7N--(full node jump)-->5N, 5N--(full node jump)-->3N, etc.
7/5 = 1.4x
5/3 = 1.67x
3/2 = 1.5x
2/1.4 = 1.43x
1.4/1 = 1.4x (aka the mentioned A14 node)
1/0.7 = 1.43x
etc.

---

Quote from: allegedly on March 23, 2026, 16:22:26What makes it worse is the rtx 4090 had a high tdp too but apparently in games it ran with way less wattage and didn't use anywhere close to it's max tdp. Could also undervolt to reduce power even further.

By contrast, the 5090 runs 24/7 at that close to max wattage under load apparently. At least accord to Tom @ MLID.

Not sure I would believe that as both are basically the same arch, on the same node (with Blackwell having the most relevant software features for consumers: MFG. hardware-wise, Blackwell supports FP4, but this requires only minor hardware changes) and having the same energy efficiency / FPS per Watt and even the mm² are is the same for some chips vs its predecessor.

Quote from: allegedly on March 23, 2026, 16:22:26I believe they do. It's called the RTX PRO 5000.

Correct, it's the same GB202-300-A1 chip (en.wikipedia.org/wiki/Blackwell_(microarchitecture)#Blackwell_dies), but obviously, the RTX PRO 5000 is more expensive/not for consumers. But good that you mentioned it, so that people who really need a 48 GB VRAM in a single card, know that it exists (4090 48GB unofficially exists). 48 GB VRAM on the RTX 5090 would cannibalize the more expensive RTX PRO 5000 sales, especially since a RTX PRO 5000 has a 384-bit memory bus width, while the 5090 has a much faster and bigger 512-bit chip. (the RTX PRO 5000 actually uses the (binned) GB202-300-A1 with disabled memory lanes (and/or broken cores))

TSMC 2N / 2nm is fortunately for us all not the end: TSMC A14 / 1.4nm (PS: nothing in the node is actually 1.4nm in size):

Quote from: tsmc.com/english/dedicatedFoundry/technology/logic/l_A14Compared with TSMC's industry-leading N2 process, A14 will offer up to a 15% speed improvement at the same power, or up to a 30% power reduction at the same speed, [..]

So let's continue/redo: 340W / 1.3 = 262W. (=575W/1.3^3 (^3 means 3 full nodes improvement))

What makes it worse is the rtx 4090 had a high tdp too but apparently in games it ran with way less wattage and didn't use anywhere close to it's max tdp. Could also undervolt to reduce power even further.

By contrast, the 5090 runs 24/7 at that close to max wattage under load apparently. At least accord to Tom @ MLID.

Quote from: 5090 using TSMC 2N on March 23, 2026, 11:29:34offer a 48 GB VRAM version

I believe they do. It's called the RTX PRO 5000.

Hey NVIDIA, RTX 5090's 575W TDP is really nasty. How expensive would it be to port 5090's 512-bit chip from TSMC 4N to 2N node? 2 node jumps, and an power efficiency improvement by 30% after each full node jump, means the 575W TDP would become: 575W / 1.3 (TSMC 3N) / 1.3 (TSMC 2N) = 575W/1.3^2 = 340W TDP.

Also, while you are at it, offer a 48 GB VRAM version (using 3 GB per chip chips, instead of the current 2 GB ones).
(RTX 5090 desktop: 512-bit/32-bit per GDDR7 chip = 16 GDDR7 chips, 2 GB each = 32 GB VRAM. Using 3 GB per chip = 48 GB VRAM. The 5090 Laptop, which is a 256-bit 5080 desktop (16 GB VRAM) in disguise) is already using 3 GB per chip chips)