Blackwell gpus

[Tinek]

Source? First I've heard about this.

Example, from August (just 1st result in search): https://www.tomshardware.com/pc-com...ll-gpus-allegedly-delayed-due-to-design-flaws Sure it's not specifically gaming GPUs, but they're all related. Considering they need about a month to fix and test things, then restart production, which then take min. 3 months to get first chips from the factory, confirm all good, then produce enough for release - yeah, January seems to be about the earliest time they could do a release, though I expect stock numbers to be relatively small initially. Then again, some other sources issue wasn't in the chip design itself but in packaging by TSMC, which would likely mean they had a whole batch of chips damaged in production most likely and not usable (depending on the exact issue), then had to wait for new batch to be delivered - that would introduce delays to resolve it, but release stock could be bigger than with fully restarting whole production.

 

[Howling]

they will have comparable yields on both products, because both chips will be cut down accordingly to match the same level of defect incidence, infact they will be pushing borders on 5080 compared to the 5090 because of the available headroom cannibalization seems to be debatable, because in an upsell situation a 5090 isnt cannibalizing the 5080, it should be the other way around
also it isnt like the situation where ps4 pro users upgrade to ps5 and then to ps5 pro midcycle, 4090 will not be upgraded to 5080 its just not viable
thats broadly my thought process

I'm not saying you're wrong here; you may well be right so I'll just explain how I thought it worked so we can see where I've made the error (I'm going to simplify it for demonstration sake).

So in chip manufacture they create a huge wafer with say 10,000 'cores'.
Each core has say a 1% failure rate.
You're cutting the wafer to 2 models - one with 100 cores and one with 200 cores.
On average you'd expect 1% of those 100 core models to have failures, and 2% of those models with 200 cores -- double the failure rate.

I assume they try and account for this by cutting to say 105 cores for the 100 core units and 210 cores for the 200 core units to minimize chip fails? Is this fully negating the effect of the larger die size failure rate?

 

[brostradamus]

I'm not saying you're wrong here; you may well be right so I'll just explain how I thought it worked so we can see where I've made the error (I'm going to simplify it for demonstration sake).

So in chip manufacture they create a huge wafer with say 10,000 'cores'.
Each core has say a 1% failure rate.
You're cutting the wafer to 2 models - one with 100 cores and one with 200 cores.
On average you'd expect 1% of those 100 core models to have failures, and 2% of those models with 200 cores -- double the failure rate.

I assume they try and account for this by cutting to say 105 cores for the 100 core units and 210 cores for the 200 core units to minimize chip fails? Is this fully negating the effect of the larger die size failure rate?

i remember this from my or class..
what they have as a parameter is average number of defects/area (d) as a parameter, then to estimate defect probability you assume that the defects are distributed as a poisson distribution with parameter d*A (chip area)
so the defect probability is 1-exp(dA).. it doesnt scale linearly with area, so they balance it by cutting a much larger area from a larger chip, so in the end they just cut enough area from both chips to deliver the same level of reliability, so as to ensure that the scale of the chip does not put a constraint on supply - without this there would be a much bigger hit on profitability and lost revenue opportunity (its an exponential response after all)
also, purely from a business strategy standpoint, they definitely need 4090 users to upgrade, and sooner the better

 

[Howling]

i remember this from my or class..
what they have as a parameter is average number of defects/area (d) as a parameter, then to estimate defect probability you assume that the defects are distributed as a poisson distribution with parameter d*A (chip area)
so the defect probability is 1-exp(dA).. it doesnt scale linearly with area, so they balance it by cutting a much larger area from a larger chip, so in the end they just cut enough area from both chips to deliver the same level of reliability, so as to ensure that the scale of the chip does not put a constraint on supply - without this there would be a much bigger hit on profitability and lost revenue opportunity (its an exponential response after all)
also, purely from a business strategy standpoint, they definitely need 4090 users to upgrade, and sooner the better

You could have just said "Yep Howl, they do that"
Man, I haven't even heard poisson distribution for at least 20 years though... all that yummy maths statistics we had to (well, technically I chose maths so it's my bad... preferred Pure maths and Mechanics though) learn and totally use every day.... xD

 

[opethdisciple]

Hint at prices here.

 

[brostradamus]

You could have just said "Yep Howl, they do that"
Man, I haven't even heard poisson distribution for at least 20 years though... all that yummy maths statistics we had to (well, technically I chose maths so it's my bad... preferred Pure maths and Mechanics though) learn and totally use every day.... xD

i too have a hazy recollection there.. d is < 0 so, the function is increasing at a decreasing rate, lol so the response is better than linear..

 

[brostradamus]

Hint at prices here.

jensen's going to make it look much better during the launch event and everyone here will be talking about how 5090 is great value for the price

 

[Calin Banc]

Pretty much all of them. Cyberpunk with PT, IJ, etc. Hub shown all these examples very well in their video too so I can refer you there now

Well yes, there are some issues, but not that big. Just RT in CB77 is pretty spotless (almost), Chernobilyte too.
It kinda depends per scene, with some tricks you fix a lot of issues. Some are just "bugs" that require more polish.

All in all, to me is worth it. The downside is less than those from "pure" raster.

 

[Tinek]

Well yes, there are some issues, but not that big.

For many it might not be game breaking. For me it's mostly very annoying visually and breaks immersion much more than just raster rendering, sans the stupid SSR. I much more preferred cube mapping and other older methods.

Just RT in CB77 is pretty spotless (almost), Chernobilyte too.

I like GI in the night with neon lights and wet reflections in the rain. Mostly because of aesthetics of the cyberpunk city. But, they look fine atheistically even without whole PT involved.

It kinda depends per scene, with some tricks you fix a lot of issues. Some are just "bugs" that require more polish.

Current denoising algos aren't good enough - too much blur or too much noise, depending on how aggressive it is. Stuff they show in scientific papers look much better but it all has performance hit again. I'd have to see how much it hits FPS in practice before judging it. It's still the thing that you need all these AI crutches to mask the fact GPUs are orders of magnitude too slow for PT, still.

All in all, to me is worth it. The downside is less than those from "pure" raster.

I never advocate for "pure" raster but to much smarter approach by devs to each game and use what works without affecting quality or FPS too much. But currently it's 1 switch on and done, 0 effort to actually polish things. All to have it done cheap and to slap RT/PT on marketing slides. Really lazy approach! Meanwhile prices of these games go up and not down.

 

[Sargatanas2511]

jensen's going to make it look much better during the launch event and everyone here will be talking about how 5090 is great value for the price

Buy 10, you'll save more.

 

[Dangerman1337]

To me that would aire on the 5080 been on the cheaper side. Otherwise all reviewers would be like "if your spending this much, just wait a few weeks".

On the otherside if this is all about the Trump tariffs, wouldn't they release the strongest card first?

It's weird but maybe Nvidia wants to soften the "blow" from Blackwell being too expensive, especially if the 5080 beats the 4090 in most if not all cases? Especially with Path-Tracing?

On the subject of tariffs... I hope a lot of Hardware Companies go "if Americans have to pay more, then so does the rest of the world" nonsense.