Discussion Nvidia Blackwell in Q1-2025

Page 181 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

basix

Senior member
Oct 4, 2024
287
570
96
Well, 16 GPCs are visible on the diagram. It does not mean, that the chip will really look like this.

But I think 16 GPCs is reasonable. I can think of a few reasons for that:
  • Better scaling in general, because 1.33x GPCs (ROPs anyone? :D)
  • Power of 2 scaling for ML/AI Workloads: A100, H100 and B200 all have 8x GPCs. Rubin HPC might have 8 or 16 and Rubin CPX would have 16 GPC as well. ML/AI workloads like power of 2 divisions (MI350X did go back to 256 CU because of that, MI300X featured 320 CU)
  • GauRast: Gaussian Splatting Acceleration (Neural Rendering) within the rasterizer. 1.33x GPCs will bring a boost there (see the GauRast paper https://arxiv.org/html/2503.16681v1)
  • Transformer Attention Acceleration (see GB300 or Rubin CPX) does benefit from exponential functions. On GB300 Nvidia says they pimped the SFU within the SM for that (or emulate SFU EX2 functions). GauRast from the previous point will introduce additional EXP-Units within the rasterizer (GPC frontend). This might create some synergies between Neural Rendering and general ML/AI.
 
Last edited:

jpiniero

Lifer
Oct 1, 2010
17,015
7,409
136
It's still a bad arch even with more memory.

Tough to get that much better performance while using the same node and a mandate to keep transistor budgets flat if not lower.

It definitely would have looked better if they had used N4P instead of N5.
 
  • Like
Reactions: MoogleW

Win2012R2

Golden Member
Dec 5, 2024
1,248
1,292
96
Tough to get that much better performance while using the same node and a mandate to keep transistor budgets flat if not lower.
Dropping 32bit PhysX was Borderlands 2 criminal...

But seriously the whole package clearly does not work as they intended, maybe it's -10% clocks, but just feels something gone wrong there, with all that "AI" money Nvidia should be able to afford making a proper gaming card.

It definitely would have looked better if they had used N4P instead of N5.
Ada was designed using special TSMC 4N, no chance they used N5 for Blackwell.
 

jpiniero

Lifer
Oct 1, 2010
17,015
7,409
136
Ada was designed using special TSMC 4N, no chance they used N5 for Blackwell.

4N is an nVidia N5 special.

Rubin, btw, looks like what I thought Blackwell is going to be. It looks like they will double down (if not more) on tensor cores, leaving little room for raster or RT improvements beyond higher clocks. And that's assuming they don't cut core counts at the lower end to keep costs in line with N3's way higher prices.
 
Last edited:

MrTeal

Diamond Member
Dec 7, 2003
3,919
2,708
136
The 5070 Ti Super at the same price point would be a nice bonus if you're looking now, but going from 16GB to 24GB isn't going to be a big difference for that card. The only Super that looks interesting at all is the 5070 Super.
 

CP5670

Diamond Member
Jun 24, 2004
5,683
783
126
It looks like you can upgrade 4090 to 5090 for just $100-300 if you buy and sell in the right places. I don't have much need for it but might do this before the 4090 loses value. The MSI 4090 liquid model actually sells for more than its original price 3 years ago, maybe because its small size makes it good for multi-gpu setups for AI. Some of the base 5090 cards are $2000 if you wait for a microcenter drop.

One issue I see is the huge size means I have to move my soundcard down to a slot where the M.2 drive goes down to half speed. Maybe a pcie riser cable would avoid this.

Update, yes the riser cable works nicely for the sound card. It's a bit clunky to fit into the last ATX slot because the motherboard's pcie slot is in the way, but I managed. Now there are easily 4 slots for a huge gpu.
 
Last edited:

CP5670

Diamond Member
Jun 24, 2004
5,683
783
126
I sold the 4090 for $2100 on ebay ($1800 after their fees) and paid $1900 for it three years ago (more in today's dollars). One of the best cards I have ever owned with the AIO and huge performance boost over anything older. The 4090 demand is driven by datacenter conversions and the China export ban. I can't remember ever seeing any card keep its value up for this long.

Looking around for an msrp 5090 now, they show up in Microcenter every few weeks. Among the NYC area ones the NJ one has a much lower tax rate. I loved the MSI AIO but the 5090 version costs way too much. The AMD IGP is useful to keep the PC usable without a card.
 
  • Like
Reactions: Tlh97 and Mopetar

EXCellR8

Diamond Member
Sep 1, 2010
4,111
925
136
The 4090 demand is driven by datacenter conversions and the China export ban. I can't remember ever seeing any card keep its value up for this long.
I was looking into getting one recently and was wondering why a GPU from late 2022 was commanding that much, so I guess that makes sense. FFS there's always something... it was a BEAST when it first launched, but still over two racks 3 years later? Bonkers.
 
  • Like
Reactions: Mopetar and CP5670

CP5670

Diamond Member
Jun 24, 2004
5,683
783
126
I got a pretty good deal on a open box 5090 Astral for $2230 at Best Buy (new price is $3360), or $2146 with tax and 10% off for signing up for their credit card. So the total upgrade cost was only $346. I had to check out 4 or 5 stores to find this, open box prices are at the store's discretion and different everywhere.

I'll set it up tomorrow. The card is gigantic compared to the tiny 4090 liquid I had before.
 
  • Like
Reactions: Mopetar and Elfear

CP5670

Diamond Member
Jun 24, 2004
5,683
783
126
I have two months to return it if it has problems. I have bought used gpus in the past and had no issues, the 4090 was also from a craigslist seller and opened.