Question NVIDIA Rubin H2-2027

[dangerman1337]

I am not sure if RTX 60 will be using TSMC N3P necessarily and 3GB GDDR7 Modules. If RDNA 5 especially AT0 comes out months before and is very performant as some rumours indicate and 3GB Modules, Nvidia may just bite the bullet with 4GB modules and treat RTX 60 like RTX 30 and go with SF2X process and price competitively Vs RDNA 5 & Next-Gen Consoles with more flexibility with margins. Like GR202 gets produced en masse with 512-bit SKUs to maybe even 320-bit SKUs because they're producing so much. Considering they where able to price the 5070 Ti at $749 with TSMC N4 at 16-18K a wafer, if DRAM prices collapses by the time RTX 60 goes into production Nvidia could even do a 32GB 6070 Ti (especially Vs a PS6) that uses the full GR203/204 die at $800 and a cut down 28GB 6070 at $650 or so if they're on 20K a wafer SF2X process node. Or do non-Ti SKUs for gamers and the Tis that fully enable the dies in 2028 just like RTX 30 did.

All assuming Open AI Circle Jerk bubble fully bursts this year which is very plausible.But all this could be wrong lol.

As for Rubin, I wouldn't be surprised they end up calling the full GR203 the 6080 Ti. Mainly because of the MSRP which will surely be much more than $999.

Only way they do that is if Nvidia goes with 4GB Modules and RDNA 5 flops.

 

[adroc_thurston]

and go with SF2X process and price competitively Vs RDNA 5 & Next-Gen Consoles with more flexibility with margins.

SF2X will be barely if any cheaper (especially when yielding big die).

 

[jpiniero]

SF2X will be barely if any cheaper (especially when yielding big die).

I have to think it'd be a lot cheaper, but yeah yield is an issue. I think they would only use it for the smaller dies like 6/7/8. It also probably won't clock as high as N3P so you'd have to take that into consideration.

Only way they do that is if Nvidia goes with 4GB Modules and RDNA 5 flops.

4 GB modules won't be available then. It's more of a 2028 thing.

 

[adroc_thurston]

I have to think it'd be a lot cheaper

Nope, they're charging normal 3nm class prices.

4 GB modules won't be available then

CPX and AT0 big boy configs are both built around 32Gb GDDR7.

 

[jpiniero]

Nope, they're charging normal 3nm class prices.

Who knows what TSMC is pricing N3P now... 20k would probably be a lot cheaper than that.

That is the problem with Samsung and Intel... it can't be too cheap to fund future nodes properly.

 

[adroc_thurston]

20k is a lot less than 30+, you know.

N3p wafer is ~$20.5k after price hikes.

 

[jpiniero]

N3p wafer is ~$20.5k after price hikes.

I'll take the over.

 

[jpiniero]

So looking at it again, i'd say the full GB203 ($1299+ depending on ram prices) be like 10-20% faster than the 4090 at 4K, and maybe the cut one ($999+) being about 4090 performance. With the full being closer to the 5090 at lower resolutions. Figure the 5090 will look better at lower resolutions with Zen 6D.

I am not expecting much other than clock speed really and DLSS 5. Maybe 36 gbps memory will help a bit.

 

[MrMPFR]

Well, 16 GPCs are visible on the diagram. It does not mean, that the chip will really look like this.

But I think 16 GPCs is reasonable. I can think of a few reasons for that:

• Better scaling in general, because 1.33x GPCs (ROPs anyone? 😀)
• Power of 2 scaling for ML/AI Workloads: A100, H100 and B200 all have 8x GPCs. Rubin HPC might have 8 or 16 and Rubin CPX would have 16 GPC as well. ML/AI workloads like power of 2 divisions (MI350X did go back to 256 CU because of that, MI300X featured 320 CU)
• GauRast: Gaussian Splatting Acceleration (Neural Rendering) within the rasterizer. 1.33x GPCs will bring a boost there (see the GauRast paper https://arxiv.org/html/2503.16681v1)
• Transformer Attention Acceleration (see GB300 or Rubin CPX) does benefit from exponential functions. On GB300 Nvidia says they pimped the SFU within the SM for that (or emulate SFU EX2 functions). GauRast from the previous point will introduce additional EXP-Units within the rasterizer (GPC frontend). This might create some synergies between Neural Rendering and general ML/AI.

Yup GauRast + attention softmax

Maybe also HW hashgrid encoding + cachemem overhaul to fuse directly with tensor cores + dedicated or leasing scratchpad for neural rendering: https://arxiv.org/abs/2303.05735
A bit of browsing suggest some sort of post- and/or modified MLP shift at some point as well. Just a few examples:
https://arxiv.org/abs/2512.19522v1, https://arxiv.org/abs/2504.12273, https://arxiv.org/abs/2407.10482, https://arxiv.org/abs/2312.09398

Can see NVIDIA leaning into all this for Rubin to counter RDNA5. One-two orders of magnitude speedup (SW+HW) over 50 series + MLPs seems doable.

 

[MrMPFR]

https://twitter.com/x/status/2008898284933517714

NOT a giant for gaming
That mean maybe another 30% compared to RTX 5090?

Any ideas of how much inference perf will get cut down for consumer GB20x?