AMD Radeon RX 9070 vs NVIDIA GeForce RTX 3070 Laptop GPU

The AMD Radeon RX 9070 is a high-end graphics card based on AMD's Navi 48 (N48 XT) that features 56 CUs (28 WGP). The model will feature 16 GB of GDDR6 VRAM with a 256-bit bus and 64 MB of total cache.

Performance

Our tests show that the card is 13% faster overall than the RTX 4080 in games, and 3% slower than the RTX 4080 Super. The higher-level RTX 9070 XT is 12% faster. 

Power consumption

The Radeon RX 9070 requires 220 watts of energy, but there are also custom versions from AiB-partners with 240-watt TDP.

The Nvidia GeForce RTX 3070 Mobile (for laptops, GN20-E5, Max-P, ID 249c / 249d) is the second fastest Ampere graphics card for notebooks in early 2021. It is based on the GA104-770-A1 Ampere chip and offers 8 GB GDDR6 graphics memory with a 256 Bit memory bus. It supports PCIe 4.0 and will be available in two different variants from 80 to 125 Watt (TGP). As package the notebook graphic card uses GB5-256 and the chip size is 40x40mm. The GPU supports eDP 1.4b to connect the internal monitor.

There is no more Max-Q variant (formerly used for the low power variants) but every OEM can choose to implement Max-Q technologies (Dynamic Boost, Whispermode).

The performance depends on the TGP setting and cooling of the laptop. On average it should be similar to the old RTX 2080 Super Mobile and therefore best suited for QHD - 4k gaming. The similar named RTX 3070 desktop card is significantly faster.

The GA104 chip offers 6,144 FP32 ALUs of which half can also execute INT32 instructions (i.e. 3,072 INT32 ALUs). The RTX 3070 however wont use all ALUs on the chip. With Turing all shaders could still execute FP32 or INT32 instructions. The raytracing and tensor cores on the chip were also improved according to Nvidia. The RTX 3070 mobile series uses 5,120 of the 6,144 cores. The Ampere chips also include an improved 5th generation video encoder (NVENC for H.264 and H.265) and a 7th generation decoder (for various formats now including AV1).

The GA104 chip is manufactured by Samsung in 8nm (8N), which is not quite able to keep up with the 7nm node at TSMC (e.g. used by AMD and also for the professional GA100 Ampere chip).