[The Exiled]

AMD Radeon RX 9070 Series Technical Deep Dive

Today AMD is announcing only the top Radeon RX 9070 series from this generation. The two aim to provide enthusiast-class gaming experience in the performance segment at prices that could disrupt NVIDIA's performance-segment offerings from the RTX 5070 series. Historically, the xx70 series have sold in large volumes, and AMD itself has had a track record of unsettling NVIDIA in this segment with the original RDNA RX 5700 series versus the RTX 20-series, and so the company could be on to something. The series consists of just two SKUs for now, the top Radeon RX 9070 XT, and the next-best Radeon RX 9070, both are performance-segment products based on a common silicon, the 4 nm Navi 48.

The Navi 48 silicon is technically a performance-segment chip, logically its predecessors would have to be the Navi 32, the Navi 22, and the Navi 10, but there's more to this story because AMD changed the nomenclature of Radeon RX with this generation. The Navi 48 silicon physically has 64 compute units, and a 256-bit wide GDDR6 memory bus driving 16 GB of memory on both cards. The RX 9070 XT maxes it out, enabling all 64 CU (4,096 stream processors), and the highest possible clock speeds, while the RX 9070 gets 56 CU (3,584 stream processors). The RX 9070 XT also comes with 128 AI accelerators, and 64 RT accelerators, besides 512 TMUs and 128 ROPs. The RX 9070, on the other hand, offers 112 AI accelerators, 56 RT accelerators, 448 TMUs, and 112 ROPs. The RX 9070 XT boosts up to 2.97 GHz, while the RX 9070 goes up to 2.52 GHz.

FSR 4 makes its debut as a Radeon RX 9000 series exclusive technology, with up to 30 game titles expected at launch and at least 75 titles in 2025. The biggest change with FSR 4 is a new AI ML-based super resolution component that vastly improves image quality at a given performance preset. FSR 4 now more closely resembles DLSS in the way it works. AMD now has its own FSR 4 AI data-center made up of AMD Instinct AI GPUs, on which the company trains custom, game-specific ML models, which the FSR 4 upscaling algorithm uses to reconstruct details in upscaled frames. On the client side, the ML-based upscaler needs an AI acceleration throughput of 779 AI TOPS, which the RX 9070 series more than sufficiently meet (around 1200 TOPS for the RX 9070, and around 1500 TOPS for the RX 9070 XT).