NVIDIA RTX 50 “Blackwell” GB202 GPU GDDR7 384-bit memory

Comixbooks

Comixbooks

Fully [H]
Joined
Jun 7, 2008
Messages
22,027
0243d82f8a9b446d8084e5598072639a.jpg



NVIDIA RTX 5000 GPUs could have GDDR7, 384-bit bus, and 2x performance compared to Lovelace
More rumors on NVIDIA's next-gen GPUs and how they could be a beefy upgrade over Lovelace - but the idea of a 512-bit bus has been dismissed.

https://www.tweaktown.com/news/9437...x-performance-compared-to-lovelace/index.html



Big Leak on Nvidia's next generation cards called Blackwell using GDDR7 Memory and GB202 GPU rumored 2024-2025

https://videocardz.com/newz/nvidia-...gpu-rumors-point-towards-gddr7-384-bit-memory
 
Last edited:
I imagine one or the other would make sense, both would have been quite the generational jump.

Going from 1000 to 1500 would be a similar jump than the 2080ti->3090 giant memory bandwith jump
 
Zinn said:
How is it even possible to make a faster card than the RTX 4090?
Click to expand...
I'm sure someone said the same thing about the RTX 3090, and then 2080 Ti before it, etc.

The Blackwell project was probably started well before the 4000 series launched, maybe even before the 3000 series launched. Nvidia tends to have many generational projects being worked on in parallel.
 
Zinn said:
How is it even possible to make a faster card than the RTX 4090?
Click to expand...
The 4090 is more than twice as fast than the 3090 in some scenario while being a very similar architecture without pushing the die a lot and having pretty much the same memory bandwith than the 3090, after such a generational gain why think it would have been where GPU improvement stop ?
 
Zinn said:
How is it even possible to make a faster card than the RTX 4090?
Click to expand...
Three ways, probably all at the same time:

1) Better design. Companies continue to improve their technology, they learn from what they did in the past, try new things, discover new ideas, etc, etc. It isn't easy, it isn't fast, but companies do improve their designs as time goes on.

2) Smaller process. TSMC has a new 3nm node out, right now only Apple is using it because it is expensive and has a high failure rate but more will or will use another high-end process like it. While the size number are BS and 3nm isn't as much smaller than 5nm as the numbers would indicate, it is still smaller, meaning more transistors in a given area, meaning more power.

3) MOAR POWER! One of the ways nVidia got gains with the 4090 was pushing its size and power consumption up to 450 watts. It wouldn't surprise me if they pushed it again with the 5090 to 600 watts.

Add all that up, chips goes fast.
 
Kind of knew the 512 bit memory bus rumour was a bit far fetched and using such a large memory bus would just increase the cost and would not have much of a huge benefit. Though still stuck on 24GB of VRAM is rather disappointing, would have been happier to see at least 32GB but that isn't achievable on a 384 bit memory bus and having 48GB of VRAM is just overkill and will also unnecessarily increase the cost of production.

But yet again, this isn't actually confirmed and who knows what the final specs and product would look like.
 
Absalom said:
I'm sure someone said the same thing about the RTX 3090, and then 2080 Ti before it, etc.

The Blackwell project was probably started well before the 4000 series launched, maybe even before the 3000 series launched. Nvidia tends to have many generational projects being worked on in parallel.
Click to expand...
Earliest rumor I recall reading that had the name Blackwell was end of 2021.
Krenum said:
Nvidia already has faster cards than even "Blackwell", they just don't release them every year. Its a marketing process.
Click to expand...
No, it's called research & development. Do you really believe that they're just sitting on product and drip feeding it to the masses?
 
Armenius said:
No, it's called research & development. Do you really believe that they're just sitting on product and drip feeding it to the masses?
Click to expand...
Like all company do since they got access to the Aliens tech in the 50s. They had ways to make a much cheaper card and have much better profit, but they decided to make a giant 600mm one and loose money for absolutely no reason...

Sycraft said:
MOAR POWER! One of the ways nVidia got gains with the 4090 was pushing its size and power consumption up to 450 watts.
Click to expand...
In a lot of game 4090 power pretty much like a 3090 I think:

power-gaming.png


Will have to see GDDR7 vs 6x and a smaller N3 vs 600mm of N5, but chance are that like for the 4090 what you push past 350-400 does not deliver a lot of performance.

https://www.anandtech.com/show/18759/cadence-derlivers-tech-details-on-gddr7-36gbps-pam3-encoding
Rumors are In general, GDDR7 promises higher performance than GDDR6 as well as lower power consumption and implementation costs than GDDR6X.


I imagine we could have said the same last time and better peak management could temp them to go there, but at some point on 15A-120v market, how much more than the 4090 can they go in that regard (without starting fires)...
 
Last edited:
LukeTbk said:
Like all company do since they got access to the Aliens tech in the 50s. They had ways to make a much cheaper card and have much better profit, but they decided to make a giant 600mm one and loose money for absolutely no reason...


In a lot of game 4090 power pretty much like a 3090 I think:

View attachment 613900

Will have to see GDDR7 vs 6x and a smaller N3 vs 600mm of N5, but chance are that like for the 4090 what you push past 350-400 does not deliver a lot of performance.

https://www.anandtech.com/show/18759/cadence-derlivers-tech-details-on-gddr7-36gbps-pam3-encoding
Rumors are In general, GDDR7 promises higher performance than GDDR6 as well as lower power consumption and implementation costs than GDDR6X.


I imagine we could have said the same last time and better peak management could temp them to go there, but at some point on 15A-120v market, how much more than the 4090 can they go in that regard (without starting fires)...
Click to expand...
Yeah, my 4090 uses less power than my 3090 ever did. My PC is drawing around 700W from the wall, so it's using between 600-700W of power. That leaves a lot of room even if you're still on 15A breakers in the US.
 
Armenius said:
Yeah, my 4090 uses less power than my 3090 ever did. My PC is drawing around 700W from the wall, so it's using between 600-700W of power. That leaves a lot of room even if you're still on 15A breakers in the US.
Click to expand...
Plus my 4090 does not power fluctuate like my 3090 does... even overclocked and max power limit (520W on my card), it still uses less power than my 3090 did with the same limit because of spikes on the 3090. Also, aside from a handful of games, even at 4K 144Hz, the card doesn't sweat and rarely uses full power... lol

So yeah, RTX 5090, take my money please nvidia! Will happily buy this card if it is every bit of a leap the 4090 was for 4K gaming.
 
Zinn said:
How is it even possible to make a faster card than the RTX 4090?
Click to expand...
Supposedly Blackwell is going to have some big changes to the architecture. Lovelace on the other hand is mostly the same architecture as Ampere with increased L2 cache and some other minor differences. And of course the biggest difference between Ampere and Lovelace being the move from Samsung 8nm (remember this is a refined 10nm process) to TSMC 4nm (a refined 5nm process), there was performance to be had.

So for Blackwell, combine the architectural improvements/changes with the move to N3 and much faster GDDR7, not hard to figure this out.

Definitely eyeing the 5090 to be my next card for 4k.
 
Sir Beregond said:
Supposedly Blackwell is going to have some big changes to the architecture. Lovelace on the other hand is mostly the same architecture as Ampere with increased L2 cache and some other minor differences. And of course the biggest difference between Ampere and Lovelace being the move from Samsung 8nm (remember this is a refined 10nm process) to TSMC 4nm (a refined 5nm process), there was performance to be had.

So for Blackwell, combine the architectural improvements/changes with the move to N3 and much faster GDDR7, not hard to figure this out.

Definitely eyeing the 5090 to be my next card for 4k.
Click to expand...
A potential Pascal revival. Awesome. I'm ready.
 
Gulkor said:
my guess end of 2024 to start of 2025.
Click to expand...

Hoping later than earlier, kind of want to give my wallet a rest after this recent 4090 purchase... if it's going to be another 2x performance increase, though, it'll be hard to resist. Could always use more cards for Stable Diffusion anyway.
 
Today, Moore's Law is Dead says Blackwell laptops are basically confirmed for January 2025. Desktop parts basically always precede laptop parts. So, Blackwell desktop could very well launch in 3rd or 4th quarter 2024.
 
In the probabilistic framework of forthcoming technological developments, Nvidia is projected to unveil the 5090 during the fourth quarter of 2024. Subsequently, the 5080 and its subordinate counterparts are envisaged to emerge in the initial quarter of 2025. This conjecture aligns with Nvidia's discernible predilection for targeted advancements in the domain of enthusiast-centric products.
 
Krenum said:
Yes absolutely.
Click to expand...
Why did they decide to lose money releasing a 600mm 4090 with a new power delivery tech, etc... instead of the superior architecture able to do the same performance with a smaller 400mm die under 350w on the same node do you think ?
 
Digital Viper-X- said:
Every time a new card comes out, it is 2x the performance... or so they say
Click to expand...


We can imagine they mean this kind of situation where they will again like last time aim to at least roughly double

performance-pt-3840-2160.png
performance-pt-3840-2160.png


If a 5090 is able to play Alan Wake 2 at 60-65 fps 4k, Phantom Liberty at 40-45 fps, they will be able to say doubled performance and in Ray reconstruction being used they will be able to say more than double.

It is a possibility that there will not have on hands the nice frame generated bar this time to use (they could only compare to Ampere again or add 1 or 2 generated frame, but maybe not)
 
GoldenTiger said:
I'm referring to the speed, not the capacity ;). Gigabits per second (gbps).
Click to expand...
That would be sweet.

As for capacity, do we know what density GDDR7 will have? Theoretically could do 36GB if each of 12 chips on 384-bit is 3GB. If not and it's just the same 2GB per chip, then I guess another 24GB card. Or maybe 48GB if front and back like the 3090, though front and back seems less than ideal to me.
 
Sir Beregond said:
That would be sweet.

As for capacity, do we know what density GDDR7 will have? Theoretically could do 36GB if each of 12 chips on 384-bit is 3GB. If not and it's just the same 2GB per chip, then I guess another 24GB card. Or maybe 48GB if front and back like the 3090, though front and back seems less than ideal to me.
Click to expand...
Back in June Samsung revealed 16Gb chips, so we have yet to move beyond GDDR6 in that respect. It's still manufactured on a 10nm node similar to GDDR6.

https://news.samsung.com/global/sam...k-the-next-generation-of-graphics-performance
 
celwin said:
Nvidia and AMD has products lined up for multiple generations. They purposely slow roll incremental improvements in each gen release to keep the profits comming in.
Click to expand...
Would it not be much more profitable to release that superior smaller die that give the same result because of a better architecture (if both has it, keeping it secret to the other would not matter much) ?
 
LukeTbk said:
Would it not be much more profitable to release that superior smaller die that give the same result because of a better architecture (if both has it, keeping it secret to the other would not matter much) ?
Click to expand...
You would think. It's not just GPUs. Major electronic manufacturers follow the same pattern. Sort of makes sense. You want your R&D waay ahead of your current and next product release.
 
celwin said:
You would think. It's not just GPUs. Major electronic manufacturers follow the same pattern. Sort of makes sense. You want your R&D waay ahead of your current and next product release.
Click to expand...
Does it, if say Nvidia has a working Blackwell substitute that is significantly superior (on the same node, that part is not on them obviously) that they could have made a 4060 performance card on a 112 mm die instead of 159mm (and so on up the stack), that make possible less power and overall cheaper pci, cooler, power delivery system, etc... and make much more profit doing so.

What would be the reason not to do it
1) Be certain that the next gen will be better than this one, ok, but that would still be true, worst case scenario they just release a bigger die that use more power to do it, they would have giant room to do so and they will have a better node and memory next gen to help as well.

2) Hide it to the competition, they could reverse engineering it, sure, but then the statement say that they all do it, the competition already have much more advanced graphic

That does not feel right to me at all, also the miss fired launch and failure (AMD next top of the line cancelled, 5090 failure vs 9700pro), why Intel is so much behind what the others have now, if themselves are 2 generation behind, Intel for sure did leave nothing in the bank.

Big clients like Sony-Microsoft for gaming or facebook-amazon-google and other for datacenter-ml and other would not go for it, they homemade solution would completely walk over amd-nvidia pro line products (and intel-amd cpus), they themselve are their own user, ever compute by watt is pure money in their pocket and have zero reason to not use the very best they can make right now, same goes for Tesla chips.

Or take Apple Arm shift, does it feel like when the M1 launched they already had a much better M2 and M3 in the bank, to be certain it will be 3 nice upgrade in a row or that the M1 was the best product they could make and they tried to make it better after with sometime so-so result ?

What is the reasoning, proof or things that point toward this ?
 
You must log in or register to reply here.
Back
Top