2025 Professional GPU Engineering Roundup

Introduction

So far in 2025, we have seen the launch of over twenty new desktop GPUs. Although we covered most of them in dedicated reviews, one major area we have been missing is NVIDIA’s new RTX PRO™ Blackwell series—apart from our initial 6000 Blackwell Workstation Edition review. These cards have been in high demand, making it difficult to obtain enough of the lineup to justify a full article. However, we now have nearly every variant of RTX PRO Blackwell in hand! Armed with them, we feel the best approach is to test as many GPUs as possible in as many workflows as possible.

For this article, we tested every RTX PRO Blackwell GPU we had, as well as nearly the entire stack of Ada Generation cards. We also included some older or non-NVIDIA cards like the NVIDIA RTX™ A6000, the AMD Radeon™ AI PRO R9700, most of the Radeon™ PRO W7000 series, and even the Intel® Arc™ Pro B50. Our primary focus is on comparing NVIDIA’s current and last-gen cards, but we’re also interested in seeing where other brands can compete.

This article will only cover our recently added architecture, engineering, and construction (AEC) benchmarks, which include CAD, BIM, and photogrammetry software. We also have a companion article looking at the performance of the same GPUs in content creation workflows. Unfortunately, we will not be examining multi-GPU setups in this article.

There aren’t many other reviews available for professional GPUs, especially the more recently released NVIDIA models. However, we do recommend StorageReview’s coverage of the 6000 Blackwell Workstation, especially if you are interested in more in-depth AI performance data. igor’sLAB also released an AMD Radeon AI R9700 review, which includes data for some previous-gen NVIDIA cards.

Below, we have listed specifications for several current and recent professional GPUs, including a few that we were not able to include in this round of testing. For more information, visit the workstation pages for NVIDIA, AMD, or Intel.

In line with most of our other recent GPU reviews, we performed all the testing for this article on an AMD Ryzen™ 9 9950X3D-based system. The 9950X3D is one of the all-around fastest CPUs available, ensuring that we minimize processor bottlenecks as much as possible. However, we do have a somewhat unique methodology that is worth mentioning. As part of our philosophy of focusing on professional workflows, we limited as many stability-affecting system tweaks as possible. This means PBO and ASUS’s MCE were disabled in the BIOS, the RAM was run at the maximum officially supported JEDEC speeds, and settings like VBS were enabled in Windows.

As far as possible, all the apps, drivers, BIOSs, and benchmarks were on their latest versions. We didn’t reuse past results, and we updated our testing platform prior to this round of tests. The applications we tested with were RFO for Revit, InvMark for Inventor, SPECapc for SOLIDWORKS, and our in-house PIX4Dmatic benchmark. Links with more information about those are available in the expandable section above.

Revit

As an application, Revit is primarily CPU-dependent. The RFO benchmark produces a variety of scores, and we have pulled out three of them that are impacted by the GPU: total model creation time, total graphics time, and a composite total export time. Model creation is the most important of these for the majority of Revit users; however, the graphics and export times can significantly impact the user experience. We have not included the rendering time, as we don’t believe it is relevant to most users and it scales poorly. We’ve also only used the “Standard” preset here, although we are exploring using the “Graphics_Expanded” preset for future GPU reviews.

In model creation (Chart #1), we found that different GPU families affected performance, but there was minimal difference between models within a given family. We understand model creation to be entirely CPU-based, so this makes sense. The family-dependent performance is interesting and may represent additional overhead, depending on the driver and GPU being used, or pipelines that are GPU-dependent. RFO showed the best performance with AMD GPUs, followed by Intel’s B50. NVIDIA’s Blackwell family was just behind the B50, with another small performance drop-off for the older Ada and Ampere GPUs. None of those differences were huge, but it was interesting to see the W7500 outperform a 6000 Blackwell.

Moving on to the graphics tests (Chart #2), you can see why we are interested in exploring the “Graphics_Extended” benchmark preset. There was almost no discrimination between GPUs until the very bottom of the performance stack. All of NVIDIA’s cards were within the margin of error, as were the W7900 and W7800. AMD’s R9700, W7600, and W7500 were slower, but probably not by enough to outweigh potential cost savings or performance in other domains. Intel’s B50 did fall to the bottom of the pack, although it is also the cheapest.

We weren’t going to include export time (Chart #3) until we saw that there appeared to be an interesting architecture-based story to tell, much like with model creation. We found that AMD’s RDNA3 and 4 GPUs were nearly 10% faster than NVIDIA’s Blackwell-based cards. Ada and Ampere cards from NVIDIA were slightly slower yet, while Intel’s B50 was far behind all of them. We’re not sure this makes a massive difference to most Revit users, but it is an interesting finding, likely related to what we observed with model creation.

Inventor

We continued to encounter issues with drawing scores in InvMark, as was the case with our last two articles, making the overall score largely unusable. Although we haven’t been able to pin down the cause of the issue, one commenter on a previous article suggested that there may be a performance bug in Windows 11 affecting this application, and specifically, drawing performance. We hope to look into that in the near future. As it stands now, though, we only have one score for this review: the graphics score.

We found that even the graphics tests in Inventor are relatively insensitive to GPUs once a certain threshold is reached. Although our B50 review saw good differentiation, that was only with entry-level GPUs. Here, once above the AMD Radeon PRO W7500, all of the GPUs perform basically the same.

We did still gather some interesting takeaways, though. First, AMD’s GPUs perform generally well. Although we would typically consider most of these results to be within the margin of error – about 5% – the clustering of the averages for all the non-W7500 AMD GPUs around the same score at the top of the chart does suggest they may have a minor performance advantage against NVIDIA. Regardless, there were no bad GPUs among those we tested for Inventor, so the GPU that is the best value or otherwise facilitates workflows is the one that should be used.

SOLIDWORKS

SPECapc for SOLIDWORKS reports more GPU scores than most of the other engineering benchmarks, but we can use the “composite” score as an overall indicator of performance. Although this tool also tests specific CPU-based workflows, we have only reported the GPU scores here. We also want to quickly note that our testing was at 4K with 150% scaling, which is not aligned with SPEC’s “official” results nor that of many other reviewers, but all of our testing is performed at this resolution.

Starting with the composite score (Chart #1), we found that AMD leads the pack again, with the W7900 and R9700 occupying the top two slots. They are only marginally faster than most of the Blackwell family, and the top Ada cards, but this is an excellent showing from a much cheaper set of GPUs! We saw only minor generational improvements for the NVIDIA cards at the high end, but more “midrange” cards like the 2000, 4000, and 4500 saw performance uplifts of 12-21% when moving from Ada to Blackwell. Despite its great performance compared to other entry-level professional GPUs, the B50 struggles to keep up with the mid-range and high-end models we tested here.

Moving on to the Shaded scores (Charts #2 and #3), we saw very similar patterns. AMD’s Radeon AI PRO R9700 topped the charts this time, with a notable lead of around 7% over the 5000 and 6000 Blackwell. While not a huge difference, the R9700 is substantially cheaper than those. The last-gen W7000 series cards are less competitive, but given NVIDIA made only very small improvements gen-on-gen, this doesn’t matter too much.

In the RealView test (Chart #4), it was instead the Radeon PRO W7900 which offered the best performance, 8% faster than the 5000 Blackwell and 10% faster than the 6000 Ada. The AI PRO R9700 sits between the 6000 Blackwell and the 4000 Blackwell.

Finally, in the drawing tests (Chart #6), AMD completely outperformed NVIDIA, with every tested AMD GPU beating the fastest NVIDIA model by at least 18%. For both AMD and NVIDIA, we see essentially no differentiation between GPUs within their respective lineups. Intel’s B50 also put up a surprisingly good showing, only slightly behind NVIDIA.

Overall, these results make it a bit difficult to recommend any particular GPU. AMD seems to be generally dominant, but whether a W7000 or AI PRO GPU is better depends on the workflow. However, it is clear that NVIDIA – even with their new Blackwell cards – doesn’t offer the best value in SOLIDWORKS.

PIX4Dmatic

PIX4Dmatic is an application we only started testing last month. At that time, we promised future and ongoing testing to try and fill out a performance picture. Although there is still more to be done, we are excited to have it as part of our AEC benchmarking suite. However, one caveat for this article is that, due to time constraints, we were only able to test with our smaller dataset. We previously found scaling to be relatively consistent between the small and medium sets, but it’s still worth noting. Additionally, PIX4Dmatic only supports NVIDIA GPUs.

The first chart looks at the overall processing time for the small image set. We found that, in terms of a complete pipeline, the impact of processing time from GPUs is relatively minor. Blackwell cards were generally faster than their last-gen counterparts, but only by a few percent, with one exception: the 2000 Blackwell was noticeably faster than the 2000 Ada. However, while useful, the overall score is influenced by a variety of sections that are not, or only weakly, influenced by GPU performance.

Moving on to one of the more GPU-sensitive portions, in calibration (Chart #2) we see much more scaling with theoretical GPU performance. The 6000 Blackwell models lead the chart – though interestingly, we found the Max-Q variant slightly faster than the 600 W Workstation card – followed by the 5000 Blackwell. Almost any of the adjacent bars are interchangeable, as they fall within each other’s confidence intervals. However, an overall trend is present: generational improvements average about 5%, though they fall off at the higher end, and the 2000 Blackwell an outlier at 13% faster than the 2000 Ada.

Our third chart looks at the time taken to generate a dense point cloud, and shows very similar trends. The Blackwell cards are faster than their Ada counterparts by a small margin, and higher-end cards are faster than lower-end cards, but the overall difference isn’t huge. Additionally, since all the cards tested (being professional models) have large VRAM buffers, the memory capacity is likely not a differentiator.

Conclusion

As we continue to flesh out our engineering tests, it is fascinating to see how support, relative performance, and value propositions for professional workstation GPUs differ from the consumer side. While there are still some applications which only work with NVIDIA GPUs, the guarantees of software support, ISV certification, and driver validation help ensure that more applications work well across the whole range of GPU architectures. We think, much like the consumer space, the professional workstation market is becoming increasingly competitive.

Overall, we were surprised to find that AMD’s professional GPUs offer the best performance in AEC workflows, outside of programs which require NVIDIA cards. Whether the Radeon PRO W7900 or Radeon AI PRO R9700 was the top GPU depended on the exact workflow within an application, but both offered outstanding performance in the applications we tested. NVIDIA’s Blackwell GPUs were solid, but struggled to differentiate themselves from their last-gen, Ada-based counterparts. On the low end, Intel lagged, but possibly not by enough to offset the B50’s low price tag for value-conscious users.