News Nvidia announces Rubin GPUs in 2026, Rubin Ultra in 2027, Feynam after

[Admin]

Nvidia provided its latest data center GPU roadmap update, showing the Rubin platform slated for release in the second half of 2026, with Rubin Ultra planned for the second half of 2027. The next generation architecture after Rubin will be named after Richard Feynman, a theoretical physicist.

Nvidia announces Rubin GPUs in 2026, Rubin Ultra in 2027, Feynam after : Read more

 

[usertests]

There will also be 1TB of HBM4e per four reticle-sized GPUs, with 100 PetaFLOPS of FP4 compute.

Has anyone figured out where the increase from 288 GB to 1 TB (maybe 1152 GB?) is coming from?

Is it from 4x more stacks than the 288 GB GPUs (which are using two dies), more layers per stack with HBM4e, density increases, etc.

Edit: Blackwell Ultra uses 8 stacks of 12-Hi HBM3e. Which is 36 GB per stack, 3 GB per layer.

I believe the new one is using 16 stacks. It's probably 16 stacks of 16-Hi HBM4e with 4 GB per layer (64 GB per stack), adding up to 1024 GB total.

 

[SonoraTechnical]

meh who cares.. the good stuff won't be affordable, nor readily available... you're dead to me nVidia..

 

[TheyStoppedit]

Each card will be 10% faster than it's same tier of the 50 series, (with a 10% power increase) besides the 6090 will be 30% faster than the 5090. The 6090 will be 675W (the extra 75 from the PCIe slot). They won't fix the current balancing (der8auers/buildzoids videos from last month). The 6070 will be advertised as faster than the 5090 with half the VRAM thanks to 8x frame gen. (8 fake frames instead of 4)

6090 - $2499
6080 - $1799
6070Ti - $1499
6070 - $1299
6060Ti - $999
6060 - $799

You read it here first

 

[usertests]

meh who cares.. the good stuff won't be affordable, nor readily available... you're dead to me nVidia..

You're dead to Nvidia too. Both sides will change their tune if the bubble bursts.

Each card will be 10% faster than it's same tier of the 50 series, besides the 6090 will be 30% faster than the 5090. The 6070 will be advertised as faster than the 5090 with half the VRAM thanks to 8x frame gen. (8 fake frames instead of 4)

Junk prediction. The reason that Blackwell performance gains are so low is simple: It's on the exact same 5nm-class node as Lovelace. The successor moving to a 3nm or 2nm node will make it a lot better.

 

[TheyStoppedit]

Junk prediction. The reason that Blackwell performance gains are so low is simple: It's on the exact same 5nm-class node as Lovelace. The successor moving to a 3nm or 2nm node will make it a lot better.

No, it's because there's no R&D going into performance uplifts anymore. Why? Because all the R&D has gone into data centers/enterprise/AI. This is where the cash cow is. Why would they put a GB202 into a 5090 and charge $2000 when they can put it into a datacenter GPU and charge $30,000?

From here on in, every gen will have 10% uplifts on each tier (besides the 90) and the rest of the performance will be subsidized by fake frames/DLSS. Part of this is because there is no competition from team red. NVidia used to be a gaming GPU company. We are in the midst of a massive AI boom. Do you think NVidia cares about me or you and your stupid frames?

Nope, from now on, all R&D is datacenter/AI/enterprise/fake frames/DLSS. Datacenter/AI/enterprise will get what they want because they pay the most (unless you want to pay $30,000 for a 6090). Gamers, you get whatever chips can't pass QC for Datacenter/AI/enterprise and more fake frames/DLSS and a 20% price hike.

The shareholders will get what they want and your frames will look like garbage to make it happen. This is a business, and there is only one thing that matters: MONEY.

 

[usertests]

From here on in, every gen will have 10% uplifts on each tier (besides the 90) and the rest of the performance will be subsidized by fake frames/DLSS.

This is incorrect for the reason I stated.

They don't care about consumer GPUs right now, but they can pivot back to it whenever they feel the need to. Such as if a bubble bursts.

 

[Giroro]

So Nvidia expects tech companies to just buy a new slightly-better multi-billion dollar AI super computer every year? Nobody's found a way to even get close to paying for their current ones yet.

Where is that kind of money supposed to keep coming from?
I mean, in the short term they're trying to get it from global-scale enshittification. But even if every app somehow maintained their current level of engagement while converting every user to a paid subscriber, and filling the screen with a ratio of 100% ads and 0% content/functionality, revenue would still be at least an order of magnitude away from breaking even. It's not a sustainable business model.

It's pointless for Nvidia to try to annualize these products. The next upgrade cycle for these machines will be in at least 5 years, probably more like 10 or 20. Every company already has their machines either built or on order so it's all one-and-done for Nvidia until one of these companies completes their supervillain plan of breaking the stock market, or cracking crypto, or whatever it is they're actually trying to accomplish.

 

[spongiemaster]

No, it's because there's no R&D going into performance uplifts anymore. Why? Because all the R&D has gone into data centers/enterprise/AI. This is where the cash cow is. Why would they put a GB202 into a 5090 and charge $2000 when they can put it into a datacenter GPU and charge $30,000?

From here on in, every gen will have 10% uplifts on each tier (besides the 90) and the rest of the performance will be subsidized by fake frames/DLSS. Part of this is because there is no competition from team red. NVidia used to be a gaming GPU company. We are in the midst of a massive AI boom. Do you think NVidia cares about me or you and your stupid frames?

Nope, from now on, all R&D is datacenter/AI/enterprise/fake frames/DLSS. Datacenter/AI/enterprise will get what they want because they pay the most (unless you want to pay $30,000 for a 6090). Gamers, you get whatever chips can't pass QC for Datacenter/AI/enterprise and more fake frames/DLSS and a 20% price hike.

The shareholders will get what they want and your frames will look like garbage to make it happen. This is a business, and there is only one thing that matters: MONEY.

Did you bother to read the article? The roadmap is for data center AI accelerators, and has nothing to do with gaming GPU's.

 

[spongiemaster]

So Nvidia expects tech companies to just buy a new slightly-better multi-billion dollar AI super computer every year? Nobody's found a way to even get close to paying for their current ones yet.

No, everyone isn't on the same upgrade cycle. Some are still trying to get into the market and the hyper scalers are constantly buying whatever is the latest and greatest to add to what they already have, not simply to upgrade what they have.

 

[spongiemaster]

meh who cares.. the good stuff won't be affordable, nor readily available... you're dead to me nVidia..

Nvidia doesn't care about individual customers like you either who aren't in the market for thousands of their $40k AI accelerators. Nothing in this article you didn't have time to read before commenting, has anything to do with you.

 

[usertests]

So Nvidia expects tech companies to just buy a new slightly-better multi-billion dollar AI super computer every year? Nobody's found a way to even get close to paying for their current ones yet.

Where is that kind of money supposed to keep coming from?

It's pointless for Nvidia to try to annualize these products.

They have had effectively unlimited demand for their AI accelerators, so anything they make sells. Maybe that will slow down, but if everything's still on backorder, anything they make = revenue/profit. What's this about companies not being able to pay? They can borrow money to pay Nvidia if they need to. As long as Jensen gets paid, the system works.

I'm not sure about the annualization strategy. It has allowed them to make physically larger accelerators using MCMs, and maybe these are more profitable. Customers can benefit from the memory capacity increases, and they may be required for inference of a larger LLM on a single chip. I believe there are some valuable efficiency gains being realized, even if perf per die area isn't going up that much from Blackwell-Blackwell-Blackwell.

If a company has enough Nvidia chips, someone else will take their place in line.

 

[jackt]

trying to distract from theyr troubles...

 

[JTWrenn]

Everyone here complaining about talk of datacenters...at a damn AI datacenter conference just needs to take a breath. This tech will trickle down to some degree but this really had nothing to do with things we are going to buy. This was all datacenter all the time. They just announced a new gpu line 2 months ago and now they are announcing upcoming and releasing stuff for ai and datacenter. If they did it all at every conference the thing would still be going on.

 

[kdw75]

Only problem is that they can't seem to make a decent amount of their GPUs. nVidia can't supply the data centers for some reason.

 

[Mr Majestyk]

You're dead to Nvidia too. Both sides will change their tune if the bubble bursts.


Junk prediction. The reason that Blackwell performance gains are so low is simple: It's on the exact same 5nm-class node as Lovelace. The successor moving to a 3nm or 2nm node will make it a lot better.

Yet AMD delivered 9070XT with near 7900XTX performance in raster with 33% less core on the exact same node as Nvidia. Amazing how AMD got IPC uplifts on N4P. Leatherman went all in on AI BS and multiframe garbage to announce bogus performance uplifts. We weere told a few years ago Blackwell would be a big change architecturally, clearly not. Even GDDR7 didn't help it much.

 

[alchemy69]

Richard GPUs would be a Dick move.

 

[renz496]

Yet AMD delivered 9070XT with near 7900XTX performance in raster with 33% less core on the exact same node as Nvidia. Amazing how AMD got IPC uplifts on N4P. Leatherman went all in on AI BS and multiframe garbage to announce bogus performance uplifts. We weere told a few years ago Blackwell would be a big change architecturally, clearly not. Even GDDR7 didn't help it much.

Same with nvidia actually. 5070Ti and 5070 for example able to perform like 4080S and 4070S with a bit less CUDA core for pure raster/RT without things like DLSS/FG/MFG. And i imagine blackwell is even more flexible/performing in some non gaming task because all nvidia CUDA cores now can do both FP32/INT32 while with ada/ampere it was FP32 + FP32/INT32.

 

[George³]

This is incorrect for the reason I stated.

They don't care about consumer GPUs right now, but they can pivot back to it whenever they feel the need to. Such as if a bubble bursts.

Maybe we must shoot on AI bubble to be bursting? Because billionaires wouldn't do it voluntarily.

 

[bit_user]

Where B300 NVL72 offers 1.1 PFLOPS of dense FP4 compute, Rubin NVL144 — that's with the same 144 total GPU dies — will offer 3.6 PFLOPS of dense FP4.

@JarredWaltonGPU , the slide says EFlops.

​

Vera will be a relatively small and compact CPU, with 88 custom ARM cores and 176 threads.

IMO, this is probably equally important to the Rubin news. So, it's back to custom Arm cores - and with SMT!

I wonder if those will show up in a consumer SoC, or if the rumored consumer offering from Nvidia is limited to their partnership with MediaTek.

The inference compute with FP4 will rocket up to 15 ExaFLOPS, with 5 ExaFLOPS of FP8 training compute. It's about 4X the compute of the Rubin NVL144, which makes sense considering it's also four times as many GPUs.

Thermal density should be a major challenge. I wouldn't be surprised if they end up being only 2.5x to 3x as fast, due to the need to clock them lower.

Presumably that means we'll get Richard CPUs with Feynman GPUs, if Nvidia keeps with the current pattern.

I still think that's incredibly lame. And, as I predicted, it created a situation where we now have "Grace Blackwell". Who is Grace Blackwell? Nobody. At least, nobody specific. Ridiculous.

I think Feynman wouldn't approve.

 

[bit_user]

They don't care about consumer GPUs right now, but they can pivot back to it whenever they feel the need to. Such as if a bubble bursts.

They actually do care about client GPUs, at least at the upper end (GB102 and GB103), because they use those same dies to make inferencing cards like this:

https://www.tomshardware.com/pc-com...en-more-demand-for-the-limited-supply-of-gpus
​

That's why they went with a 512-bit GDDR7 datapath - not because gamers needed it!

 

[valthuer]

Yet AMD delivered 9070XT with near 7900XTX performance in raster with 33% less core on the exact same node as Nvidia. Amazing how AMD got IPC uplifts on N4P. Leatherman went all in on AI BS and multiframe garbage to announce bogus performance uplifts. We weere told a few years ago Blackwell would be a big change architecturally, clearly not. Even GDDR7 didn't help it much.

Look at the bright side; at least Nvidia managed to present us with a better flagship.

What did AMD manage to do? Create a GPU that can't even surpass their own best card, from 3 years back?

If that's their notion of making progress, we shouldn't be surprised with the bogus performance uplifts and crappy prices we 're getting from Nvidia.

I mean, what are they supposed to do; compete with themselves?

 

[bit_user]

What did AMD manage to do? Create a GPU that can't even surpass their own best card, from 3 years back?

If that's their notion of making progress, ...

AMD has been clear that RX 9070 isn't meant to be a flagship. It's a mid-range. They actually had a proper RDNA4 flagship that they cancelled.

we shouldn't be surprised with the bogus performance uplifts and crappy prices we 're getting from Nvidia.

I mean, what are they supposed to do; compete with themselves?

Not everyone can afford a Nvidia flagship GPU. For those that can't, what AMD and even Intel are doing still have relevance, because they are competitive, further down the stack.

 

[valthuer]

AMD has been clear that RX 9070 isn't meant to be a flagship. It's a mid-range. They actually had a proper RDNA4 flagship that they cancelled.

Yes, we already knew they weren't gonna release a flagship, but that doesn't make it any less disappointing.

AMD has been clear that RX 9070 isn't meant to be a flagship. It's a mid-range. They actually had a proper RDNA4 flagship that they cancelled.


Not everyone can afford a Nvidia flagship GPU. For those that can't, what AMD and even Intel are doing still have relevance, because they are competitive, further down the stack.

Not denying AMD/Intel's relevance, but, as things stand right now, i'm getting the sense that the consumer GPU market could benefit from some competition in the high-end segment.

Not that it really matters to either Nvidia, or AMD, as their earnings from discreet GPUs pale in comparison to their profit from AI/datacenters.

So, i'm guessing we'll be seeing more of the same down the road.