Question NVIDIA Rubin H2-2027

[Tangopiper]

Let's collect and discuss all things Rubin here!

Thanks to kopite7kimi on X, we know that client Rubin will use the GR20x codenames, Rubin CPX GR21x and Rubin AI GR10x.

We also know that GR100 is made of two GR102 dies. It is unknown if GR102 will be shipped as a dedicated product - going by Blackwell, it will likely not.

Everything else is my speculation assuming similar binning and memory configuration to Blackwell and a move to 1 GPC = 8 TPC / 16 SM for most parts.

Memory configurations based on the fact that ~16GB will be the minimum required for mid-range and up. NVIDIA way well skimp and give is less on 5060 and 5050 models.

Codename	Node	Chip design	Full GPC / SM configuration	Model	Memory Configuration
GR100	TSMC N3 family	2x GR102 MCM		R100	8192b 288GB HBM4
GR102	TSMC N3 family	Monolithic		???	4096b 144GB HBM4
GR212	TSMC N3 family	Monolithic		Rubin CPX	512b 128GB GDDR7
GR202	TSMC N3 family	Monolithic	12 / 192	6090	512b 48GB GDDR7
				6080 Ti	384b 36GB GDDR7
GR203	TSMC N3 family	Monolithic	6 / 96	6080	256b 24GB GDDR7
				6070 Ti	256b 24GB GDDR7
GR205	TSMC N3 family	Monolithic	4 / 64	6070	192b 18GB GDDR7
GR206	TSMC N3 family	Monolithic	3 / 48	6060 Ti	160b 15GB GDDR7
				6060	160b 15GB GDDR7
GR207	TSMC N3 family	Monolithic	2 / 24	6050	128b 12GB GDDR7

 

[jpiniero]

Memory configurations based on the fact that ~16GB will be the minimum required for mid-range and up.

Assume the memory bus will be less and not more.

 

[Tangopiper]

Assume the memory bus will be less and not more.

You are very likely correct - but I’m hopeful we might actually get some increases next gen. If the shortage abates by then…

 

[Win2012R2]

I’m hopeful we might actually get some increases next gen.

36-40 Gbs GDDR7 should be normal in 2027, plus likely 4Gbit modules will appear, which should make 3Gbits "cheap".

Looking at how efficient RDNA4 is with GDDR6 one would imagine Nvidia will do more with less

 

[MoogleW]

My first guesstimations based on the specs from CES.

~45% average FPS uplift for a chip at the same die size (60% higher FP32 TFLOPS).

Memory modules may be 28-36 gbps. Memory bandwidth is currently not a pressing concern unless the gains are huge which they won't be. Global caches may see changes but in design rather than sizes. I hope the focus as well on local caches and register files but I dooubt it unless it is to bring tensor cores and shaders even closer for that neural rendering workload.

Based on my modeling;
GR203: ~ 5090, but more like 90-95% of the performance
GR205: ~ 4090
GR206: ~rx 7900XT level

GR207 is tough, die size may shrink or stay the same,
Shrink: ~rtx 5060, (around 25% gain)
Same: ~5060ti, around 45% gain

 

[SolidQ]

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

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

 

[jpiniero]

~45% average FPS uplift for a chip at the same die size (60% higher FP32 TFLOPS).

If they spend transistors on anything, it'll be mainly tensor cores. I'd expect the GR205 and below to be smaller dies too.

 

[CakeMonster]

OP's assumption that the top chip would be used for both the 6090 and 6080Ti part (instead of AI SKU's) seems very optimistic, and so is the VRAM amounts. We had a VRAM bump with 5090, and would be very lucky to get another one with 6090 and 6080Ti as well. I hope I'm wrong of course.

 

[jpiniero]

OP's assumption that the top chip would be used for both the 6090 and 6080Ti part (instead of AI SKU's) seems very optimistic, and so is the VRAM amounts. We had a VRAM bump with 5090, and would be very lucky to get another one with 6090 and 6080Ti as well. I hope I'm wrong of course.

They will use 3 GB chips, so you will get a memory capacity increase (except for models where they cut the bus of course)

4 GB chips might be available later too.

 

[Golgatha]

 

[Tangopiper]

OP's assumption that the top chip would be used for both the 6090 and 6080Ti part (instead of AI SKU's) seems very optimistic, and so is the VRAM amounts. We had a VRAM bump with 5090, and would be very lucky to get another one with 6090 and 6080Ti as well. I hope I'm wrong of course.

I should have put a note, 6080 Ti is only going to happen in the event RDNA 5 AT0 is actually getting close to Rubin GR202. Very much in my dreams.

Memory will be getting a bump just based on the use of 3GB capacity modules. As noted in the original post, RTX 6050 and 6060 might get less if NVIDIA are stingy as usual. Such as:
RTX 6050 = 96b 9GB GDDR7
RTX 6060 = 128b 12GB GDDR7

6060 Ti will still need to be approx. 16GB otherwise it will be another 3060 > 4060 situation.

 

[jpiniero]

I'm wondering if they will revive the 8 die and make that 64-bit.

Then when 4 GB chips are available, then you will see 6050 with 8 GB still, with two 4 GB chips.

 

[adroc_thurston]

in the event RDNA 5 AT0 is actually getting close to Rubin GR202

that's really really really really easy.
Winning is the hard part.

 

[jpiniero]

that's really really really really easy.
Winning is the hard part.

Just like Zen 5 has 50% better IPC, eh?

 

[adroc_thurston]

Just like Zen 5 has 50% better IPC, eh?

yeah, it does, in server. next.

 

[jpiniero]

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.

 

[Tangopiper]

that's really really really really easy.
Winning is the hard part.

I’d hope so. But it’s Radeon, snatching defeat from the jaws of victory is their thing

 

[Tangopiper]

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.

Don’t give leather jacket man ideas. But it’s absolutely a possibility the stack gets moved around again.

GR203 for 80 and 80 Ti
04 code name returns with GR204 for 70 and 70 Ti

 

[Tigerick]

H2-2027?

Here are my reasons Rubin might launch in Q4-2026:

• https://www.techpowerup.com/340207/...-production-at-tsmc-ceo-jensen-huang-confirms. 6 Rubin chips. NVIDIA Rubin CPX is expected to be available at the end of 2026.
• The cancellation of RTX50 Super series.
• Intel and AMD are going to launch desktop CPU in H2-2026: Nova Lake and Zen6. That's mean NVL-HX and Gator Range should launch in Q1-2027. NV always align the launching of GPU with flagship desktop and mobile CPU; do you really think NV will miss the cycle? FYI, Mobile RTX50 was launched in Feb 2025 aligned with ARL-HX and Fire Range launching.
• Samsung has started sampling 24Gb 36Gbps GDDR7 in Nov 2025.

 

[jpiniero]

The cancellation of RTX50 Super series.

Doesn't seem like it's cancelled cancelled... just indefinitely delayed.

 

[MrMPFR]

As noted in the original post, RTX 6050 and 6060 might get less if NVIDIA are stingy as usual. Such as:
RTX 6050 = 96b 9GB GDDR7
RTX 6060 = 128b 12GB GDDR7

They might be even more stingy thanks to 36-40gbps GDDR7:
- 9-12GB 30SM GR207 96bit 6060, no 6050 (except mobile)
- 12-16GB 48SM GR206 128bit 6060TI

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

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

30% is very little and also Kopite7kimi post prob also rules out this:

IIRC CPX extrapolation was 8x raw NVFP4 dense vs 50 series. IF DLSS (currently FP8) works with NVFP4 then that's 16X matmul at iso-core + clocks.

Thinking CPX = 6090 die with all gfx stripped out except ROPS for AI slop generation + massively boost ML + SFU.
How much NVFP4 gain per SM for GB20x series. Any guesses?

that's really really really really easy.
Winning is the hard part.

Wondering what AT0 and GR202 full die will be?
Both prob 192 SM/CU but how many SEs (AT0) and GPCs (GR202)? 8 and 16?

 

[adroc_thurston]

Both prob 192 SM/CU but how many SEs (AT0) and GPCs (GR202)? 8 and 16?

Probs.
GB202 had scaling issues so it definitely needs more balance.

 

[jpiniero]

Probs.
GB202 had scaling issues so it definitely needs more balance.

it's power limited

 

[adroc_thurston]

it's power limited

no, it struggles with shader count scaling.

 

[ToTTenTranz]

Doesn't seem like it's cancelled cancelled... just indefinitely delayed.

And if the 50 Super are delayed but still coming, this only hints at Rubin coming later, not sooner.


Nvidia has little to gain with launching consumer Rubin early:

- Consumer GPUs being made out of latest-gen nodes that would (not) compete with server Rubin chips sold for a much higher markup;
- Expensive DRAM for which the markup wouldn't go towards Nvidia (they actually stopped bundling the RAM with the GPU chips and are now out of that pipeline) and the Rubin chips are probably planned to have more VRAM (I don't think the 6060 is coming with 8GB, nor the 6070 is coming with 12GB) which means they'd be a lot more expensive at this point, or Nvidia would have to cut on their own margins.
- AMD hasn't shown any plans or hints of a RDNA4 refresh nor a 2026 release of RDNA5, so it's not like Nvidia is feeling any pressure.
- Releasing Rubin early would place more pressure to releasing its successor early as well.



The only valid reason I see for rushing Rubin is if the lower end models use cheaper LPDDR6 like AMD's AT3 and AT4 versions of dGPUs, for which the savings would be directed at Nvidia increasing the markup for their chips while maintaining similar price-per-FPS to the audience.