Discussion Nvidia Blackwell in Q1-2025

[SolidQ]

That was here?

https://twitter.com/x/status/1799009730796585395

 

[MoogleW]

You don't need >50% higher BW for that.

@gdansk and what is the speed of the worst GDDR7? Even if it's only 24gbps, that's still 33% more than what AD106 has.

That would be 28gb/s. Any less and GDDR6X would make more sense

 

[blackangus]

That would be 28gb/s. Any less and GDDR6X would make more sense

You have to consider yields and cost.
24Gbps GDDR6x may have lower yields and cost more than the equivalent GDDR7x.
I dont know Im just saying there are other considerations.(each way!)

 

[Heartbreaker]

Hynix won't be volume shipping any form of GDDR7 until 2025.

SK hynix confirms GDDR7 memory mass production starts in Q1 2025 - VideoCardz.com

SK hynix GDDR7 in Q1 2025 at earliest Update (June 13, 2024): SK hynix clarifies that the company intends to start mass production in the fourth quarter of this year, contrary to what AnandTech posted. Original story: AnandTech had an opportunity to ask an SK hynix representative when the...

videocardz.com

While there have been some rumors of late 2024 RTX 5000, I think most rumors had them as 2025 cards from the start. Even though Samsung/Micron could be ready before that, I'd bet on these being 2025 cards.

 

[jpiniero]

Hynix won't be volume shipping any form of GDDR7 until 2025.

Micron and Samsung are sampling right now. I could see them lead with ProViz/AI cards first but I think you will see something from GB202 at least this year.

 

[gdansk]

Doesn't Nvidia usually prefer Micron?

 

[Heartbreaker]

Doesn't Nvidia usually prefer Micron?

On leading edge it often appears so. Weren't GDDR5X and GDDR6X essentially Micron/NVidia exclusives?

 

[del42sa]

NVIDIA CEO to deliver a keynote at SIGGRAPH 2024 on July 29

Keynote Outline:
1. Introduction
2. Discuss what Nvidia has done with AI
3. Discuss what Nvidia is currently doing with AI
4. Say "the more you buy, the more you save"
5. Discuss the future of AI
6. Ensure AI is said a few more times to raise stock price
7. Conclusion

 

[jpiniero]

NVIDIA CEO to deliver a keynote at SIGGRAPH 2024 on July 29

Be a good place to announce the Quadro Blackwell.

 

[Mopetar]

NVIDIA CEO to deliver a keynote at SIGGRAPH 2024 on July 29

Keynote Outline:
1. Introduction
2. Discuss what Nvidia has done with AI
3. Discuss what Nvidia is currently doing with AI
4. Say "the more you buy, the more you save"
5. Discuss the future of AI
6. Ensure AI is said a few more times to raise stock price
7. Conclusion

I'd laugh if JHH just walks out on stage to start a video recording of his presentation at Computex.

Anyone who complains gets strapped to their seat and made to watch again.

 

[blckgrffn]

I'd laugh if JHH just walks out on stage to start a video recording of his presentation at Computex.

Anyone who complains gets strapped to their seat and made to watch again.

It's an AI generated video and recap of the earlier presentation. Presented with a hologram.

We do what we must because we can.

 

[pcslookout]

$3000 to $4000 minimum for Rtx 5090

Enjoy

 

[Mopetar]

I think the price for the 5090 won't go up all that much. Instead we just get a deeper cut of the die. Since AMD isn't going to have anything that would even compete with the biggest Blackwell, Nvidia has almost no reason to release a consumer part outside of just needing something to top the existing 4090.

Instead expect whatever lets them salvage does that aren't even sold as cutdown professional parts. They make more money selling non-gaming chips and they can always release a 5090 SUPER or 5090 Ti later on once the high margin markets are saturated.

 

[jpiniero]

I think the price for the 5090 won't go up all that much. Instead we just get a deeper cut of the die.

Presumably this is going to be a monster sized die, close to the recticle limit. GB203 is going to be closer to AD102's size than AD103. The costs involved are going to be way higher.

 

[CakeMonster]

We'll see, but if the 5090 is far from the previous xx90 increase both with regards to performance and VRAM, I can't imagine it will sell well. Not that it matters, NV might just throw up their hands and say 'this is what it costs, nodes/demand etc forced our hand with pricing'. But if I end up skipping it, I doubt a later 5090Ti will very appealing, as the next generation draws closer anyway.

 

[jpiniero]

We'll see, but if the 5090 is far from the previous xx90 increase both with regards to performance and VRAM, I can't imagine it will sell well.

The 5090 should be fine and people will complain about the 5080's price but it should deliver a decent improvement over the 4080S. It's the rest of the lineup that will be questionable.

 

[poke01]

The 5090 should be fine and people will complain about the 5080's price but it should deliver a decent improvement over the 4080S. It's the rest of the lineup that will be questionable.

It will be questionable till the Super lineup arrives

 

[CP5670]

I like what I'm reading about the 5090 but not sure what I would even play on it. The 4090 already runs everything very well. Maybe VR sims could be improved further, but they are mainly CPU bottlenecked.

 

[igor_kavinski]

The 4090 already runs everything very well.

Nvidia will have snuck pathtracing into some under development titles by now so look forward to those making your 4090 look inadequate. Plus, texture memory compression (we have a dedicated hardware compression engine in there!) and some enhanced AI fakery to make you want to torture your 4090 so it dies an early death and you have a good excuse to get the 5090.

 

[poke01]

I like what I'm reading about the 5090 but not sure what I would even play on it. The 4090 already runs everything very well. Maybe VR sims could be improved further, but they are mainly CPU bottlenecked.

Path tracing could be one but that’s probably 5-6 years out. I would skip 5090 if you have 4090 unless you are a professional or hobbyist with money and it saves you time.

 

[poke01]

Nvidia will have snuck pathtracing into some under development titles by now so look forward to those making your 4090 look inadequate. Plus, texture memory compression (we have a dedicated hardware compression engine in there!) and some enhanced AI fakery to make you want to torture your 4090 so it dies an early death and you have a good excuse to get the 5090.

Or here’s a money saving idea. Turn PT off and enjoy your 4090. It will still excel at RT and Raster.

 

[CP5670]

I generally buy new cards that have a big improvement like the last two generations were, and sell my existing card before the value drops too much. In Cyberpunk PT just looks a bit different from RT, not necessarily better, and is definitely not worth the performance hit. Aside from PT, it's pretty hard to stress the 4090 at 4K. Some games do slow down but it's almost always due to a single thread CPU bottleneck.

 

[AdamK47]

Path tracing could be one but that’s probably 5-6 years out. I would skip 5090 if you have 4090 unless you are a professional or hobbyist with money and it saves you time.

A professional hobbyist that likes to save rendering time between frames.

 

[coercitiv]

True professional hobbyists keep their frames onscreen longer, so they can truly appreciate the image quality.

 

[SteinFG]

Looking at SM counts leaked by kopite, here's my guess for the lineup:

gpu	die	sm count, L2	gpu	die	sm count, L2
4090, 24GB	AD102 (144 SM)	128, 72MB	5090, 28GB	GB202 (192 SM)	~168, 84MB
4080 Super, 16GB	AD103 (80 SM)	80, 64MB	5080, 16GB	GB203 (84 SM)	84, 64MB
4070 Ti Super, 16GB	AD103 (80 SM)	66, 48MB	5070 Ti, 16GB	GB203 (84 SM)	~74, 48MB
4070 Super, 12GB	AD104 (60 SM)	56, 48 MB	5070, 16GB	GB203 (84 SM)	~60, 48MB
4060 Ti, 8/16 GB	AD106 (36 SM)	34, 32MB	5060 Ti, 12GB	GB205 (50 SM)	~46, 36MB
4060, 8GB	AD107 (24 SM)	24, 24MB	5060, 8GB	GB206 (36 SM)	~30, 32MB

5060 can get 12GB with 3GB GDDR7 packages, but who knows when they're coming