AMD RDNA 5 GPUs reportedly taped out on TSMC N3P, mid-2027 launch hinted

[Redaktion]

Speculation about AMD switching future Radeon GPUs to Samsung Foundry has been shut down by a trusted leaker, who says the chips are already taped out at TSMC. The same source hints at a mid-2027 debut, with little to expect from Radeon in 2026.

https://www.notebookcheck.net/AMD-RDNA-5-GPUs-reportedly-taped-out-on-TSMC-N3P-mid-2027-launch-hinted.1192698.0.html

[p.e.]

TSMC's N3P vs prev N4P node, will hopefully give the usual full node power efficiency improvement of 25% to 33%, not really expecting more (this is not N2P, after all), but ofc it would be nice.

Full node improvement over previous gen:
+15% power efficiency improvement at same power.
+25-33% power efficiency improvement at same performance.
(this is what I roughly remember, numbers might be different)

The full node power efficiency improvement at same performance is especially nice, according to TSMC's slides..imagine a 5090' 575W TDP down to 385W TDP. I have not been interested in a 5090 due to its almost 600W TDP (well, and price ofc), but now the "at same performance" metric becomes interesting. Many don't like the 575W TDP. Ofc, unfortunately many also expect a performance boost, which would reduce the full node 33% efficiency improvement.

5090' 32 GB VRAM (at 512-bit) are not enough per card (certain interesting and capable LLMs only fit into 48 GB VRAM (or RAM ofc)). A 5090 version using the 3 GB VRAM modules, which would give it 48 GB VRAM, would be the minimum I'd be interested in. Indeed, a 5090 48 GB VRAM would eat into the much more expensive RTX PRO 5000 48 GB VRAM (at 384-bit) and 72 GB VRAM (at ?-bit) and be faster.

[p.e.]

and 72 GB VRAM (at ?-bit)

Nvm, it's also at 384-bit (12 GDDR7 chips (384-bit/32-bit per chip), each GDDR7 chip containing 3 GB, at both sides of the PCB (clamshell design): 72 GB = 12 chips * 3 GB per chip * 2)

[Second Class Citizen]

Meh, that's too far away for me. Even when it comes will deal with the usual initial out of stock / lack of volume overpricing. So basically 2028-2029 really.

I just wish they brought rdna4 to igpus faster, as that would be good enough for me. (rdna5 is meant to be only 10% faster in raster but much bigger gains in RT)

It's kind of sad that the first rdna4 igpu to market won't be in an AMD x86 APU but Samsung's Exynos 2600 arm SoC.

You can clearly see what AMD value, they don't care about their own APU's at all (still planning to use rdna3.5 in 2026 and medusa zen 6 apus in 2027). But contracts with Sony and Samsung are far more important and lucrative for them.