Discussion Nvidia Blackwell in Q1-2025

[MrTeal]

I'm hoping for a $699 5090, but I am fully expecting that to be only slightly less likely than a $1600 32GB 5090. We're a year and a half in and you can't even buy a 4090 for $1600 today. Papa Jensen isn't going to bless us with 33% more shaders, bus width and VRAM on a brand new process for the same price as AD102 unless the AI market collapses in the next 6 months. I'd be pretty happy with $1800 and not surprised at 2 grand.

 

[jpiniero]

I'm hoping for a $699 5090, but I am fully expecting that to be only slightly less likely than a $1600 32GB 5090. We're a year and a half in and you can't even buy a 4090 for $1600 today. Papa Jensen isn't going to bless us with 33% more shaders, bus width and VRAM on a brand new process for the same price as AD102 unless the AI market collapses in the next 6 months. I'd be pretty happy with $1800 and not surprised at 2 grand.

Even if AI collapses, you are still talking about an expensive node and expensive memory.

I'd still say 512-bit is the realm of Titan.

 

[Aapje]

I'd still say 512-bit is the realm of Titan.

Nvidia seems to have decided since Ampere that it works better to market that level of cards as the x090, rather than use the Titan branding. The bigger question is whether they even want to sell that level of bandwidth and VRAM capacity in the prosumer/consumer market, or whether they only want to sell that to the professional market.

However, it seems like a huge waste to cut that big bus down to 384 bit, and would probably starve the GPU. But I can see them cut the 5090 down to 448 bits, like they took 64 bits off the 3080. Then it would be a 28 GB 5090, so an improvement over the 4090, but not too much extra VRAM either. And if 3 GB modules come available in a year, they could just keep 2 GB modules on the 5090, but put 3 GB modules on the pro cards, for a total of 96 GB with a clamshell. That would actually increase the gap between the expensive professional and 'cheap' prosumer/consumer cards.

 

[jpiniero]

Nvidia seems to have decided since Ampere that it works better to market that level of cards as the x090, rather than use the Titan branding.

I think the 5090 is going to be the deepest cut GB202, which the 4090 is not. 384-bit @ 28 gbps is still 33% more bandwidth than the 4090.

Could do 448-bit for the 5090 Ti.

 

[Aapje]

I think the 5090 is going to be the deepest cut GB202, which the 4090 is not.

Yes it is, unless you count the special version due to the Chinese sanctions. But that is still called a 4090.

 

[jpiniero]

Yes it is, unless you count the special version due to the Chinese sanctions. But that is still called a 4090.

I was referring to the not-released 4080 Ti... but there's an Quadro that's even deeper of a cut than what the 4080 Ti likely would have been.

So what I was thinking is that the 5090 would be about 30% faster than the 4090... and if nVidia felt like they needed to impress people, they could go ahead and release the Titan as well at a crazy price.

 

[Zor Prime]

I'm paying whatever nVidia wants for a 5090 at release. nVidia just needs to get it out the door, I could use an upgrade from my 3090 (I skip generations.)

 

[CakeMonster]

People do want the best chip for the whole generation when they get the xx90 though. The price also reflects that. If NV cripples it too much, with the plan of releasing a successor with significantly more unlocked cores as well as significantly higher memory bandwidth, there will be backlash. It would not be the same as the relatively tame 3090Ti or a theoretical 4090Ti.

 

[Zor Prime]

^ I agree. That's why I'm willing to pay whatever nVidia wants for a 5090, particularly if performance is off the chart. Just buy it and don't have to think about a thing for approx 2 years. Can totally coast through the next generation by putting some settings on medium or low, disable shadows and the like while leaving resolution and textures on max -- what I do with latest release titles and still getting 60+ FPS in 4K with my 3090. Totally willing to pay the price. Let's hope it's under $2.5K, though.

So how about it nVidia, I got money for ya. Where's my 5090?

 

[jpiniero]

or a theoretical 4090Ti.

The 4090 is seemingly so memory bottlenecked that nVidia could have easily just made a new SKU with all the cache and 24 gbps GDDR6X and gotten another 15%. The 3090 vs 3090 Ti doesn't really compare.

 

[SteinFG]

NV has decided to go for GDDR7 for top 3 dies, then we should be seeing 512-bit GDDR7 after all:-

1. RTX-5080 20GB GDDR7 320-BIT 80 MB
2. RTX-5080 24GB GDDR7 384-BIT 96 MB
3. RTX-5090 32GB GDDR7 512-BIT 112 MB ?

Pending confirmation...

384 bit not confirmed by leaks, only 512 and 256

 

[Aapje]

People do want the best chip for the whole generation when they get the xx90 though. The price also reflects that. If NV cripples it too much, with the plan of releasing a successor with significantly more unlocked cores as well as significantly higher memory bandwidth, there will be backlash. It would not be the same as the relatively tame 3090Ti or a theoretical 4090Ti.

I am quite sure that if the top RDNA5 card beats the 5090 at the end of 2025, there will be a Super-refresh or Ti to put Nvidia on top again. If you take the most competitive person you now, multiply his competitiveness by ten and inject cow DNA into him, you get Jensen. He really, really, really doesn't like to lose.

And frankly, I think that people like you would grumble a bit, but will then just get the 6090 or 7090, next time, assuming that Nvidia is still on top then. The very fact that you go for the top-end card even for these prices, makes me disbelieve that you will let this kind of thing impact your next buying decision. You will again say: "Only $2999 for a 6090, where do I sign up!"

 

[jpiniero]

NVIDIA Blackwell "B100" to feature 2 dies and 192GB of HBM3e memory, B200 with 288GB - VideoCardz.com

NVIDIA to announce Blackwell at GTC 2024 The next-gen HPC GPU to feature 192 GB of capacity. Tomorrow at the GTC 2024 keynote, Jensen Huang is said to announce the next-generation GPU architecture called Blackwell. Similarly to what we observed in the past, NVIDIA will use the same underlying...

videocardz.com

Apparently GB100 is 2 dies and has 192 GB of memory (same as MI300). A version with 288 GB will be released later.

And yes, they are announcing it before they even ship GH200.

 

[CP5670]

^ I agree. That's why I'm willing to pay whatever nVidia wants for a 5090, particularly if performance is off the chart. Just buy it and don't have to think about a thing for approx 2 years. Can totally coast through the next generation by putting some settings on medium or low, disable shadows and the like while leaving resolution and textures on max -- what I do with latest release titles and still getting 60+ FPS in 4K with my 3090. Totally willing to pay the price. Let's hope it's under $2.5K, though.

So how about it nVidia, I got money for ya. Where's my 5090?

The 4090 already feels like this to me. It handles anything I play perfectly, except for a few badly optimized games that look the same on high vs ultra settings or with RT on/off. I would get a 5090 but can't think of any new experiences it would enable. The 4090 actually did enable a lot of new things that the 3090 struggled with, like VR sims.

I would expect it to be at least $2500. They have zero competition at the high end except the 4090, which still goes for $2000 or more on most sites.

 

[Zor Prime]

The 4090 already feels like this to me. It handles anything I play perfectly, except for a few badly optimized games that look the same on high vs ultra settings or with RT on/off. I would get a 5090 but can't think of any new experiences it would enable. The 4090 actually did enable a lot of new things that the 3090 struggled with, like VR sims.

I would expect it to be at least $2500. They have zero competition at the high end except the 4090, which still goes for $2000 or more on most sites.

You're in good territory to do what I do, leap-frog generations so you're not spending money each time a new generation comes out unless you positively have to have everything on max in the interim between generations. This is why I'm now just buying the best at the advent of a new generation after seeing how far my 3090 carried me (it runs faster than a stock 3090Ti with ease. I'll miss EVGA.)

 

[Tigerick]

The NVIDIA GTC 2024 Keynote Live Blog (Starts at 1:00pm PT/20:00 UTC)

www.anandtech.com

Blackwell B100 launching is happening now..

 

[Saylick]

Looks like the rumors were true: B100 is an MCM (not quite chiplet imo) comprised of two reticle limit TSMC N4P dies connected with a high-bandwidth interconnect, likely NVLink 5 C2C. 8 stacks of HBM3E memory total, or 4 stacks per die. It's an engineering marvel and an absolute beast but I'm kind of numb to all this AI stuff tbh, especially since it's the new buzzword so it's impossible not to get reminded of it on a daily basis. This stuff doesn't impact me directly so while it is impressive, it's largely forgettable for me, the average consumer casual gamer.

 

[poke01]

https://twitter.com/x/status/1769866533818273966

Looks like gaming blackwell will get new SMs.

 

[jpiniero]

Looks like the rumors were true: B100 is an MCM (not quite chiplet imo) comprised of two reticle limit TSMC N4P dies connected with a high-bandwidth interconnect, likely NVLink 5 C2C

I think the quoted numbers include both dies AND it looks like they gutted fp64 performance.

 

[tamz_msc]

they gutted fp64 performance.

CPUs with AVX-512 and TBs worth of memory are faster at FP64 in practical terms anyway.

 

[Hitman928]

The NVIDIA GTC 2024 Keynote Live Blog (Starts at 1:00pm PT/20:00 UTC)

www.anandtech.com

Blackwell B100 launching is happening now..

Announcement, not a launch.

 

[SmokSmog]

Predictions?

GB202 4NP (N4P) 800mm2~ mono
192SM
128MB L2
512bit 28Gb/s 1792GB/s
3ghz real in game boost.

 

[igor_kavinski]

Predictions?

It will be hot. HOT, I tell you. In the very literal sense. Prepare to be toasted in your socks!

 

[SteinFG]

I think the 5090 is going to be the deepest cut GB202

There's no reason for it to be really cut down. Unless "RTX 5000 Ada Generation" (what a stupid name) is selling out, and I don't think it is.

 

[Aapje]

Very interesting and unexpected to see that Server Blackwell sticks with N4, although N4P is probably a bit of a step up in performance compared to the N4 used in Ada.

I guess they chose this because N4P outperforms N3E in speed, at the expense of efficiency.

I still see a possibility that they chose N3E for Consumer Blackwell, although with the already pretty good efficiency of Ada, and the even higher prices for N3, I can see them using N4P for consumer as well.