AMD Radeon AI R9700 aka RX 9070XT in the big workstation graphics card test – What are the benefits of double the memory when working?

1 - Introduction and technical data 2 - Test system and equipment 3 - Autodesk AutoCAD 4 - Autodesk Inventor Pro 5 - PTC Creo 6 - Dassault Systèmes Solidworks 7 - Autodesk Maya 8 - SPECviewperf 15 (2025) 9 - Adobe Photoshop 26.10 10 - Adobe After Effects 2025 11 - Adobe Premiere Pro 25.41 12 - KI Benchmarks (AI Vision, Image, Text) 13 - Rendering 14 - Temperatures, clock rate, fans, noise and power draw 15 - Summary and conclusion

PugetBench for Premiere Pro is a practical benchmark tool that measures the performance of workstations in typical Adobe Premiere Pro workflows. The aim is to record the actual performance during import, editing, effects and export and thus create a basis for comparing different hardware configurations. Of particular relevance here is the mapping of GPU-accelerated effects and the measurement of encoding and decoding performance, as these often represent the biggest bottleneck in video production.

A significant part of the benchmark focuses on export performance, as this is crucial for productivity in practice. In order to emphasize the influence of encoding as much as possible, Puget Systems relies on a special setup: The source material is a DNxHR LB clip with 480×270 pixels, which is scaled on a UHD timeline (3840×2160). DNxHR LB is a very simple codec for processing; at the same time, there is no platform with hardware decoding support for this codec. This ensures that decoding is not the limiting factor, but that encoding is the bottleneck during export.

To prevent modern codecs such as H.264 or HEVC from becoming “too smart” due to their efficient compression algorithms with repeating pixel patterns and thus unintentionally achieving better results, a fractal noise image in PNG format is also superimposed on the clip. It is important to note that this is not the fractal noise effect from Premiere itself, but a prepared image that is applied to the clip. This ensures that each pixel is unique and that there are no redundancies in the encoding.

With this setup, the benchmark can very clearly capture the encoding performance of Premiere Pro in various codecs such as H.264, HEVC or DNxHR. In addition, it tests GPU-accelerated effects so that differences in real-time preview and rendering are also visible. The results provide a direct comparison of which hardware is best suited to which workflow. This makes the benchmark a valuable tool both for users who want to optimize their systems and for IT departments that need to select hardware for creative teams.

To evaluate the performance of a system when exporting to different codecs, the benchmark is designed to make the encoding part of a rendering as much of a bottleneck as possible. To achieve this, one uses a DNxHR LB 480×270 source clip scaled to a UHD (3840×2160) timeline. DNxHR LB is a very simple codec to process (especially at this low resolution) and does not support hardware decoding on any platform, making it a good base codec. To ensure that each pixel is unique and codecs like H.264 or HEVC don’t get too “smart”, one also applies a simple fractal noise image in PNG format (NOT the fractal noise effect) to the clip to ensure there are no repeated pixels.

• H.264 50Mbps 8-bit UHD, H.264 50Mbps 8-bit (software encoding) UHD
• HEVC 50Mbps 8-bit UHD, HEVC 60Mbps 10-bit UHD, HEVC 50Mbps 8-bit (software encoding) UHD, HEVC 60Mbps 10-bit (software encoding) UHD
• DNxHR LB UHD, DNxHR SQ UHD,,DNxHR HQX UHD
• ProRes 422 Proxy UHD, ProRes 422HQ UHD, ProRes 4444 UHD.

For the processing tests, the benchmark uses a wide range of codecs in different resolutions and creates a timeline in the native resolution of the clip. It then exports each timeline to DNxHR LB in HD (1920×1080) resolution. This is where the hardware decoding support comes into play. The benchmark results from this method are very consistent and accurately reflect the relative performance between systems when working with different codecs.

• 4K H.264 150Mbps 4:2:0 8-bit, 4K HEVC 100Mbps 4:2:2 10-bit, 8K HEVC 100Mbps 4:2:0 8-bit,
• 4K ProRes 422 Proxy, 4K ProRes 422
• 4K DNxHR LB, 4K DNxHR SQ
• 4K Cinema RAW Light ST
• 4K ARRIRAW
• 5K Sony X-OCN
• 4K RED, 8K RED

The final category of tests looks at performance for GPU-accelerated effects. Many of the effects in Premiere Pro are relatively easy to process individually, so even a low-end GPU will not be taxed if only a single instance is applied. To address this, the benchmark applies each effect four to forty times, depending on the requirements. Again, a DNxHR LB UHD (3840×2160) clip is used as the basis and exported to DNxHR LB HD (1920×1080) – just like the “4K DNxHR LB” processing test. The difference is that the following effects are also applied for each test:

• Lumetri Color x40
• Gaussian Blur x40
• Sharpen x40,
• VR Digital Glitch x20, VR De-Noise x4.

Overall Score

With 140590 points, the Radeon AI R9700 is positioned just behind the RX 9070XT, which takes the top position with 145362 points. Both cards clearly dominate the field and clearly set themselves apart from the professional Radeon models and NVIDIA. The fact that the RX 9070XT performs slightly better here is not a contradiction, as Premiere Pro contains workloads that react more strongly to clock frequencies and individual acceleration paths than to large memory configurations. The Radeon Pro W7800 follows at a distance with 112221 points, but remains competitive within the scope of its driver architecture. NVIDIA can only keep up to some extent with the RTX 4000 Ada, but is around 39000 points behind the R9700. Premiere Pro is heavily GPU-optimized and rewards those architectures that offer high single-pipeline efficiency and good media engine integration, giving AMD a visible advantage in this generation.

LongGOP Score

With LongGOP material, which is a considerable challenge for the GPU because it relies on efficient intermediate image calculation, the RX 9070XT achieves the clear best value with 198 points. The Radeon AI R9700 follows with 176 points and is therefore well ahead of the professional Radeon models. The RX 9070XT’s lead is due to slightly higher clock rates, which have a greater effect on LongGOP decoding and effect evaluation than the R9700’s larger amount of memory. The differences within the Pro models remain moderate and reflect their more conservative optimization. NVIDIA clearly falls behind in this test scenario, which suggests a less aggressive acceleration of certain GPU-supported video paths.

Intraframe Score

The weighting shifts in the intraframe test. The Radeon Pro models W7700 and W7600 lead with 159 and 152 points, while the Radeon AI R9700 and RX 9070XT follow at the same level with 132 points each. This can be attributed to the fact that intraframe workloads benefit less from high raw performance and more from deterministic drivers such as those offered by the Pro series. The R9700 and RX 9070XT are nevertheless close together and show that Premiere Pro also scales well with the RDNA 3.5 architecture in the intraframe path. NVIDIA once again remains in the midfield, which confirms the general trend of the benchmarks.

Raw Score

In the RAW workloads, the Radeon AI R9700 and the RX 9070XT deliver a practically identical result of 181 points. These tasks benefit particularly strongly from high memory bandwidth and efficient compute acceleration, allowing the R9700 to fully exploit its advantage over models with lower bandwidth. The fact that the RX 9070XT is still exactly on par shows that Premiere Pro works heavily on GPU compute in this area and does not focus on pure memory throughput in isolation. The Radeon Pro W7800 achieves 124 points and thus remains far behind the two consumer cards, while NVIDIA is once again unable to provide an equivalent path.

GPU Effects Score

GPU-intensive effects such as transitions, color transformations, blurs and motion effects show the largest gap. The RX 9070XT achieved the top score of 94 points, closely followed by the Radeon AI R9700 with 93.3 points. This proves that Premiere Pro can utilize the architecture of both cards almost identically and only minor clock differences account for the minimal lead. The Radeon Pro W7800 follows with 78.1 points, but benefits less from professional drivers, as Premiere Pro makes greater use of the consumer shader paths. NVIDIA falls significantly behind with 48.7 points (RTX 4000 Ada), which clearly confirms the trend of the GPU effect evaluation.

Interim conclusion

Premiere Pro benefits exceptionally from AMD’s current GPU architecture. The Radeon AI R9700 and the RX 9070XT lead almost every sub-benchmark and distance themselves from both NVIDIA and the professional Radeon models by a clear margin. Differences between the R9700 and RX 9070XT remain small and only become apparent in specific scenarios such as LongGOP decoding or effect pipelines. The Pro models can only score points in highly deterministic tasks, while NVIDIA clearly falls behind overall. For users in the editing sector who rely on high efficiency in GPU effects, RAW workflows and modern codecs, the R9700 is one of the most powerful options on the current market and offers consistent performance throughout.