Posted April 08, 2025
The 5000 series, the "consumer" Blackwell chips without the 5090, are made with the slightly older N4 = 5 nm chips, so they do not exactly use the most pricey waver TSMC got to offer, rather the second highest priced waver. This means, although they still use "fab capacity" in some way, they are not directly competing with the same node the AI industry is using. The 5090, it seems... are simply using the most crap "AI chips" who got so many defective cores so they are "reduced" and offered as a "gamer flagship chip". All the other 5000 series chips are made separately on a "lower node".
The 9070 series, and the 5090 along with the datacenter/AI/server-chips seems to be made of a slightly improved node, the N4P = 5 nm.
Info here: https://fuse.wikichip.org/news/6439/tsmc-extends-its-5nm-family-with-a-new-enhanced-performance-n4p-node/
The datacenter versions of Blackwell as well are offering HBM3e memory, which is faster than GDDR7, at a way higher capacity... so the "apparently awesome 5090" is not in the same "ballpark" vs. the true Blackwell which is not for gamers.
The "Normal Industry" Blackwell"= GB 202 using N4P with 24756 cores and 96 GB HBM3e RAM/VRAM (full scale, it could be reduced at the final stage). Operating at 600 W or more, dependable on final design. Higher efficiency than the gamer version it seems. https://videocardz.com/newz/nvidia-gb202-blackwell-760mm%C2%B2-gpu-die-shot-revealed-24756-cores-and-512-bit-bus
The "Gamer Blackwell"= GB 202-300 using N4P, with 21760 cores and 32 GB GDDR7 VRAM, not a single core more. Operating at around 600 W dependable on design. "-300" simply means "cut down design".
There is probably a "GB202-300-A1" variant with the difference that it has been built with the lesser N4 node as well (comparable with the other 5000 series) which means... lower efficiency.
Not even top of the line because...
The Server-Blaclwell = GB 102x2 using N4P, is offering 33792 cores and 192 GB HBM3e RAM, although no detailed specs are known. Its for special customers on a "above consumer market": https://www.techpowerup.com/gpu-specs/b100.c4275
Sure the "Server Blackwell" needs 1000 W, but it got 33792 cores, so its efficiency is actually higher than what we know on the "Gamer Blackwell". Of course the Server-Blackwell can not play games because it was not made for it... it is handling datacenter tasks and everything involving AI on the industry level.
Anyway, you see.... even a 5080 which is second highest gaming-tier seems like a "Baby Blackwell" , as it is way out of "this ballpark" with its around 10 000 cores only, even on a lower N4 fab. Still not free for sure.... because 1000 cowboy coins is surely not enough... it still needs way more.
Of course if someone wants to get a 5070 using... uhm... 6144 cores and 12 GB VRAM... then go for it but i would feel betrayed buying "such a crap". It seems to be made for those "eating to much breadcrumbs" and always thanking for it.
You could as well go to Steam, paying lot of bucklets, asking them to provide some DRM locked data you never own... provided by a "33792 core Blackwell" while your own Blackwell is 6144 cores with 12 GB VRAM and please do not forget to "give the game back, as soon as played or account banned... no matter which comes first".
Regarding PS5 Pro: I do not know exactly the node used there but it is most likely the most advanced node, so N4P (same as the RDNA 4 node), else it would be difficult for Sony able to provide this performance with not much more than 200 W TDP.
Sure, in theory N3P is even more advanced, but it is currently not used on non portable GPUs because of some complicated issues with a very high transistor count. GPUs are already expensive enough and it would add even more costs and some other issues (cooling, higher failure rate at higher transistor count).
AMD has been a very loyal customer of TSMC (for Nvidia it was less stable in the past...), so they certainly are a "premiere customer" with good offers provided to them.
The 9070 series, and the 5090 along with the datacenter/AI/server-chips seems to be made of a slightly improved node, the N4P = 5 nm.
Info here: https://fuse.wikichip.org/news/6439/tsmc-extends-its-5nm-family-with-a-new-enhanced-performance-n4p-node/
The datacenter versions of Blackwell as well are offering HBM3e memory, which is faster than GDDR7, at a way higher capacity... so the "apparently awesome 5090" is not in the same "ballpark" vs. the true Blackwell which is not for gamers.
The "Normal Industry" Blackwell"= GB 202 using N4P with 24756 cores and 96 GB HBM3e RAM/VRAM (full scale, it could be reduced at the final stage). Operating at 600 W or more, dependable on final design. Higher efficiency than the gamer version it seems. https://videocardz.com/newz/nvidia-gb202-blackwell-760mm%C2%B2-gpu-die-shot-revealed-24756-cores-and-512-bit-bus
The "Gamer Blackwell"= GB 202-300 using N4P, with 21760 cores and 32 GB GDDR7 VRAM, not a single core more. Operating at around 600 W dependable on design. "-300" simply means "cut down design".
There is probably a "GB202-300-A1" variant with the difference that it has been built with the lesser N4 node as well (comparable with the other 5000 series) which means... lower efficiency.
Not even top of the line because...
The Server-Blaclwell = GB 102x2 using N4P, is offering 33792 cores and 192 GB HBM3e RAM, although no detailed specs are known. Its for special customers on a "above consumer market": https://www.techpowerup.com/gpu-specs/b100.c4275
Sure the "Server Blackwell" needs 1000 W, but it got 33792 cores, so its efficiency is actually higher than what we know on the "Gamer Blackwell". Of course the Server-Blackwell can not play games because it was not made for it... it is handling datacenter tasks and everything involving AI on the industry level.
Anyway, you see.... even a 5080 which is second highest gaming-tier seems like a "Baby Blackwell" , as it is way out of "this ballpark" with its around 10 000 cores only, even on a lower N4 fab. Still not free for sure.... because 1000 cowboy coins is surely not enough... it still needs way more.
Of course if someone wants to get a 5070 using... uhm... 6144 cores and 12 GB VRAM... then go for it but i would feel betrayed buying "such a crap". It seems to be made for those "eating to much breadcrumbs" and always thanking for it.
You could as well go to Steam, paying lot of bucklets, asking them to provide some DRM locked data you never own... provided by a "33792 core Blackwell" while your own Blackwell is 6144 cores with 12 GB VRAM and please do not forget to "give the game back, as soon as played or account banned... no matter which comes first".
Regarding PS5 Pro: I do not know exactly the node used there but it is most likely the most advanced node, so N4P (same as the RDNA 4 node), else it would be difficult for Sony able to provide this performance with not much more than 200 W TDP.
Sure, in theory N3P is even more advanced, but it is currently not used on non portable GPUs because of some complicated issues with a very high transistor count. GPUs are already expensive enough and it would add even more costs and some other issues (cooling, higher failure rate at higher transistor count).
AMD has been a very loyal customer of TSMC (for Nvidia it was less stable in the past...), so they certainly are a "premiere customer" with good offers provided to them.
Post edited April 08, 2025 by Xeshra