We have read actively and passively according to leakers that NVIDIA and AMD will compete with the same weapons in the new generation of graphics cards. That the fight will be over other key aspects and therefore the better architecture will win, but… It seems that it won’t, since NVIDIA and TSMC they have a secret deal that we’ve already seen in Hopper that will naturally extend to the RTX40 under the Ada Lovelace architecture: the 4N knot.
The 5nm has several high-volume candidates and TSMC’s problems launching it on its high-performance node (Apple consumes the so-called Low Power N5 LP) don’t seem to end completely despite rumors that all is well. In addition to this fact that has AMD and Apple as protagonists, we must add a key fact from the past that will make us understand what has been revealed: payment in advance.
NVIDIA and TSMC have had an agreement since last year
We will try to explain what is happening and not confuse anyone with this, because it is easy to fall into error. When NVIDIA and TSMC signed an exclusivity deal for edge chips late last year, it was thought to be a matter of reserving production of the N5 lithographic process like AMD or Apple had done.
The reservation was true, the knot was wrong, and here the controversy begins. NVIDIA did not want to compete with AMD on the same node because its Ada Lovelace architecture was not MCM like that of its rival (we will come back to this later) and therefore asked TSMC for something better and above all more precise.
Notice I didn’t mention PCIe 5.0 in my Lovelace slides: https://t.co/n7sdIL1Uku
Speaking of “4”, #NVIDIALovelace is indeed 4nm!
🤠🎉😋 More to say about the next Broken Silicon… https://t.co/jjJ1GSod0j
— Moore’s Law is dead (@mooreslawisdead) April 24, 2022
From this need arose a secret agreement whereby TSMC would manufacture in large volume an exclusive node for NVIDIA called 4N, but beware, as they say it is easy to make a mistake because this 4N node is not the N4 which will succeed the N5 that AMD uses. The NVIDIA 4N node is something specific and designed under CoWoS 2.5D with one thing in mind: the best power efficiency without compromising performance and also a small improvement in transistor density if we compare this node fairly with the N5 that AMD will be using.
NVIDIA RTX 40 4N vs AMD RX 7000 5N + 6N: TSMC as cornerstone
This is the ultimate battleground that the two companies will use to go against each other. NVIDIA will take this 4N which was already featured in Hopper graphics cards and which we thought was exclusive to them and implement it in any GPU of this new series.
Instead, AMD will opt for an MCM approach that will host two or more chips (depending on the GPU model) and have the chips as the main attraction. GCD (Graphic Compute Die) and MCD (Memory Compute Die) in two different packaging and substrates, where the first will be manufactured in the above 5nm as the seconds come 6nm.
Therefore, the war is centered between the best available node for a graphics card against an architecture that divides resources and chips to manage efficiency and power in a manner similar to what Ryzen does against current Intel processors.
A secret agreement between TSMC and NVIDIA that will mark the future of the next generation of graphics cards for games and data centers, AI and DL.