Nvidia reportedly leaves RTX 5090 out in the cold as GB202 finds a more lucrative home

[Redaktion]

Nvidia has reportedly stopped allocating new GB202 GPU supplies to the GeForce RTX 5090, reserving future production for its professional RTX Pro graphics cards. If true, the development could worsen RTX 5090 availability, which is already pretty spotty.

https://www.notebookcheck.net/Nvidia-reportedly-leaves-RTX-5090-out-in-the-cold-as-GB202-finds-a-more-lucrative-home.1420923.0.html

[good decision]

good decision, no gamer needs a 5090's 32 GB VRAM really, let the AI people use the 512-bit GB202 chips in RTX PRO 5000, 5500 and 6000

[good decision]

even for 3000 I wouldn't get a 5090 due to risk of melting connector, let alone for an even higher price

5090: 575W/32GB = 18 TDP Watt per GB
TDP Watt per GB = (1-((WattTDP/GB)/18))*100:
5060 Ti 16G: 180W/16GB = 11.25 TDP Watt per GB (-37.5%)
5070 Ti Super: ~320/24GB = 13.3 TDP Watt per GB (-26%)
5080 Super: ~380W/24GB = 15.8 TDP Watt per GB (-12%)

[good decision]

If you still want/need 32 GB VRAM because of those useful SOTA[1](other models are also in the comparison) Qwen3.8-27B dense, Qwen3.6-35B MoE or other models in a similar size, on the most recent GPU architecture, your best bets would be:

• 2x 5060 Ti 16 GB
• 2x 5070 Ti (Super) 24 GB / 5080 (Super) 24 GB (not mentioned in this news and not released yet, but is rumored to replace the 5090, at least for the time being)

in tensor parallel mode (VRAM and memory speed add up and 2 GPUs act like 1) on a motherboard that has built-in bifurcation of the main PCIe 5.0 x16 slot into x8/x8[2], see also e.g. reddit.com/r/LocalLLaMa/search/?q=tensor parallel.

Strictly speaking you don't need such a motherboard if you are ok with not utilizing those GPUs to their fullest[3](someone actually did some testing).

Note: If you decide to go for bifurcated PCIe 5.0 x16 into x8/x8, make sure that the GPUs are also PCIe 5.0. On NVIDIA it's only recent RTX 50 series, and AMD's RDNA4 is still on GDDR6:
NVIDIA 128-bit: 5060 Ti (16GB): 448 GB/s (+39% faster token generation due to GDDR7 memory).
AMD 128-bit: 9060 XT (16GB): 322 GB/s.

[1] artificialanalysis.ai/models/open-source/small?models=qwen3-8-27b-medium%2Cqwen3-8-27b-low%2Ck2-horizon-mova-36b-a4b%2Cg9v3-39a5b%2Ck2-horizon-7b%2Cqwen3-8-27b-non-reasoning%2Cqwen3-6-35b-a3b%2Cmuse-glimmer%2Cgemma-4-26b-a4b%2Cqwen3-6-35b-a3b-non-reasoning%2Cgemma-4-31b%2Cqwen3-8-27b#artificial-analysis-intelligence-index (K2 Horizon has expensive KV cache and may reason above average
[2] geizhals.eu/?cat=mainboards&xf=2833_AMD+AM5~2873_ab+2&pagesize=30&promode=true
[3] reddit.com/r/LocalLLM/comments/1uqbk9k/comment/owof7gx/

[good decision]

Using 2x 5070 Super 18 GB (same chip as a 5070, just using 3 GB GDDR7 dense memory chips, just like with the other Supers, instead of the 2 GB ones) would also be an option, but whether it's going to be released, nobody knows (it was supposed to be released at the beginning of 2026, but then AI memory-geddon happened).

[good decision]

(not mentioned in this news and not released yet, but is rumored to replace the 5090, at least for the time being)

And here is the first news about it (that's my first and last post about it):
notebookcheck.net/Nvidia-s-rumoured-RTX-5080-24-GB-could-become-its-most-powerful-GeForce-GPU-available-in-2027.1422225.0.html
(notebookcheck's own anti-spam doesn't even allow to add https to notebookcheck.net's own site..what a dumb system)

To add to the Watt per GB metric:
RTX PRO 4500 Blackwell: 200W/32GB = 6.25 TDP Watt per GB (-65%), or in other words, for the same 32 GB VRAM, the 5090 has a 2.875 times (=(575/32)/(200/32)) higher TDP.

Of course, the 5090 is 1.8 TB/s and up to twice the performance, while a 5070 Ti (Super) (24GB), 5080 (Super) (24GB) and Pro 4500 Blackwell all use the half-sized 256-bit GB203 chip at an also halved speed of 0.9 TB/s (and the halved token generation that results of the halved TB/s) (en.wikipedia.org/wiki/Blackwell_(microarchitecture)#Blackwell_dies), but remember in tensor parallel, two 256-bit GPUs act like one, and then you'd have 48 GB VRAM at 1.8 GB/s ;-) (NGREEDIA knows this and so do we. Or wait when AMD offers GDDR7 GPUs.)