no matter 5090 will be wow

 

[Broken Processor]

I must have gotten mixed up but did I just read 5% perf uplift for 33% more cores?

 

[bonehead123]

SO, as always:

 

[freeagent]

Gaming is no different than any other hobby now a days. How far you want to take your experience depends on how much you want to spend.

I think people forget this, or haven't had other hobbies

 

[scooze]

5070 will be faster then 4090 slower then 4070 Super

 

[AusWolf]

The Blackwell SM doubles the INT32 bandwidth and throughput by giving all shader cores the ability to run either INT32 or FP32

Finally we have a clear and singular idea of what a "CUDA core" is.

 

[robb]

Yeah that stuff adds a reason for the app, hopefully nvidia profile tweaker adds it as well.

Still waiting for this stuff to work on DLSS1 games.

How are they going to do that from the app when they haven't even added many of the basic features from the control panel to the app yet? I'm sorry but the app is a complete and utter piece of crap and can't even detect a lot of games I've played and list games I haven't played in years or sometimes ever at the top of the list of recently played games.

 

[oxrufiioxo]

Finally we have a clear and singular idea of what a "CUDA core" is.

Interestingly I think Turing did it in a similar way to Blackwell vs what Ampere/ADA was doing but INT32 I believe benefits non gaming workloads more.

NVIDIA's Turing architecture, "INT32" refers to a dedicated processing unit for 32-bit integer calculations, while "FP32" refers to a dedicated processing unit for 32-bit floating-point calculations; the key difference is that Turing allows for simultaneous execution of both integer and floating-point operations on separate cores within the same streaming multiprocessor (SM)

This sort of thing is beyond me but I was getting Turing flashbacks when looking at slides. Seems performance uplifts will also be Turing like at least when it comes to price to performance..... Maybe this is the true successor to Turing lmao...

 

[AusWolf]

Interestingly I think Turing did it in a similar way to Blackwell vs what Ampere/ADA was doing but INT32 I believe benefits non gaming workloads more.

NVIDIA's Turing architecture, "INT32" refers to a dedicated processing unit for 32-bit integer calculations, while "FP32" refers to a dedicated processing unit for 32-bit floating-point calculations; the key difference is that Turing allows for simultaneous execution of both integer and floating-point operations on separate cores within the same streaming multiprocessor (SM)

This sort of thing is beyond me but I was getting Turing flashbacks when looking at slides. Seems performance uplifts will also be Turing like at least when it comes to price to performance..... Maybe this is the true successor to Turing lmao...

Turing had separate INT32 and FP32 cores, and only an INT32 core was counted as a CUDA core (or the two together, I'm not sure, but it doesn't matter).

Then, Ampere made the FP32 cores capable of doing INT32 as well. That's how Nvidia doubled the number of CUDA cores without actually doubling it. I called it a cheat back then.

Now, every core can do both operations, which is just a lot clearer, imo. Let's see how it plays out in actual performance.