AMD launched its new generation RDNA 4 RX 9070 series graphics cards in February, with both cards featuring the Navi 48 chip.
Following this, the RX 9060 series, based on the Navi 44 chip, is expected to be released in the second quarter of 2025. To bridge the performance gap between these two series, AMD has prepared the RX 9070 GRE to fill the product line.
According to the latest information obtained by VideoCardz, the RX 9070 GRE will be equipped with the Navi 48 XL chip, featuring 48 Compute Units (CUs), which translates to 3072 stream processors. This indicates that roughly a quarter of the GPU core has been disabled.
Its boost clock speed reaches 2.79 GHz, which is lower than the RX 9070 XT’s 2.97 GHz but higher than the RX 9070’s 2.52 GHz.
The RX 9070 GRE will be paired with 12GB of GDDR6 memory, the same type used in the earlier RX 9070 XT and RX 9070. However, the memory speed will be reduced from 20 Gbps to 18 Gbps.
This means that with a 192-bit memory bus, the RX 9070 GRE’s memory bandwidth will be approximately one-third lower than that of the RX 9070 XT and RX 9070, dropping to 432 GB/s.
In its default state, the peak FP32 compute performance of the RX 9070 GRE can reach about 95% of the RX 9070. Furthermore, users can further enhance performance through memory overclocking.
AMD is expected to officially announce this graphics card in the coming weeks. It is currently unclear whether it will be launched exclusively in the Chinese market or will be available globally.
With the exposure of the RX 9070 GRE’s specifications, foreign media outlets have also made performance predictions.
The preliminary results suggest that the RX 9070 GRE’s ray tracing performance is comparable to NVIDIA’s RTX 5060 Ti 16GB, but in the Steel Nomad benchmark, the RX 9070 GRE’s performance is about 34% higher.
This implies that if the pricing of the AMD RX 9070 GRE is slightly higher than the RTX 5060 Ti 16GB, this card will be extremely competitive in the mid-range market.
As a gaming card, AI performance is also crucial nowadays.
NVIDIA’s strength in AI lies not only in its excellent hardware performance but also in its ecosystem optimization and adaptation, allowing it to achieve optimal efficiency. AMD lags behind in this regard but is actively catching up.
Recently, Stability AI released an ONNX framework-optimized Stable Diffusion model, which can achieve speedups of up to 3.8 times on compatible AMD Radeon graphics cards and Ryzen integrated graphics.
AMD stated that over the past year, it has been collaborating with OSV, OEM, and ISV partners to optimize AI applications on AMD hardware, including hardware design, drivers, compilers, and machine learning models.
According to official data, after optimization, the inference performance of SDXL 1.0 and SDXL Turbo on AMD hardware has increased by up to 3.8 times compared to the basic PyTorch model, while SD3.5 can see an increase of up to 2.6 times.
In real-world image and video generation applications, the efficiency of SDXL 1.0, SDXL Turbo, SD3.5 Large, and SD3.5 Large Turbo on the RX 9070 XT graphics card has improved by 3.1 times, 1.5 times, 3.3 times, and 2.1 times, respectively.
The efficiency of SDXL 1.0, SD3.5 Large, and SD3.5 Large Turbo on the Ryzen AI MAX+ 395 has increased by 1.4 times, 3.3 times, and 3.0 times, respectively.
The AMD hardware-optimized version of the Stable Diffusion model has a special “_amdgpu” suffix (for example, stable-diffusion-3.5-large_amdgpu) and is now available for download on Stability AI or Hugging Face.
You can also use the latest version of Amuse 3.0, jointly developed by AMD and TensorStack AI.
Note that the AMD GPU driver version should not be lower than the 24.30.31.05 preview version, or you can wait for the upcoming 25.4.1 official version.
