Discussion Nvidia Blackwell in Q1-2025

[Win2012R2]

From nvidia.com - 5090:

	Tensor Cores (AI)	5th Generation 3352 AI TOPS
	Ray Tracing Cores	4th Generation 318 TFLOPS
	Boost Clock (GHz)	2.41
	Base Clock (GHz)	2.01"

4090 -

	Ray Tracing Cores	3rd Generation 191 TFLOPS
	Tensor Cores (AI)	4th Generation 1321 AI TOPS
	Boost Clock (GHz)	2.52
	Base Clock (GHz)	2.23

Clock regression, ray tracing supposedly a lot better, but why then Cyberpunk barely faster if you discount new frame gen?

What a missed opportunity for AMD to catch up at top end in proper raw perf.

 

[gdansk]

Hell no, these are always recent. Benchmarketing and comp analysis guys never sleep before launch.

I remember checking 2 month old footnotes for the Zen 5 slides. So *good* comp analysis guys never sleep but AMD's were napping.

 

[gdansk]

Doesn't seem like it based upon the benchmarks NV provided.

But it does. stuff below GB202 seems pretty good?

Can either of you post the article(s)? I couldn't find any raster examples.

 

[adroc_thurston]

ray tracing supposedly a lot better

Very incrementally so, actually.

I remember checking 2 month old footnotes for the Zen 5 slides. So *good* comp analysis guys never sleep but not AMD's.

AMD ones don't sleep either, but CPU timelines are far lazier than GPU ones.

 

[Hulk]

Yes. It's well know that Fake Frame generation adds lag. Generally 1 full frame time, of the real frame rate. So if you are running 60 FPS real, then you add about 17 ms of lag when you turn on fake frames.

Not much should change in lag by adding more fake frames.



I don't like predicted, because it's not a future looking operation. They are interpolations between two known frames. You are always figuring out positions in between two frame you already have.




I would think it the other way around. The higher your native frame rate, the smaller the window between frames, to insert your fake ones.

60 FPS = 17ms window to insert fake frames.
240 FPS = 4ms windows to insert fake frames.

I'm not following how they can have interpolated frame without looking into the future. You would be playing in the present and seeing the past. It doesn't make sense.

Or I don't understand it.

 

[Win2012R2]

I assume we are at 128MB L2 now.

Considering piss poor perf improvement it seems like a very optimistic assumption... seems more likely they indeed cut it since mem bandwidth was higher.

 

[reaperrr3]

It's lower than I expected. Only 1.77x membw...

What were you expecting?
For all the flak they're getting, leakers have said 28Gbps memory for many months, and they were right on this one. 1.33x interface + 1.33x mem clock = 1.77 bw, completely in line with expectations.

Leakers have been wrong on details that can be easily changed very late in the game or are easy to hide, but they were spot on with memory configs.

It's fine but outside of moar interpolated frames even gb202 seems kinda mid I guess?

5090 seems a bit off, given the fattest spec bump of all.
Like with 4090 it was very evident the chungus is more chungus than the usual.

The 5090 being CPU limited I could believe. But the others? Nah...

If NV didn't significantly reduce driver/CPU overhead vs. Ada, 5080 (assuming it's actually close to 4090 in perf) won't be that much less CPU-limited than a 4090, and the 5090 would obviously be even more so.

But yeah, I'd expect a bigger uplift for the 5090 at 8k.
Maybe Nvidia considers native 8k too niche, or - more likely - conflicting too much with their talking up of DLSS.

 

[adroc_thurston]

But yeah, I'd expect a bigger uplift for the 5090 at 8k.

yeah it's weird for them to not flex a major membw bump at high res.

 

[Jan Olšan]

How is the performance on most of those cards only "2X" if the marketing figures include multiframe generation (which quadruples FPS compared to DLSS3's doubling)?

Isn't that kinda silly? Or is it so inefficient you get nowhere near linear scaling? If that is with cherry picking ON even, that is kinda surprising.

 

[adroc_thurston]

How is the performance on most of those cards only "2X" if they include multiframe generation (which quadruples FPS compared to DLSS3's doubling)?

Well generating fake frames isn't quite free.

Or is it so inefficient you get nowhere near linear scaling?

Well yeah. You need to actually spend time to conjure the frames.

 

[gdansk]

What were you expecting?

32 @ 30!
For AI needs.

 

[Win2012R2]

I couldn't find any raster examples.

Far Cry 6 is the only one I've seen, with ray tracing on too

 

[igor_kavinski]

you can buy n-CCD V$ in 9684X.

I just need the 9184X. Can't find it at a decent price anywhere!

 

[gdansk]

How is the performance on most of those cards only "2X" if the marketing figures include multiframe generation (which quadruples FPS compared to DLSS3's doubling)?

Isn't that kinda silly? Or is it so inefficient you get nowhere near linear scaling? If that is with cherry picking ON even, that is kinda surprising.

4 frames (1+3) now vs 2 frames (1+1) last gen. 2x, right?

 

[Hulk]

From nvidia.com - 5090:

	Tensor Cores (AI)	5th Generation 3352 AI TOPS
	Ray Tracing Cores	4th Generation 318 TFLOPS
	Boost Clock (GHz)	2.41
	Base Clock (GHz)	2.01"

4090 -

	Ray Tracing Cores	3rd Generation 191 TFLOPS
	Tensor Cores (AI)	4th Generation 1321 AI TOPS
	Boost Clock (GHz)	2.52
	Base Clock (GHz)	2.23

Clock regression, ray tracing supposedly a lot better, but why then Cyberpunk barely faster if you discount new frame gen?

What a missed opportunity for AMD to catch up at top end in proper raw perf.

Real compute substitued for fake AI compute. Fake frame rate increase, real frame rate not so much.
We'll know the deal soon enough when the hardware hits the streets and reviewers give it a real workout.

At this point I'm very interested in how the AMD GPUs will perform as well as the forthcoming Intel B770/750.

 

[Heartbreaker]

I'm not following how they can have interpolated frame without looking into the future. You would be playing in the present and seeing the past. It doesn't make sense.

Or I don't understand it.

It's because they are accumulating frames in a buffer instead of showing them right away. You are seeing 1 frame late.

Instead of showing the current frame right when it happens, they hold it in a buffer and use that new frame, with the previous real frame, to interpolate the frame(s) between old and the new one.

Now instead of showing you the new frame, they show you the one(s) they interpolated between old and new.

Then after they show you that interpolated frame(s), they finally show you the real frame they had buffered.

Nothing is predicted, it's just in between frames.

You are seeing the past, you are lagging by about 1 frame time. If your real frame rate is good enough, this lag won't be too noticeable. Old LCD often used to lag by several frames before they started having game modes.

 

[coercitiv]

I'm not following how they can have interpolated frame without looking into the future. You would be playing in the present and seeing the past. It doesn't make sense.

Or I don't understand it.

You're always playing in the past. There's always latency between your actions and system response (the computer, the game). Playing games is in fact a process of continuous corrections based on information about past actions we receive from the system (what we see on the monitor and consider as "present").

This is why shooter players value latency reduction:

• less corrections they have to make before aiming for a perfect shot
• less surprises like getting shot before they get to see their opponent (before it gets rendered on their system)

LE: Frame Generation adds to that latency, and to compensate for this loss, both Nvidia and AMD make optimizations to driver & game behavior in an attempt to win back as much as possible from that loss. The catch is the two features are independent of each other, the idea that latency optimizations make FG a win-win is false from the start.

 

[GodisanAtheist]

Bad news is that $549 seems a bit high for fake frame generation card and only 12GB VRAM.

Good news is that gives AMD and Intel up to $500 price point to deliver a compelling product. Hope they don't mess up.

-Narrator: They will.

 

[gdansk]

At this point I'm very interested in how the AMD GPUs will perform as well as the forthcoming Intel B770/750.

Do you not have a graphics card? I'm curious

 

[Win2012R2]

Will wait for 3rd party reviews, but most likely will skip 5090 now, and I've stopped playing games for months waiting for it!

 

[Ranulf]

4090 owners react:

 

[Jan Olšan]

Well generating fake frames isn't quite free.

Well yeah. You need to actually spend time to conjure the frames.

Yeah but to this extent? Getting 2X at the end is like they blew all the RTX 5090 raw performance improvement just for the extra fake frames, no real FPS boost, lol.

That's ridiculously MEH if true, I can't see this being real.
That's more MEH than I would have ever expected (or hoped for as an AMD fan).

 

[Heartbreaker]

4 frames (1+3) now vs 2 frames (1+1) last gen. 2x, right?

Plus those numbers seldom worked out as direct double of base frame rate for original frame generation either.

 

[GodisanAtheist]

4090 owners react:

-Beautiful thing about NV marketing is conning folks into unloading their current gen hardware for fire sale prices before the next gen lands.

There are definitely guys out there selling their 4090's for $1000 thinking they got a killer deal cause "5070 is 4090 for $550".

 

[adroc_thurston]

https://twitter.com/x/status/1876503660693512365

only 25% moar TSE score? rough.