• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Blackwell gpus

I don't think we'll get a massive upgrade over the 4090 this time, what is pushing Nvidia to do that and even if they can with these interlinked die's there is no way in hell they are selling that to consumers at an already phat margin. They will want to sell that stuff to the datacentre & AI boyZ who charge a bomb for their services....

Its just my opinion of course I don't pretend to know anything unlike certain YT channels :rolleyes:
 
I remember saying that to myself over a hundred times about the 2080ti, 3090 & 4090 gpu's after seeing the stupid prices

And i still ended up buying a 2080ti & now a 4090 :o:o:o

I'm not in the same financial situation I was when the 4090 came out though as now I'm planning for the next few years, Home, Kids etc... so £2000 for something to draw pictures on a screen is a no go.
 
I'm not in the same financial situation I was when the 4090 came out though as now I'm planning for the next few years, Home, Kids etc... so £2000 for something to draw pictures on a screen is a no go.
Especially when you can have just as good of a time playing at 2k rather than 4k for one quarter of the cost. And I'm saying that as someone who has both types of displays.
 
At GDC Jensen said the two dies talk to each other like there is only one. Probably MCM design becomes doable.


As far as we know based on public information, unless a microchip architect or engineer can interject and correct us, using MCM to make a gaming GPU with two full dies on the same silicon substrate requires around 100TB/s of connection bandwidth between the two dies (50TB/s each way) and right now Nvidia's GB Blackwell GPU only has a 10TB/s connection (5TB/s each way). 10TB/s is enough for generative AI, inference and computational workloads because the workloads are not sensitive to latency and don't need to output full image frames within small frametimes, but gaming is sensitive to this and so requires much more bandwidth to work without issues.

And while I do not understand the exact technical reasons for why its hard to scale up the connection bandwidth between the dies; the dumbed down consumer friendly version is that there is lots of little wires that have to run between the two dies to transport the data and if you want to increase bandwidth you have to lay more wires between the two dies and its very very hard fitting all the wires into that tiny space.
 
Last edited:
As far as we know based on public information, unless a microchip architect or engineer can interject and correct us, using MCM to make a gaming GPU with two full dies on the same silicon substrate requires around 100TB/s of connection bandwidth between the two dies (50TB/s each way) and right now Nvidia's GB Blackwell GPU only has a 10TB/s connection (5TB/s each way). 10TB/s is enough for generative AI, inference and computational workloads because the workloads are not sensitive to latency and don't need to output full image frames within small frametimes, but gaming is sensitive to this and so requires much more bandwidth to work without issues.

And while I do not understand the exact technical reasons for why its hard to scale up the connection bandwidth between the dies; the dumbed down consumer friendly version is that there is lots of little wires that have to run between the two dies to transport the data and if you want to increase bandwidth you have to lay more wires between the two dies and its very very hard fitting all the wires into that tiny space.
I was thinking that, too. But who knows, maybe they've done some more magic to go around the limitations. We'll see.
 
  • Like
Reactions: TNA
GN's situational take on NV:


"ah yes my favorite programming languages: Java, Python, and 4chan" :cry:


Get a load of this guy

He has a Radeon 7 in the background, bottle of Pepsi on the right and a family photos of 3 monkeys on the left wall

 
Last edited:
Yeah so GN didn't really make any predictions it was more commentary and thoughts where as NAAF rightly points out a 5090 'like' gaming GPU will be on the same process node and therefore based on historical data can expect up to a 30% improvement unless they radically up the die size possibly using multiple cores. However going back to my earlier point are Nvidia going to sell a mGPU to gaming consumers whilst passing on the savings once you've deducted the interlinks?

My answer to that is simply 'not unless they feel they have to' and guess what AMD are feasibly absent in the top tier again.

I think its far more likely Nvidia will try to sell Blackwell on the basis of path tracing and AI features than raw performance, however what they might do is another 3080/3090 type cycle where the 5080 is the better option getting you that 4090+ performance for much less where as the 5090 is the whale chum i.e. a bit more perf for a lot more $$
 
GN's situational take on NV:


"ah yes my favorite programming languages: Java, Python, and 4chan" :cry:

A rare moment of realization from a tech jurno.

If you're up against a company with vastly more resources to put in to their GPU eco system then you as the underdog will find it extremely difficult to keep up, a monopoly is not a good thing for competition.

If they are now starting to realize this then its too late, the damage is done and baked in.

 
Last edited:
Back
Top Bottom