New generations from AMD and NVIDIA explode the number of transistors per square millimeter.
On January 30, NVIDIA launched its GeForce RTX 50 series graphics cards, built around Blackwell architecture GPUs etched using the N4 process of the Taiwan Semiconductor Manufacturing Company (or TSMC). Since 2021, TSMC has been the world's largest independent semiconductor foundry, with impressive technological leverage enabling it to sign contracts with the world's leading companies. NVIDIA is taking advantage of the N4 to shatter the record number of transistors integrated into a GPU with the GB202 of the RTX 5090, which integrates a whopping 92.2 billion transistors.
The GB203 has "only" 45.6 billion and the GB205 "only" 31 billion, but the chips are considerably smaller than the RTX 5090 monster: we're talking 750 mm², 378 mm² and 263 mm² respectively for the GB202, GB203 and GB205. NVIDIA thus seems to be leading the way in the GPU world, all the more so as AMD has made it clear that its new generation of RDNA 4 architecture is not intended for the high-end market. Instead, AMD is aiming for the entry- and mid-range markets, where most people are concerned.
Transistor count and density of the Blackwell, RDNA 4 and RDNA 3 © VideoCardz
However, AMD is not sitting idly by, and the American group is also benefiting from TSMC's technological advances. The N4 process is also used on RDNA 4 GPUs, and although the Navi 48, AMD's future flagship, has "only" 53.9 billion transistors, this is almost as many as on the Navi 31 (57.8 billion), the top-of-the-range chip of the RDNA 3 generation. But to achieve such a total, the Navi 31 had to occupy a surface area of 529 mm², whereas the Navi 48 occupies 350 mm². The latter therefore boasts a density of 154 billion transistors per square millimeter, compared with 109.3 MTr/mm² for Navi 31.
In a nod to NVIDIA, which will be quiet on the high-end segment, AMD can even boast a higher transistor density than Blackwell chips. The 154 MTr/mm² are well above the 117.9 MTr/mm² of the GB205, the 120.6 MTr/mm² of the GB203 and even higher than the 122.9 MTr/mm² of the GB202. A small victory for AMD, a triumph for TSMC!
