Nvidia Blackwell Architecture Speculation

[homerdog]

Could Blackwell be the worst GPU generation of all time?

People hated Turing back in the day, but atleast the 2060 was on par with a GTX 1080 and atleast it introduced exciting new features (although they weren't perceived like that back in the day due to missing software) like RT, DLSS, Tensor Cores, DX12U feature set etc.

Pretty much the entire problem with Blackwell is insufficient supply to meet demand and the resulting pricing. On paper 5000 series is a welcome improvement over 4000 series, especially compared to 4000 pricing at launch. But in real price terms, perf/$ is either stagnant or a regression from where we were a year ago. Still it's hard to say it's a terrible generation if NVIDIA literally can't make them fast enough.

"It sucks"
"Why?"
"Because it's always sold out"


Now when the HD2900 came out, that was a truly bad GPU. Slow compared to the competition, inefficient, and it either lacked MSAA hardware or it was bugged. Here it is losing to a X1950XTX with 4xMSAA:

ATI Radeon HD 2900 XT: Calling a Spade a Spade

www.anandtech.com

And remember R600 had a 512bit memory bus

GeForce FX5800 (NV30 mentioned by @trinibwoy) was another legendary stinker.

 

[pjbliverpool]

Also the 5090 is an absolute monster. It completely annihilates any competition. I'm still impressed whenever I see a multi GPU performance bar chart at how far ahead of the rest of the pack it is. There's a reasonably gentle curve all the way up to the 4090 and then boom, the 5090 is having a party out there all on it's own. Hard to call that the worst GPU family ever.

 

[DuckThor Evil]

Fermi

Fermi wasn't that bad Yeah 480 had a high power draw for the time, making the reference model noisy, but it was the fastest card around and 470 was pretty good all around. Fermi then had a re-spin later on the same year and the 500-series pretty much fixed everything, offering great products in all categories.

 

[techuse]

NV30, Fermi, R600 were all worse I think. Despite all the bellyaching Blackwell is still the fastest and most efficient architecture on the market. So not sure how it can be the worst of all time.

Fermi doesn’t belong with the other two and I would consider it better than Blackwell.

 

[trinibwoy]

Fermi doesn’t belong with the other two and I would consider it better than Blackwell.

I had a 580 and thought it was great. It’s nowhere near 5090 level though.

 

[DegustatoR]

When/how does it get updated? Is it the driver or the NV App that does it?

When Nvidia pushes updates on their side. I.e. users have no control or visibility on this process, it happens in the background without any indication, similar to how updates to game profiles can be done via the same mechanism.

 

[trinibwoy]

When Nvidia pushes updates on their side. I.e. users have no control or visibility on this process, it happens in the background without any indication, similar to how updates to game profiles can be done via the same mechanism.

First time hearing that. Is this behavior documented somewhere?

 

[techuse]

I had a 580 and thought it was great. It’s nowhere near 5090 level though.

Fermi offered larger performance increases over its predecessor up and down the stack at better prices. It also brought a brand new architecture with a lot of foundational improvements.

 

[homerdog]

First time hearing that. Is this behavior documented somewhere?

Yea this is pretty interesting. I assumed it was tied to driver updates. Using the DLSS indicator in FF7 Rebirth with the "Latest" DLSS override it says:

Render Preset K
DLSSv3 v310.2.1 (nvapp_override)

I think this is DLSS4 despite saying DLSSv3, because the image is clearly better (less blurry) when using the override. Anyway through a couple of driver updates the DLSS indicator text hasn't changed. DLSSv3 v310.2.1

BTW the way FF7 Rebirth does dynamic resolution scaling with DLSS is amazing. I wish every game was like this. It's way harder to notice the internal resolution dropping than to notice dropped frames.

 

[DegustatoR]

I assumed it was tied to driver updates.

Driver is still shipping with DLSS SR v3.8.10 (with limited number of models in it) presumably for DX-SR support.
Interestingly the latest driver also has DLSS FG v310.2.1 DLL, no idea what for. Possibly so that MFG overrides would work on PCs w/o any Internet connection? Still need to d/l the driver somehow though.
DLLs for NvApp overrides are downloaded by driver updater after installation and are kept in %ProgramData%\NVIDIA\NGX\models\ separately from any driver install.

 

[trinibwoy]

Fermi offered larger performance increases over its predecessor up and down the stack at better prices. It also brought a brand new architecture with a lot of foundational improvements.

NVIDIA GeForce RTX 5090 Founders Edition Review - The New Flagship

NVIDIA's GeForce RTX 5090 is the fastest graphics card ever released. It comes with a whopping 32 GB VRAM, and support for multi-frame generation, which achieves hundreds of FPS easily. We also managed to disassemble the card, and our review includes performance testing without upscaling or...

www.techpowerup.com

NVIDIA GeForce GTX 480 Fermi Review

Today marks the release of NVIDIA's new GeForce Fermi architecture. After excruciating months of delays, NVIDIA has finally given the green light for their new products. The GeForce GTX 480 offers all the latest features like DirectX 11, Tesselation, gaming on multiple monitors and GPU...

www.techpowerup.com

 

[trinibwoy]

Driver is still shipping with DLSS SR v3.8.10 (with limited number of models in it) presumably for DX-SR support.
Interestingly the latest driver also has DLSS FG v310.2.1 DLL, no idea what for. Possibly so that MFG overrides would work on PCs w/o any Internet connection? Still need to d/l the driver somehow though.
DLLs for NvApp overrides are downloaded by driver updater after installation and are kept in %ProgramData%\NVIDIA\NGX\models\ separately from any driver install.

It’s crazy how applications can do this without user knowledge. Kinda scary actually.

 

[homerdog]

Driver is still shipping with DLSS SR v3.8.10 (with limited number of models in it) presumably for DX-SR support.
Interestingly the latest driver also has DLSS FG v310.2.1 DLL, no idea what for. Possibly so that MFG overrides would work on PCs w/o any Internet connection? Still need to d/l the driver somehow though.
DLLs for NvApp overrides are downloaded by driver updater after installation and are kept in %ProgramData%\NVIDIA\NGX\models\ separately from any driver install.

Is it supposed to use 310.2.1 when I set the override to "Latest"?

 

[techuse]

NVIDIA GeForce RTX 5090 Founders Edition Review - The New Flagship

NVIDIA's GeForce RTX 5090 is the fastest graphics card ever released. It comes with a whopping 32 GB VRAM, and support for multi-frame generation, which achieves hundreds of FPS easily. We also managed to disassemble the card, and our review includes performance testing without upscaling or...

www.techpowerup.com

NVIDIA GeForce GTX 480 Fermi Review

Today marks the release of NVIDIA's new GeForce Fermi architecture. After excruciating months of delays, NVIDIA has finally given the green light for their new products. The GeForce GTX 480 offers all the latest features like DirectX 11, Tesselation, gaming on multiple monitors and GPU...

www.techpowerup.com

PPW isn’t the sole defining characteristic of a GPU.

 

[trinibwoy]

PPW isn’t the sole defining characteristic of a GPU.

Didn’t say it was. However Fermi was a disaster on that metric. There’s nothing wrong with Blackwell architecture. The problems are price and availability.

 

[Subtlesnake]

PPW isn’t the sole defining characteristic of a GPU.

Fermi wasn't great for PPA either. Whereas Blackwell leads in both PPA and PPW and is a considered a disappointment simply for not making a bigger improvement over Ada, on the same process.

 

[DegustatoR]

Is it supposed to use 310.2.1 when I set the override to "Latest"?

As of right now that's the latest version with all models and it's what the driver downloads for overrides so yes.

 

[techuse]

Didn’t say it was. However Fermi was a disaster on that metric. There’s nothing wrong with Blackwell architecture. The problems are price and availability.

Never said there was anything wrong with Blackwell, only that I consider Fermi better.

 

[trinibwoy]

Never said there was anything wrong with Blackwell, only that I consider Fermi better.

Sure and I disagree. Fermi wasn’t the best architecture on the market when it launched while Blackwell is.

 

[troyan]

And it doesnt make sense to use Fermi. Fermi was on 40nm. GT200 is a better comparision to Blackwell while it was on the same 65nm as G92b.