Jump to content

Will the RTX 5090 use GB202?

Will the RTX 5090 use GB202 or end up a fully unlocked GB203?  

7 members have voted

  1. 1. Will the RTX 5090 use GB202 or end up a fully unlocked GB203?

    • GB202, fully or mostly unlocked
      4
    • GB202, absolute bottom bin
      2
    • GB203, fully or mostly unlocked
      1


From a previous thread discussing RTX 4000 series binning scheme, a reminder of how upsold the RTX 4000 series was relative to RTX 3000 series:

Why was it like this? Probably because Nvidia could. Nvidia seems to be content with having a 30-50% performance jump between generations, even if the 1:1 intergenerational comparison is closer to 70-100%. They've probably figured out precisely what people will tolerate for generational bumps in performance, not like they don't have advanced AI hardware or anything 馃槃

I don't think they'll get enough pressure from Radeon or Intel to change this course and I believe we'll see another 1-2 tier upsell on the binning scheme.

Personally, I'm pessimistic and think we won't see GB202 at all on Geforce, maybe only a mid gen refresh with a RTX 5090 ti. The optimist in me thinks Nvidia will at least make the RTX 5090 good to sell the generation.聽My suspicion could be wrong entirely because of the yields. If Nvidia ends up with a lot of bottom bin GB202 and not that many GB203's fully unlocked, they might end up just tossing them to Geforce to get a strong marketing pitch in the RTX 5090.

Current leak regarding the specs:

image.png.53a960f853bb6c6d5216ac83156cbfeb.png

Is anyone aware of GDDR7 capacity? Is there a higher amount of DRAM per chip? Nvidia practically needs to have higher VRAM than their scheme for RTX 4000 series, otherwise they'd fall behind on that alone. If GDDR7 has higher capacity per chip, then they can upsell lower memory bus widths to meet that. That could even involve an RTX 5090 with the GB203 at 256-bit with 24GB, if possible. Even then, I wouldn't be surprised of the RTX 5090 ends up with less VRAM than 24GB, because if it gaps the RTX 4090 enough, it won't matter to consumers who want more performance. They could also end up quadrupling up the bus like what they did with the RTX 4060ti 16GB.

What does give more credence to the possibility that the RTX 5090 won't run off GB202 is the credible leak that it'll be a dual slot card.

Most of Nvidia's profits aren't Geforce, their quarterly earnings from 22 May 2024 showed their data center market getting 8.7x the profits compared to Gaming.

Data Center

  • First-quarter revenue was a record $22.6 billion, up 23% from the previous quarter and up 427% from a year ago.

Gaming and AI PC

  • First-quarter Gaming revenue was $2.6 billion, down 8% from the previous quarter and up 18% from a year ago.聽

NVIDIA Announces Financial Results for First Quarter Fiscal 2025 | NVIDIA Newsroom

I just don't see a reason why Nvidia would waste any GB202, even the bottom bin on Geforce if they can make an order of magnitude more money from the same die on the data center market.

Ryzen 7950x3D Direct Die NH-D15, CCD1 disabled

RTX 4090 @133%/+230/+500

Builder/Enthusiast/Overclocker since 2012聽 //聽 Professional IT since 2017

Link to comment
Share on other sites

Link to post
Share on other sites

5090 will have to beat 4090. 256-bit bus would require 3GB/chip to match the 24GB of 4090, which may not be available early on. I can't see them dropping to 16GB. It seems more likely they'll keep using 2GB/chip with a wider bus which would also help performance. Doubling up 2GB chips on 256-bus could give 32GB, and with GDDR7 speeds it might keep up but wouldn't beat 4090 in bandwidth. I don't feel that a good fit.

The claim of a dual slot cooler is more curious. A dual slot is easily good for 250W, quite a step down from the 450W of 4090. If Nvidia have found a way to make it more efficient and improve the cooling it could close the gap.

Gaming system: R7 7800X3D, Asus ROG Strix B650E-F Gaming Wifi,聽Thermalright Phantom Spirit 120 SE聽ARGB, Corsair Vengeance 2x 32GB 6000C30, RTX 4070,聽MSI MPG A850G, Fractal Design North,聽Samsung 990 Pro 2TB, Alienware AW3225QF (32" 240 Hz OLED)
Productivity system:聽i9-7980XE, Asus X299 TUF mark 2, Noctua D15, 64GB ram (mixed), RTX 3070, NZXT E850, GameMax Abyss, Samsung 980 Pro 2TB, iiyama ProLite XU2793QSU-B6 (27"聽1440p 100 Hz)
Gaming laptop: Lenovo Legion 5, 5800H, RTX 3070, Kingston DDR4 3200C22 2x16GB 2Rx8, Kingston Fury Renegade 1TB + Crucial P1 1TB SSD, 165 Hz IPS 1080p聽G-Sync Compatible

Link to comment
Share on other sites

Link to post
Share on other sites

5 minutes ago, porina said:

5090 will have to beat 4090. 256-bit bus would require 3GB/chip to match the 24GB of 4090, which may not be available early on. I can't see them dropping to 16GB. It seems more likely they'll keep using 2GB/chip with a wider bus which would also help performance. Doubling up 2GB chips on 256-bus could give 32GB, and with GDDR7 speeds it might keep up but wouldn't beat 4090 in bandwidth. I don't feel that a good fit.

The claim of a dual slot cooler is more curious. A dual slot is easily good for 250W, quite a step down from the 450W of 4090. If Nvidia have found a way to make it more efficient and improve the cooling it could close the gap.

What if they just run a 256bit bus and quadruple it to 32GB? GDDR7 might not struggle as much as GDDR6 when doing so, since we didn't see this on any GDDR6x cards. Seeing a theoretical doubling of memory bandwidth with GDDR7.

I think the double slot = smaller die, closer to the RTX 4080's bin. Gaming cards want that high clock speed, something you don't get lowering the TDP on a high core count GPU.

Ryzen 7950x3D Direct Die NH-D15, CCD1 disabled

RTX 4090 @133%/+230/+500

Builder/Enthusiast/Overclocker since 2012聽 //聽 Professional IT since 2017

Link to comment
Share on other sites

Link to post
Share on other sites

Me thinks the 5090 will be a meh bin from the gb202, a bit worse than the 4090 wrt its ad102. Don't think we'll be seeing a 5090ti either.

1 hour ago, Agall said:

Is anyone aware of GDDR7 capacity? Is there a higher amount of DRAM per chip?

16Gb at first, 24Gb to follow soon-ish. It can go up to 64Gb in theory, but I don't think this will happen.

1 hour ago, Agall said:

That could even involve an RTX 5090 with the GB203 at 256-bit with 24GB

Rumor is that it's going to be a 448-bit bus, so that would be 28GB of vram using the initial GDDR7 specs. Seems reasonable.

1 hour ago, Agall said:

I just don't see a reason why Nvidia would waste any GB202, even the bottom bin on Geforce if they can make an order of magnitude more money from the same die on the data center market.

Small margin or not, it's still money, and if they can have enough production to make it sell, then why not profit some extra couple billions?

Tbh so far the 5090 seems like a nice upgrade from a 4090. Seeing that I'm still on 2x3090, I might make the jump to two blackwells if those are indeed dual slot with 28gb each.

FX6300 @ 4.2GHz | Gigabyte GA-78LMT-USB3 R2 | Hyper 212x | 3x 8GB + 1x 4GB @ 1600MHz | Gigabyte 2060 Super | Corsair CX650M | LG 43UK6520PSA
ASUS X550LN | i5 4210u | 12GB
Lenovo N23 Yoga

Link to comment
Share on other sites

Link to post
Share on other sites

1 minute ago, igormp said:

Me thinks the 5090 will be a meh bin from the gb202, a bit worse than the 4090 wrt its ad102. Don't think we'll be seeing a 5090ti either.

16Gb at first, 24Gb to follow soon-ish. It can go up to 64Gb in theory, but I don't think this will happen.

Rumor is that it's going to be a 448-bit bus, so that would be 28GB of vram using the initial GDDR7 specs. Seems reasonable.

Small margin or not, it's still money, and if they can have enough production to make it sell, then why not profit some extra couple billions?

Tbh so far the 5090 seems like a nice upgrade from a 4090. Seeing that I'm still on 2x3090, I might make the jump to two blackwells if those are indeed dual slot with 28gb each.

That would make sense with a 448-bit bus, so a mid range bin of a GB202, similar to the RTX 4090's bin of the AD102. I think the only major reason to make the RTX 5090 actually good would be because those halo products sell the generation. People look at RX 7900 XTX vs RTX 4090 and see how massive the gap is at 4K which puts the idea in their head that any RTX 4000 series card is probably better than any RX 7000 at the same price point.

Nvidia might simply not care though, their gaming profits are generally up, but not nearly as much as data center. If they can sell the same GB202 GPU for 3x as much to data center clients, then why would they bother? Especially if TSMC fab time simply can't keep up with demand.

Ryzen 7950x3D Direct Die NH-D15, CCD1 disabled

RTX 4090 @133%/+230/+500

Builder/Enthusiast/Overclocker since 2012聽 //聽 Professional IT since 2017

Link to comment
Share on other sites

Link to post
Share on other sites

18 minutes ago, Agall said:

What if they just run a 256bit bus and quadruple it to 32GB? GDDR7 might not struggle as much as GDDR6 when doing so, since we didn't see this on any GDDR6x cards. Seeing a theoretic doubling of memory bandwidth with GDDR7.

I think the double slot = smaller die, closer to the RTX 4080's bin.

4090 runs 21 Gbps. Early GDDR7 can go up to 32 Gbps, with rumours of 50 series using slower than that. So about +52% bandwidth per chip best case. 4090's 384 bit bus is 50% wider. Same or less bandwidth overall sounds a tough one for next gen flagship unless they throw in a bucket of cache. Which is not impossible.

Bigger reason I don't feel GB203 is viable for a 5090 is it probably wont have enough cores in it. Rumours do point to it being a good fit for 5080.

Gaming system: R7 7800X3D, Asus ROG Strix B650E-F Gaming Wifi,聽Thermalright Phantom Spirit 120 SE聽ARGB, Corsair Vengeance 2x 32GB 6000C30, RTX 4070,聽MSI MPG A850G, Fractal Design North,聽Samsung 990 Pro 2TB, Alienware AW3225QF (32" 240 Hz OLED)
Productivity system:聽i9-7980XE, Asus X299 TUF mark 2, Noctua D15, 64GB ram (mixed), RTX 3070, NZXT E850, GameMax Abyss, Samsung 980 Pro 2TB, iiyama ProLite XU2793QSU-B6 (27"聽1440p 100 Hz)
Gaming laptop: Lenovo Legion 5, 5800H, RTX 3070, Kingston DDR4 3200C22 2x16GB 2Rx8, Kingston Fury Renegade 1TB + Crucial P1 1TB SSD, 165 Hz IPS 1080p聽G-Sync Compatible

Link to comment
Share on other sites

Link to post
Share on other sites

10 minutes ago, porina said:

4090 runs 21 Gbps. Early GDDR7 can go up to 32 Gbps, with rumours of 50 series using slower than that. So about +52% bandwidth per chip best case. 4090's 384 bit bus is 50% wider than. Same or less bandwidth overall sounds a tough one for next gen flagship unless they throw in a bucket of cache. Which is not impossible.

Bigger reason I don't feel GB203 is viable for a 5090 is it probably wont have enough cores in it. Rumours do point to it being a good fit for 5080.

Yeah, I don't think we'll ever see another abX02 RTX xx80 card again. It'll likely end up a top big of the next tier every generation now. RTX 3080 seemed more like a fluke than a standard, likely from Samsung's fab having poor yields.

What gives credence to the RTX 5090 being GB202 are some of the leaks that its not being released first, that the RTX 5080 will come first. They'd need at least a few million of a mid-tier bin to justify a release. RTX 5080 being a top-ish bin GB203 would make more sense, since yields on smaller dies should be better. They'd only need to gap or equal the RTX 4090 with an RTX 5080 to sell, regardless of its probability for a lot lower power draw while doing so.

Maybe we'll even see a bottom bin GB202 with a 256bit bus, then quadrupled up for 32GB. Something slightly more than a fully or mostly unlocked GB203 that would slot the RTX 5080 next to it with 16GB to have a gap to the RTX 3080 vs 3090 on launch. Allow them to justify selling the RTX 5080 at $1200 and RTX 5090 at $1600. If they end up using a good bin of the GB202 for the 5090, I feel like we'll see that $2000 MSRP that they could've used for the RTX 4090 but didn't for some reason.

Ryzen 7950x3D Direct Die NH-D15, CCD1 disabled

RTX 4090 @133%/+230/+500

Builder/Enthusiast/Overclocker since 2012聽 //聽 Professional IT since 2017

Link to comment
Share on other sites

Link to post
Share on other sites

1 hour ago, Agall said:

That would make sense with a 448-bit bus, so a mid range bin of a GB202, similar to the RTX 4090's bin of the AD102. I think the only major reason to make the RTX 5090 actually good would be because those halo products sell the generation. People look at RX 7900 XTX vs RTX 4090 and see how massive the gap is at 4K which puts the idea in their head that any RTX 4000 series card is probably better than any RX 7000 at the same price point.

Even a capped 5090 is still going to be a halo product. It just has to deliver minor gains from the 4090 since there will be no competitors to it, so a meh 5090 (considering what it could have been) will still sell like hot cake and be the top tier product in the gaming market, which benefits nvidia since they can just make use of those bad gb202 bins without worries.

2 hours ago, Agall said:

Nvidia might simply not care though, their gaming profits are generally up, but not nearly as much as data center. If they can sell the same GB202 GPU for 3x as much to data center clients, then why would they bother? Especially if TSMC fab time simply can't keep up with demand.

Because defective gb202s can still be sold by a lot with way less compromises and still helping them PR-wise.

1 hour ago, Agall said:

If they end up using a good bin of the GB202 for the 5090, I feel like we'll see that $2000 MSRP that they could've used for the RTX 4090 but didn't for some reason.

i bet it'll be a 2k MSRP model even with a bad bin just because they can, and because it'll sell nonetheless.

Remember that many folks also buy those for professional usage (like me) since they're really cheap compared to the actual pro options.

FX6300 @ 4.2GHz | Gigabyte GA-78LMT-USB3 R2 | Hyper 212x | 3x 8GB + 1x 4GB @ 1600MHz | Gigabyte 2060 Super | Corsair CX650M | LG 43UK6520PSA
ASUS X550LN | i5 4210u | 12GB
Lenovo N23 Yoga

Link to comment
Share on other sites

Link to post
Share on other sites

33 minutes ago, igormp said:

Even a capped 5090 is still going to be a halo product. It just has to deliver minor gains from the 4090 since there will be no competitors to it, so a meh 5090 (considering what it could have been) will still sell like hot cake and be the top tier product in the gaming market, which benefits nvidia since they can just make use of those bad gb202 bins without worries.

Because defective gb202s can still be sold by a lot with way less compromises and still helping them PR-wise.

i bet it'll be a 2k MSRP model even with a bad bin just because they can, and because it'll sell nonetheless.

Remember that many folks also buy those for professional usage (like me) since they're really cheap compared to the actual pro options.

I'm going to guess they'll bin it for ~50% more performance over the 4090, whatever GB202/GB203 bin that comes out to. Then sell it for $1.8k-$2k as effectively their halo product to sell the whole generation to consumers. Maybe only market it as ~50% more in ray tracing and AI, and not even bother talking about rasterization performance.

Ryzen 7950x3D Direct Die NH-D15, CCD1 disabled

RTX 4090 @133%/+230/+500

Builder/Enthusiast/Overclocker since 2012聽 //聽 Professional IT since 2017

Link to comment
Share on other sites

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now