News The RTX 5090's GB202 GPU will reportedly be the largest desktop chip from Nvidia since 2018 coming in at 744mm-squared — 22% larger than AD102 on t...

[Admin]

Rumor has it that Nvidia's RTX 5090 wielding the GB202 chip has a massive 744mm-squared die size. Nvidia is expected to announced the Blackwell RTX 50-series GPUs at CES 2025 in January.

The RTX 5090's GB202 GPU will reportedly be the largest desktop chip from Nvidia since 2018 coming in at 744mm-squared — 22% larger than AD102 on t... : Read more

 

[bit_user]

If you look back through the history of Nvidia GPUs, their die sizes always got bigger when they released another generation on the same manufacturing node. That's both because nodes tend to get cheaper over time and because it's really hard to offer substantially more performance on the same node without using more die area. Sure, they can tweak with the microarchitecture, but that's not going to deliver the kind of generational improvements people have come to expect of GPUs.

That said, I expect Nvidia isn't going to do us any favors on pricing. If the launch price of the GB102 isn't 22% higher than the AD102's, I bet it'll at least be close! I also suspect peak power could be as much as 22% higher, as well. Especially when you consider it's got 33% more channels of GDDR6X and it's running at a higher data rate.

 

[acadia11]

Ok, since I’m doing the rational thing and not splurging on a B200 for my AI needs … fine I’ll be practical and settle for a Rtx 5090, …

 

[Mama Changa]

This is close to the theoretical maximum die size of ~858mm^2 on low NA EUV. This is why they have no choice but to move to chiplets eventually, as high 0.55 NA EUV halves that maximum size, so no giant power hungry monolithic chips for 2nm. I'm not sure if Rubin is on N3P, but it seems likely if Nvidia's timeframe for release is on shcedule, as I doubt 2nm will be ready in time.

 

[dimar]

I'm super hyped about independent benchmarks of RTX 5090 along with Ryzen 9 9950X3D and whatever Intel's top Ultra CPU will be. I'm prepared to be disappointed as well 🙂 Also hoping AMD to make something competitive to 5090

 

[jlake3]

I'm super hyped about independent benchmarks of RTX 5090 along with Ryzen 9 9950X3D and whatever Intel's top Ultra CPU will be. I'm prepared to be disappointed as well Also hoping AMD to make something competitive to 5090

Got some bad news for ya on an AMD equivalent to the 5090.

Rumours have been that AMD’s top RDNA4 die ran into development issues that were gonna cost a lot of money/engineering hours to fix, and since their datacenter/AI plans don’t utilise RDNA (they have compute-optimised CDNA dies for that) and gaming GPUs at that price are a low sales volume, they decided just to cancel it and that’s how “focus on the midrange” became the plan for RX 8000.

 

[Giroro]

Got some bad news for ya on an AMD equivalent to the 5090.

Rumours have been that AMD’s top RDNA4 die ran into development issues that were gonna cost a lot of money/engineering hours to fix, and since their datacenter/AI plans don’t utilise RDNA (they have compute-optimised CDNA dies for that) and gaming GPUs at that price are a low sales volume, they decided just to cancel it and that’s how “focus on the midrange” became the plan for RX 8000.

On the other hand, AMD is copying Nvidia's strategy of missing their planned GPU launch window by 3-6 months. They couldn't have just launched this year, because AMD has no idea what to sell their cards for until Nvidia fixes the prices.
They gotta do something with their engineers with all that extra time on their hands.
It's not like AMD needed that time to let manufacturing catch up harvesting their infinite AI money pit, like Nvidia.

 

[Giroro]

If the RTX 5090 has a smaller die than the RTX 2080 Ti, than the 5090 will be cheaper than the 2080 Ti, obviously.

It's not like the mark ups on these GPUs are so insanely high over the base cost of the die that a 20% increase is die cost would be a barely noticeable blip compared to the price is the product, or anything.

I mean, they'll still probably raise the price of a 5090 by a lot, Because it's going to be for AI, not gaming.
Or they'll deliberately gimp AI performance in their gaming cards, which is still on the table... But then why would Nvidia waste the silicon on gamers at all?
Nvidia is not a gaming company anymore. We're just going to have to deal with it.

 

[jp7189]

If the RTX 5090 has a smaller die than the RTX 2080 Ti, than the 5090 will be cheaper than the 2080 Ti, obviously.

It's not like the mark ups on these GPUs are so insanely high over the base cost of the die that a 20% increase is die cost would be a barely noticeable blip compared to the price is the product, or anything.

I mean, they'll still probably raise the price of a 5090 by a lot, Because it's going to be for AI, not gaming.
Or they'll deliberately gimp AI performance in their gaming cards, which is still on the table... But then why would Nvidia waste the silicon on gamers at all?
Nvidia is not a gaming company anymore. We're just going to have to deal with it.

Nvidia's CUDA dominates specifically because it works even on entry level cards. Cost conscious college kids get to play with it, and then after a few years become influential with company money. Nvidia won't do anything to jeopardize that. They'll use VRAM sizes to entice people and companies to go with higher tiers. RTX 6000ada is pretty much the same as a 4090 but with double the VRAM and 4x the price.

I could see a world where they release a 24GB 5090 with 2GB chips and later a ti or titan with 36GB, 3GB chips. (I don't believe the 512 bit memory bus rumors.)

Likewise a RTX 6000 black might have 48GB to start and a 72GB option down the road.

 

[bit_user]

RTX 6000ada is pretty much the same as a 4090 but with double the VRAM and 4x the price.

They also back off the clock speed and each card comes with a substantially longer warranty that's rated for 24/7 compute workloads. Some of their professional (formerly Quadro) cards also include ECC memory, such as that one. In terms of power, it's rated for only 300 W.

At my job, we once used Quadro P1000 cards for a certain project. It bugged me that we didn't use GTX 1050 Ti, which had higher clocks and ran up to 75 W instead of like 51 W, while costing only half as much. The answer I got back is that we had to use the Quadro version for its longer warranty, as well as that our workload potentially voided the warranty and CUDA ToS of the GTX product.

 

[jp7189]

They also back off the clock speed and each card comes with a substantially longer warranty that's rated for 24/7 compute workloads. Some of their professional (formerly Quadro) cards also include ECC memory, such as that one. In terms of power, it's rated for only 300 W.

At my job, we once used Quadro P1000 cards for a certain project. It bugged me that we didn't use GTX 1050 Ti, which had higher clocks and ran up to 75 W instead of like 51 W, while costing only half as much. The answer I got back is that we had to use the Quadro version for its longer warranty, as well as that our workload potentially voided the warranty and CUDA ToS of the GTX product.

Can't comment on the tos.. never dug in to it. Regarding warranty on specific products I've used, the MSI suprim 4090 is 3 years same as the pny 6000ada. The former has an option to enable ECC in the nvidia control panel though I'm not sure how to verify it actually works. The later also uses much slower non-X vram, but that's still way faster than swapping to system ram if you need more than the 24GB on the 4090.

 

[jlake3]

Can't comment on the tos.. never dug in to it. Regarding warranty on specific products I've used, the MSI suprim 4090 is 3 years same as the pny 6000ada. The former has an option to enable ECC in the nvidia control panel though I'm not sure how to verify it actually works. The later also uses much slower non-X vram, but that's still way faster than swapping to system ram if you need more than the 24GB on the 4090.

Never made a claim with MSI, but I’ve had to make warranty claims with both PNY (who is the partner for Quadro cards in North America) as well as Asus, and the number of years of warranty on paper does not tell the full warranty experience story.

PNY I was doing something I probably shouldn’t have been, I shipped them the card, and no questions they express shipped a new retail box to me.

ASUS my hardware just failed under normal usage, and they accused me of user damage, then when I pointed out how that didn’t add up they attempted a repair, which didn’t work, then they claimed my configuration was the issue. I then swapped out every other major component one at a time to prove wasn’t the case, then they attempted to repair the hardware again, which still didn’t work… then said I could send it in again (postage at my own expense) and they’d attempt another repair but they could not authorize a full replacement. When I pointed out how many months in we were and how they’d failed two repairs, they pointed at the legal text of the warranty and said it give them the choice of repair or replace, and their choice is to repair again and if you don’t like it, you can buy a replacement item from our eShop.

 

[bit_user]

Regarding warranty on specific products I've used, the MSI suprim 4090 is 3 years same as the pny 6000ada.

You might find that the warranty on their consumer products explicitly excludes 24/7 workloads, like crypto mining and AI training. Crypto can kill a card in less than 3 years, quite easily. I'm sure they started adding such exclusions, back when that took off. I don't know how they'd be be able to tell, but I wouldn't put it past them to have a way.

The former has an option to enable ECC in the nvidia control panel though I'm not sure how to verify it actually works.

You mean the RTX 4090 does? That's interesting. I wonder if it uses something like the in-band ECC method Intel supports in some of its embedded CPUs.

 

[bit_user]

ASUS ...

Wow, thanks for sharing! If I buy another ASUS product, I'll probably make sure to test it well within the 30 return period of the store, then. Hopefully, it'll fail early, if at all, so that I can take advantage of that.

I don't buy a lot of ASUS products, but I did get one of their workstation motherboards and a gaming monitors, both in the past year. Both are fine, so far. Actually, I'm disappointed with the contrast ratio of their monitor, especially for a display that's rated HDR600, but that's another story for another thread.

 

[spongiemaster]

Got some bad news for ya on an AMD equivalent to the 5090.

Rumours have been that AMD’s top RDNA4 die ran into development issues that were gonna cost a lot of money/engineering hours to fix, and since their datacenter/AI plans don’t utilise RDNA (they have compute-optimised CDNA dies for that) and gaming GPUs at that price are a low sales volume, they decided just to cancel it and that’s how “focus on the midrange” became the plan for RX 8000.

AMD is unlikely to admit it any time soon, but it's well established they tried to go multi chiplet for highend RDNA 4 and it didn't work. The technology doesn't exist currently to reduce the inter-chiplet latency enough to make it work. It boggles the mind they ever thought they could get it working when they haven't gotten multichiplet designs working properly in gaming for CPU's going back to the original Threadripper released in 2017.

 

[jp7189]

You might find that the warranty on their consumer products explicitly excludes 24/7 workloads, like crypto mining and AI training. Crypto can kill a card in less than 3 years, quite easily. I'm sure they started adding such exclusions, back when that took off. I don't know how they'd be be able to tell, but I wouldn't put it past them to have a way.


You mean the RTX 4090 does? That's interesting. I wonder if it uses something like the in-band ECC method Intel supports in some of its embedded CPUs.

Warranty claims aside, the msi card can run 24/7 and gpu and memory temps hang out 50-60C while sucking down 400+ watts. With a nice relaxed air flow. Of course you can't stick a lot of water cooled cards in one chassis.

The pny 6000ada caps hard at 300 watts, but the gpu temps hang out in the 80s with mem around 100C. Though I've never seen them thermally throttle, perf limit is always power. Those blower cards blast air out the back and are a single point of failure for that design. Fun fact, they will heat up the metal doors on the back of a server cabinet to the point where you have to be careful where you touch them.

Either way I havent see any gpu component failures on any card, but I dont have 3 years 24/7 on any card, and my sample size isn't statistically significant anyway. I'll just say the pny cards seem more at the limit despite having lower power usage.

 

[bit_user]

Warranty claims aside, the msi card can run 24/7 and gpu and memory temps hang out 50-60C while sucking down 400+ watts. With a nice relaxed air flow. Of course you can't stick a lot of water cooled cards in one chassis.

The pny 6000ada caps hard at 300 watts, but the gpu temps hang out in the 80s with mem around 100C. Though I've never seen them thermally throttle, perf limit is always power. Those blower cards blast air out the back and are a single point of failure for that design. Fun fact, they will heat up the metal doors on the back of a server cabinet to the point where you have to be careful where you touch them.

Either way I havent see any gpu component failures on any card, but I dont have 3 years 24/7 on any card, and my sample size isn't statistically significant anyway. I'll just say the pny cards seem more at the limit despite having lower power usage.

I'll bet you'd find the PNY cards run lower voltages and voltage tends to be a bigger issue for silicon longevity than temperatures.

 

[jp7189]

I'll bet you'd find the PNY cards run lower voltages and voltage tends to be a bigger issue for silicon longevity than temperatures.

Very true re: voltage vs longevity.

I just ran and checked and here's what I see
Ada avg 0.909 volts looks like 0.900 almost constant with some excursions to 1.060; 287watts avg; 301 peak

4090 0.910 avg with excursions to 1.050; 448watts avg, 470 peak

 

[bit_user]

I just ran and checked and here's what I see
Ada avg 0.909 volts looks like 0.900 almost constant with some excursions to 1.060; 287watts avg; 301 peak

4090 0.910 avg with excursions to 1.050; 448watts avg, 470 peak

So, it's the same workload? If so, that's a little weird.