Nvidia Blackwell Architecture Speculation

[DegustatoR]

I don't see anything "disappointing" in getting a +10-20% performance increase at the same price.

 

[Charlietus]

Guess nobody is really interested in talking about the 5080 reviews, eh? lol

Turns out we didn't need to wait for reviews to know this was gonna be extremely disappointing.

If you consider that the 1000$ MSRP will not exist in Europe, and will be almost non existent in the US, things are worse than what reviewers are saying. 1200€$ should be considered the real price for this thing (and that's probably conservative).

Nvidia MSRP prices are a smokescreen.

Ps: just saw the official price for the 5080 in Italy, and it's 1200€ base price

 

[CarstenS]

I have been keen to find out the L2 cache for months, and I agree it’s a bit odd to leave it out. They would have had to shrink the relative size of the cache footprint to keep the same ratio on GB202 (ie 128MB) - not impossible (AMD did it with Zen 5) but once we knew the chip was 22% bigger it never seemed likely. My guess is it’s still 96MB (maybe a bit smaller even) and they don’t want silly stories about less cache for more money or some such nonsense. Worth noting that unless they needlessly disable cache on the 5090 again, it’ll probably end up with more overall than the 4090 (even if the SM/L2 ratio is a bit smaller).

Edit: Replied before seeing the post on the next page, no surprises there.

128 MByte L2 for GB202 has been confirmed in the meantime since my posting below.

Nvidia Blackwell Architecture Speculation

My guess is that the outer yellow one is the I/O daughterboard mezzanine connector, and the green one is for the PCIe Gen5 daugtherboard. Could be vice versa but this seems to make it easier to route the ribbon cables. NVIDIA's Unreleased TITAN/Ti Prototype Cooler & PCB | Thermals, Acoustics...

forum.beyond3d.com

 

[Scott_Arm]

If you consider that the 1000$ MSRP will not exist in Europe, and will be almost non existent in the US, things are worse than what reviewers are saying. 1200€$ should be considered the real price for this thing (and that's probably conservative).

Nvidia MSRP prices are a smokescreen.

The same was probably true of 4080 and 4080 super. At least I never saw it. Gets super hard when you start comparing prices of AIB models, but I do agree that using the FE pricing as some kind of comparison point seems limited.

 

[troyan]

Guess nobody is really interested in talking about the 5080 reviews, eh? lol

Turns out we didn't need to wait for reviews to know this was gonna be extremely disappointing.

Honestly, even Blackwell as an architecture is super underwhelming. You can argue it's setting up some improvements that will show up later down the line, but by the time these things are more commonly incorporated into games, there will be better GPU's out.

How is the 5080 a disappointment when it delivers more for the same price?

 

[digitalwanderer]

I don't see anything "disappointing" in getting a +10-20% performance increase at the same price.

Considering nVidia's promo materials claimed it was more than 2x the performance of the 4080, yes 10-20% is a let down.

I know the used MFG for their comparison but that's on them for getting the hype too high.

 

[DegustatoR]

Considering nVidia's promo materials claimed it was more than 2x the performance of the 4080, yes 10-20% is a let down.

I know the used MFG for their comparison but that's on them for getting the hype too high.

Well it is 2X with MFG vs FG, and this is still a valid comparison.

 

[DegustatoR]

You wouldn't see anything disappointing if Jensen personally ran over your dog.

Reported. This is seriously tiring.

And it's more like 10% average

As with 5090 there is zero point in looking at "average" which seem to be CPU limited to a significant degree or just full of buggy results.

 

[troyan]

Blackwell delivers more for the same price. So it is actually an improvement.

 

[troyan]

It delivers ~85% of the 4090 performance for 62% of the price. Good luck finding a faster GPU for less.

 

[Scott_Arm]

Techreport it's about 13% faster than 4080 super at 4k. Really comes down to the games being tested.

NVIDIA GeForce RTX 5080 Founders Edition Review

The NVIDIA GeForce RTX 5080 is out now. Priced at $1000 it includes all the new GeForce RTX 50 features and offers good gaming performance. On the other hand, the gen-over-gen performance increase is smaller than expected, but NVIDIA didn't raise their pricing either.

www.techpowerup.com

 

[GLS28R]

Based on AW2 - RTX 2080 Ti Runs about 5 ms better with it on vs off

That's amazing. I wonder what the Blackwell advantage is. Nvidia basically put all of their resources into neural rendering and improving the RT cores. You would hope that Blackwell has a significant performance advantage in titles that support RTX Mega Geometry.

 

[troyan]

And Blackwell seems to perform better in UE5 games with Software Lumen. So the difference between 5080 and 4090 in 1440p will be around 10%.

 

[homerdog]

Techreport it's about 13% faster than 4080 super at 4k. Really comes down to the games being tested.

NVIDIA GeForce RTX 5080 Founders Edition Review

The NVIDIA GeForce RTX 5080 is out now. Priced at $1000 it includes all the new GeForce RTX 50 features and offers good gaming performance. On the other hand, the gen-over-gen performance increase is smaller than expected, but NVIDIA didn't raise their pricing either.

www.techpowerup.com

You got me excited for a second. RIP TechReport

Does the 5090 really have 128MB of L2? Techpowerup still lists 98MB.

 

[troyan]

98MB is active: https://images.nvidia.com/aem-dam/S...ell/nvidia-rtx-blackwell-gpu-architecture.pdf

 

[raytracingfan]

DXR 2.0 must be coming. This is a pretty significant departure from how ray-tracing is currently being handled.

I'm really curious how Mega Geometry works on a gpu that lacks the cluster intersection testing in the RT core. You'll still get big benefits on the CPU side, but I wonder how the intersection testing is handled and what the performance penalty is.

The new Sphere and Linear Swept Spheres intersection testing is a big change to RT as well. It's the first time in a long time (forever?) that mass market graphics cards have come with hardware support for non-triangle 3D graphic primitives. This post ended up being rather prescient.

What are the chances of adding support for intersecting curve primitives in the next iteration of DXR? Apple recently added it to Metal but it’s unclear whether it’s hardware accelerated. Optix supports it in CUDA.

Theoretically it will be faster to intersect curve primitives for stuff like hair which I believe currently uses highly tessellated lines & triangles.

DXR 2.0, standardization of SER, Cooperative Vectors, and Shader Model 7 with SPIR-V support are enough new features for Microsoft to announce DirectX 13.

 

[digitalwanderer]

Well it is 2X with MFG vs FG, and this is still a valid comparison.

Seems more deceptive than valid to me but you are welcome to your own opinion on the matter.

 

[arandomguy]

You got me excited for a second. RIP TechReport

Does the 5090 really have 128MB of L2? Techpowerup still lists 98MB.

GB202 has 128MB.

RTX 5090 as configured has 98MB.

As an aside RTX 5090 and RTX 4090 are both roughly cut down proportionally the same relative to the full die in all aspects I believe.

 

[DegustatoR]

Seems more deceptive than valid to me but you are welcome to your own opinion on the matter.

Nothing deceptive in comparing 2X FG with 4X FG.

 

[Flappy Pannus]

Nothing deceptive in comparing 2X FG with 4X FG.

As a measure of performance, there is. 4X produces more artifacts than 2X so they are not directly comparable.

Obviously generated frame quality matters, or else you would dismiss DLSS4 FG because we have the lossless scaling app to do that (and more) already. Few people are going to make that argument because the resulting quality isn't up to DLSS4.