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

AutoCAD has been one of the most important and most widely used CAD programs in the professional environment for decades. It serves as a reference application for architecture, mechanical engineering, electrical engineering and construction planning, whereby performance under real conditions depends heavily on GPU driver optimization and hardware architecture. Since AutoCAD is used as standard software in many engineering offices, planning agencies and educational institutions, the Cadalyst benchmark has established itself as a recognized test to objectively evaluate the performance of workstation graphics cards. The Cadalyst test covers different workloads that simulate real-world 2D and 3D design tasks, making it an accurate tool for assessing GPU suitability for professional CAD environments.

Total Score

In the overall picture of the AutoCAD benchmark, the Radeon AI R9700 is clearly at the top of the mixed 2D and 3D workloads. With a score of 1318 points, it is visibly ahead of the Radeon RX 9070XT, which only just falls behind with 1280 points. This close difference confirms the technical proximity of both cards, which is reflected in the identical GPU configuration. Nevertheless, it is clear that the larger memory of the R9700 offers a measurable advantage with more complex CAD loads, as AutoCAD increasingly benefits from larger VRAM in higher resolutions and with extensive projects. The older Radeon Pro W7800 achieves 1224 points and is thus clearly behind the two RDNA-4 models. The result makes it clear that the new architecture not only has a positive effect on specialized workstation functions, but also on general CAD performance. In terms of efficiency per watt and per shader, the R9700 thus offers a coherent overall picture that sets it apart from both the RX 9070XT consumer variant and the RDNA 3 workstation generation.

3D performance

In terms of pure 3D performance, the R9700 once again performs very strongly and achieves 3113 points, which puts it practically on a par with the NVIDIA RTX 2000 Ada. Compared to the RX 9070XT, there is a small but understandable lead, as it lands at 2965 points. The difference can again be explained by the memory expansion, as 3D representations in AutoCAD generate data structures that are partially VRAM-bound and therefore benefit from higher reserves. The gap to the Radeon Pro W7800, which achieved 2814 points, is slightly larger than in the overall test. This is plausible because RDNA 4 offers improvements in geometry processing and cache efficiency, while the W7800 is still based on the heavy MCM implementation of RDNA 3. For workloads that contain many individual objects or repeated transformation calculations, this results in a visibly more stable frame rate and shorter redraw times in practice.

In this test field, the R9700 thus demonstrates a 3D performance that is clearly suitable for workstations and easily holds its own in the professional application environment.

2D performance

In the 2D component of the benchmark, the R9700 and the RX 9070XT dominate the test almost equally. The scores of 1038 to 1035 points are practically identical, which confirms the architectural identity of the cards. AutoCAD’s 2D rendering paths hardly require any GPU-internal computing power and scale more strongly with the clock behavior, memory latency and driver optimization. The memory advantage of the R9700 does not apply here, as 2D drawings generally only generate small amounts of data. The W7800 follows with 981 points and thus lags slightly behind, which can be explained by the lower clock rate of the RDNA-3 workstation cards.

The results show that AMD has been offering stable performance in classic 2D workloads for years and that there are only minor differences between the generations. For drawing and layout tasks, this means that both RDNA-4 models perform above average and work practically without bottlenecks. The R9700 does not utilize its extra memory here, but still offers the best single-user performance in the test field.

Disk performance

The disk performance is only indirectly attributable to the GPU, but nevertheless shows how efficiently the driver supports data movements and caching operations. It is interesting to note that the Radeon Pro W7700 leads here with 552 points, while the R9700 and the RX 9070XT achieve exactly 541 points. This confirms that the underlying architecture does not hinder data processing and that both RDNA-4 models provide the same performance level. The W7800 is just below this with 519 points, which is due to the more conservative driver paths of the professional RDNA-3 generation. For the user, this means that the R9700 can be used in data-intensive workflows without any problems despite its classification as a creator model, without falling behind classic workstation cards. However, the tie with the RX 9070XT shows that the memory expansion does not play a role in this area.

CPU performance

The CPU part of the benchmark naturally hardly reacts to the GPU used. Nevertheless, it shows how the GPU drivers can relieve the CPU or generate additional overheads. The R9700 achieves 579 points and is thus on a par with the Radeon Pro W7800 and W7500 as well as the RTX A1000 and A2000 Ada. The gap to the RX 9070XT with 541 points is noticeable, even if GPUs only have a secondary influence here. This difference is most likely due to the lower driver overheads of the R9700, which may use more optimized paths for CAD workloads due to its workstation proximity. The RX 9070XT, on the other hand, is listed more as a gaming card, which puts the focus more on gaming workloads. The W7800 remains in the upper midfield, although its slightly lower score compared to the R9700 may be an indication of its older architecture. Overall, the picture of consistent and cleanly optimized drivers for the R9700, which does not generate any unnecessary additional loads despite its high raw performance, is also confirmed here.

Interim conclusion

The results from the AutoCAD benchmarks so far show very clearly that the Radeon AI R9700 represents a consistently optimized further development of the familiar RDNA 4 foundation and is clearly positioned against both the almost identically constructed RX 9070XT and the Radeon Pro W7800 from the previous workstation generation. The card is ahead in all GPU-relevant sub-areas, whereby its strengths are particularly evident in mixed 2D/3D workloads and in pure 3D rendering. Although the larger VRAM does not have an effect in every scenario in AutoCAD, it does play a role where more complex models or extensive textures have to be stored in memory.

Even though the RX 9070XT comes close in many disciplines, the R9700 shows an overall more stable and professional performance, which suggests more optimized driver paths and lower overheads. The Radeon Pro W7800 often lags behind despite its high computing power, which can be explained by both the architecture-related limitations and the lower clock frequency. The R9700 demonstrates here that a more modern architecture and a more efficient cache structure can achieve more in practice than a nominally larger range of functions of the previous generation.

This gives the impression of a very balanced and powerful CAD graphics card, which is clearly positioned above the RX 9070XT and at the same time closes the gap to the classic workstation line in several areas. The results so far suggest that the R9700 can not only keep up in many professional workflows, but even offers advantages where memory size, driver optimization and orderly computing load are required.