Discussion Nvidia Blackwell in Q1-2025

[jpiniero]

Not for the 5070 relative to the 4070. It has higher advertised boost clock and a 50W higher TDP to brute force any problems.

That's why I said Game clock (heh). NV GPUs can automatically go above the advertised boost clock in games easy.

 

[Win2012R2]

Process node is the same or almost the same. I don't understand why people expected a huge performance increase?

Better architecture? Better perf in 4k due to much bigger mem bandwidth?

 

[CastleBravo]

I don't think so. It should be the same at minimum.

33% more SMs and a wider memory bus on 21% more transistors. I think an IPC decrease is possible.

 

[Heartbreaker]

Due to the extra competition at the low end I at least thin the RTX 5060 Ti will be a beefier card, maybe even RTX 5060. The MSRP of 5070 seems to hint that way. But let's wait and see ...

I don't. It looks like 5060 will still have a 128 bit bus, so 8GB of VRAM unless it uses the new 3GB GDDR7 chips, and this might be the same for the 5060 Ti.

Typical NVidia pattern, is the go insane for the top card, and each step down, they give progressively less improvement.

You can see it again here, comparing to the Super cards where available:

5090 has about 30% more Shader cores, than 4090.
5080 has about 5% more Shader cores, than 4080 Super.
5070 Ti has about 6% more Shader cores, than 4070 ti Super.
5070 has about 15% LESS Shader cores, than 4070 Super!

Other than 5090, most cards are probably small improvement at Raster compared to 40 Series Super cards, maybe 15%...

But 5070 actually has significantly LESS Shader cores than 4070 Super. IMO it will be lucky to even match the 4070 Super at Raster. It will probably be a trading blows type situation where the extra memory BW helps in enough benches to eek out a win, but I'd bet 4070 Super will be faster in some games than a 5070.

This pattern does NOT bode well for 5060. The 5060 Ti might have a chance if it's built on a same die as 5070 and has a 192bit memory bus...

 

[gdansk]

That's why I said Game clock (heh). NV GPUs can automatically go above the advertised boost clock in games easy.

Yep, but the 25% higher TDP will help clock rates for the 5070 relative to the 4070 even if your theory about a dense-prioritized implementation are correct.

 

[Gideon]

Beware NVidia marketing. That video appears to be a mockup, as there is no more detail until they zoom, you should be able to see the difference before that.

Also somewhere next to all the new Transformers models, I read "Beta", so IMO, this just a work in progress with faked video.

Basically more misleading NVidia marketing until something real shows up.

Well Digital Foundry (yeah-yeah Nvidia shills, whatever) also seems to have had access to it:

some artifacts are still very noticable, but all-in-all it looks quite good

 

[Tigerick]

Jensen said something interesting there about the changes in the SM.

"Two dual shaders, one is for floating point and one is for integer" Which he also clarified saying something like 'concurrent execution. Equal performance for both'. Which sounds like an evolution of Turing and Ampere, with Ampere/Lovelace being half a step.

Maybe 128FP32(64FP32+64FP32) + 128INT32(64INT32+64INT32) per SM. That
s a lot of Integer performance, but at least 1 clock cycle would not be conditionally half rate in floating point). This would of course make TFLOPs scale differently again in games vs previous 2 gens

Hopefully we get a whitepaper but its not yet certain

Yep, something doesn't add up. I could use the formula to calculate Shader FP32 numbers of RTX-4090, ie. 82.6 TF. But using the same formula, I get 104.9 TF: far lower than 125 TF shown in the presentation.

BTW, shader performance numbers are not listed in the official specs of RTX-5090.

PS: CUDA capability is getting big jump in the version number to 12.9, really interested to see what changes under hood.

 

[coercitiv]

Well Digital Foundry (yeah-yeah Nvidia shills, whatever) also seems to have had access to it:

some artifacts are still very noticable, but all-in-all it looks quite good

Haha, it's time for Nvidia to tell the press to tell the consumers how prone to image artifacts the previous technology was. All kinds of issues will pop up from the DLSS version that was supposed to be "better than native" just 24 hours ago.

 

[gdansk]

far lower than 125 TF shown in the presentation.

I think that's for the full fat GB202 which is - of course - not shipping.

 

[Gideon]

This pattern does NOT bode well for 5060. The 5060 Ti might have a chance if it's built on a same die as 5070 and has a 192bit memory bus...

You are probably right about the 5060. The reasons why i think 5060 Ti will be different this time (probably indeed using the same die and 192 bit bus) is because :

1. Last time there really was no competition (Navi 32 was too expensive to produce and Navi 33 was a stupid chip. barely faster than last gen) - this time there probably is and from both vendors
2.They'll want to have at least some sub 600$ SKUs to endice more budget oriented RTX 2060 SUPER; RTX 3060 12GB and possibly RTX 3070 holdouts to upgrade who don't want to dish out around 600 bucks for a 5070

 

[gdansk]

Haha, it's time for Nvidia to tell the press to tell the consumers how prone to image artifacts the previous technology was. All kinds of issues will pop up from the DLSS version that was supposed to be "better than native" just 24 hours ago.

That is one of the major advantages of Nvidia Neural Rendering Real Time Ray Tracing technology. It gets better every generation unlike boring old raster which doesn't remove artifacts every generation.

 

[Gideon]

Haha, it's time for Nvidia to tell the press to tell the consumers how prone to image artifacts the previous technology was. All kinds of issues will pop up from the DLSS version that was supposed to be "better than native" just 24 hours ago.

That is SO true. And Digital Foundry is *very* guilty in this. I still remember they only ever showed the DLSS 2.0 trailing artifacts in Death Stranding once DLSS 2.2 or sth was out. *Never* did they mention any serious artifacts on DLSS 2.0 prior to that. Rince, repeat.

 

[Heartbreaker]

Well Digital Foundry (yeah-yeah Nvidia shills, whatever) also seems to have had access to it:

some artifacts are still very noticable, but all-in-all it looks quite good

I don't consider DF Shills, but I do note that when they get these exclusive first looks, they tend to be very "uncritical" (never bite the hand that feeds you). It's like when Ferrari flies people to Wine country when it's miserable winter elsewhere, everyone is very uncritical of the cars they are testing, well except that one guy that was and was never invited back.

I have strong memory of the first DF exclusive on 40 series Frame Generation, where they basically only praised it and said nothing critical. Later when they had their own copy when everyone did, they did a proper critical deep dive.

Will watch the video later.

 

[In2Photos]

It might be a good time to start talking about IPC decreases. If 5090 was doing well in raw perf (raster/RT) then they would not have focused so much on new fake frames. 4090 got away with it because it was actually good double normal perf and x4 if you really want fake frames, so that was ok.

Why add 1 frame in Raster when you can add 2 using AI!!!

 

[Heartbreaker]

You are probably right about the 5060. The reasons why i think 5060 Ti will be different this time (probably indeed using the same die and 192 bit bus) is because :

1. Last time there really was no competition (Navi 32 was too expensive to produce and Navi 33 was a stupid chip. barely faster than last gen) - this time there probably is and from both vendors
2.They'll want to have at least some sub 600$ SKUs to endice more budget oriented RTX 2060 SUPER; RTX 3060 12GB and possibly RTX 3070 holdouts to upgrade who don't want to dish out around 600 bucks for a 5070

What competition? AMDs 9060 series uses Navi 44, a very tiny chip which has a 128 bit bus and AMD is apparently sticking with slower GDDR6. Which means even if NVidia sticks with 128 bit bus, it will have much more memory BW than AMD...

 

[coercitiv]

I don't consider DF Shills

Me neither, I just see them as temporary Nvidia employees during the first month of each Nvidia release.

 

[CastleBravo]

I don't consider DF Shills, but I do note that when they get these exclusive first looks, they tend to be very "uncritical" (never bite the hand that feeds you). It's like when Ferrari flies people to Wine country when it's miserable winter elsewhere, everyone is very uncritical of the cars they are testing, well except that one guy that was and was never invited back.

I have strong memory of the first DF exclusive on 40 series Frame Generation, where they basically only praised it and said nothing critical. Later when they had their own copy when everyone did, they did a proper critical deep dive.

Will watch the video later.

Well, they seem to be very critical of DLSS3 in this video. I wonder why they didn't show all these flaws with DLSS3 when RTX4k was launched? Funny how that works.

 

[jpiniero]

Yep, but the 25% higher TDP will help clock rates for the 5070 relative to the 4070 even if your theory about a dense-prioritized implementation are correct.

GDDR7 probally draws more (even tho it's more efficient) and perhaps the design also runs hotter.

 

[SteveGrabowski]

So is 5070 basically a 4070 Super minus $50 with moar framegen?

 

[gdansk]

So is 5070 basically a 4070 Super minus $50 with moar framegen?

A pre-OC'd 4070 with more memory bandwidth and more fake frames. Also no optical flow hardware (greatest thing since sliced bread about 26 months ago).

Edit: not that there's anything wrong with most of that. But Nvidia saying it has 4090 performance is such crap.

 

[Gideon]

What competition? AMDs 9060 series uses Navi 44, a very tiny chip which has a 128 bit bus and AMD is apparently sticking with slower GDDR6. Which means even if NVidia sticks with 128 bit bus, it will have much more memory BW than AMD...

I was specifically talking about RTX 5060 Ti. A card that will probably cost 400-450$ USD (a 12GB version a bit more, but they might wait a year and do a SUPER card with that). If AMD indeed chooses to market Navi 44 in that price range they can GTFO. Considering that RTX 5070 MSRP is 550$ RX 7700XT MSRP is 450$, I'm quite confident that the competition will instead be cut-down Navi 48.

Not to argue but just to reiterate what I meant.

 

[Gideon]

Well, they seem to be very critical of DLSS3 in this video. I wonder why they didn't show all these flaws with DLSS3 when RTX4k was launched? Funny how that works.

Not only that, there are plenty of flaws visible in the B-roll they are intentionally not talking about. Look at the lower right part of the screen (the tree especially) in this shot for instance (timestamped to 4:55):

Ironically in the section they claim "ghosting and tearing" is fixed (yet you can still see ghosting on the character's hand, just improved considerably).

But all in all, both technologies certainly seem to have some merit. The transformer based upscaling does look better, despite its faults (and I have seen Ray Reconstruction in CP2077 at 1440p DLSS Quality and didn't like the blurry painting like resolve at all, thing they are finally showing here) and the 3x - 4x "fake frames" can also be quite useful for 240-300+ Hz monitors.

Not nearly as groundbreaking as Nvidia is hyping IMO, but definitely useful

 

[SolidQ]

AMD bite NV and they make own RDNA3?

https://twitter.com/x/status/1876544335044358184

 

[CakeMonster]

There is the additional memory bandwidth on 5090 at least. But yeah, I'm a tad disappointed about it still being on N4

 

[Gideon]

AMD bite NV and they make own RDNA3?

Speaking of this, this meme is quite well made: