Table of Contents
Introduction
At GTC this year, NVIDIA announced its next generation of professional graphics cards: NVIDIA RTX PRO™ Blackwell. Based on NVIDIA’s all-new Blackwell architecture, currently powering NVIDIA GeForce RTX™ 50-series GPUs, it promises strong performance improvements over the last-gen Ada Lovelace-based cards. Although announced about two months ago, the rollout of this family of GPUs has been slow, which is traditional for professional-class cards. In this article, we will be reviewing the flagship model of this product family: the NVIDIA RTX PRO™ 6000 Blackwell Workstation Edition.
Image
Uniquely, this generation, NVIDIA is releasing three versions of its top-end card. The NVIDIA RTX PRO™ 6000 Blackwell Workstation Edition places NVIDIA’s GB202 GPU in a housing nearly identical to the GeForce RTX™ 5090. It is a two-fan, two-slot, blow-through cooler designed for use in desktop workstations and configured to draw up to 600 Watts. A second version, the NVIDIA RTX PRO™ 6000 Blackwell Max-Q Workstation Edition, has desktop multi-GPU in mind. It features a two-slot cooler in a blower-fan configuration, and limits the power draw to 300 Watts. Finally, NVIDIA is offering a passively-cooled two-slot NVIDIA RTX PRO™ 6000 Blackwell Server Edition for use in proper servers with heavy-duty server fans, at an adjustable TDP of up to 600 Watts.
Other than the cooler and designed TDP, these cards are identical. Each features the same GB202 GPU die and memory subsystem. In this review, we will only be looking at the standard Workstation Edition. However, in the future, we plan to also review the Max-Q variant; we are excited to see how the lower-TDP can enable multi-GPU configurations that are currently limited with the Workstation Edition due to the high-TDP.
Much like the GeForce 50-series GPUs we reviewed earlier this year, the new RTX PRO 6000 Blackwell cards feature more and improved CUDA cores, fourth-generation Ray Tracing and fifth-generation Tensor cores, and the latest NVIDIA NVENC/NVDEC media engines. This brings better performance in traditional compute, ray-traced compute, and matrix compute. Of particular note is support for FP4 calculations (with sparsity) using the Tensor cores. This potentially allows for a doubling of performance on top of the standard generational improvements. Additionally, on the video editing side, the new media engines allow for hardware acceleration of H.264 and H.265 4:2:2 10-bit media.
Perhaps the most exciting part of the NVIDIA RTX PRO 6000 Blackwell is the memory subsystem. These GPUs feature a 512-bit memory bus at 28 Gbps, allowing for a maximum bandwidth of nearly 1.8 TB/s. While memory bandwidth itself can have a large impact on performance, even more impressive is the VRAM attached to it: 96 GB of GDDR7. These cards have twice the VRAM of the last-gen 6000 Ada, and thrice that of the GeForce RTX 5090!
For your convenience, we have listed the most relevant GPU specifications from the last few generations of professional video cards released by NVIDIA, AMD, and Intel. For more information, you can visit Intel Ark, NVIDIA’s RTX PRO Page, or AMD’s Radeon PRO Page.
Price- and specification-wise, the NVIDIA RTX 6000 Ada Generation and NVIDIA RTX PRO 6000 Blackwell cards are in a class of their own. Even the NVIDIA RTX A6000, which launched with a price around $4,700, is hard to compare to their $6,800 and $8,500 price tags. Due to this, we have chosen only to compare the above-mentioned cards in this review. However, as more of the Blackwell family becomes available, we will continue testing them. We will also test relevant comparators and include all the results in one chart so you can see how the family also compares within itself.
The standout feature of this generation (beyond the media encoders we discussed above) is probably the memory subsystem. The 6000 Blackwell cards feature a 512-bit bus with a maximum theoretical bandwidth of 1,792 GB/s. This supports the huge 96 GB VRAM buffer offered by these cards, both of which are nearly twice what the 6000 Ada generation cards were capable of.
However, NVIDIA’s new professional Blackwell cards also offer substantial theoretical performance improvements: 37% higher traditional FP32 performance, 80% higher RT TFLOPS, and a questionably arrived at 174% higher “AI Tops” matrix (Tensor) performance. Though these values rarely translate one-to-one with real-world performance, the “AI TOPS” claim buggers belief. One new innovation of the Blackwell architecture is for FP4 Tensor calculations. Assuming that their FP4 performance scales as expected, we predict a theoretical performance increase closer to 37% in Tensor calculation at FP8 (which they used for the 6000 Ada); there are some similar shenanigans occurring with the jump from A6000 to 6000 Ada.
Regardless, these are all impressive specification increases, though they come with a substantial increase in price and power draw. The power draw is perhaps the most noteworthy of these, as the increase from 300 W to 600 W makes it substantially more difficult to put multiple of these GPUs in one desktop workstation.
Test Setup
Test Platform
| CPUs: AMD Ryzen™ 9 9950X |
| CPU Cooler: Noctua NH-U12A |
| Motherboard: ASUS ProArt X670E-Creator WiFi BIOS Version: 2604 |
| RAM: 2x DDR5-5600 32GB (64 GB total) |
| PSU: Super Flower LEADEX Platinum 1600W |
| Storage: Samsung 980 Pro 2TB |
| OS: Windows 11 Pro 64-bit (26100) Power Profile: Balanced |
NVIDIA GPUs
Benchmark Software
| Lightroom Classic 13.1 – PugetBench for Lightroom Classic 0.96 |
| After Effects 25.2.2 – PugetBench for After Effects 1.0 |
| DaVinci Resolve 20.0 beta – PugetBench for DaVinci Resolve 1.2.0-beta |
| Topaz Video AI 6.2 |
| Unreal Engine 5.5 |
| V-Ray 6.00.01 |
| Blender 4.0.0 |
For our GPU testing, we have shifted to an AMD Ryzen™ 9 9950X-based platform from our traditional Threadripper™ platform. The 9950X has fantastic all-around performance in most of our workflows. This should let the video cards be the primary limiting factor where there is the possibility of a GPU bottleneck. For testing, we used the latest available GPU drivers. Our software packages are pretty typical for our GPU reviews, including most of the PugetBench benchmarks for Adobe applications, PugetBench for DaVinci Resolve, our in-development Unreal Engine benchmark, and industry standard Blender, V-Ray, and Topaz Video AI benchmarks. Unfortunately, we had to skip testing Premiere Pro due to some updates from Adobe around handling H.265 media. That change prevents PugetBench for Premiere Pro from working with application versions that fully support the new NVIDIA Blackwell GPUs. As such, we were unable to collect Premiere Pro results for this review. If you want more information on this, we have a blog post available explaining the situation in detail.
Raw Results Tables
We choose our benchmarks to cover many workflows and tasks to provide a balanced look at the application and its hardware interactions. However, many users have more specialized workflows. Recognizing this, we like to provide individual results for benchmarks as well. If a specific area in an application comprises most of your work, examining those results will give a more accurate understanding of the performance disparities between components. Otherwise, we recommend skipping over this section and focusing on our more in-depth analysis in the following sections.
Photography: Lightroom Classic
System Image
We would be very surprised if anyone was considering a GPU of this calibre specifically for Lightroom Classic work, but it is a competent GPU if that was a secondary application. As is typical, there isn’t a lot of distinction in the Overall score (Chart #1). However, we do see an acceptable increase of 18% in the Export JPEGs test (Chart #2), which is hardware accelerated. This isn’t worth the higher cost, of course, but it may be a happy accident for some users. Future testing of the Blackwell family, especially once we get into the lower-end cards, will be more relevant for this particular test.
Motion Graphics: Adobe After Effects
System Image
After Effects is similarly ill-positioned, in most cases, for a high-end professional GPU. Much of the application, in particular the “traditional” 2D workflows, is heavily CPU-bound. However, Adobe has started to add GPU-dependent 3D capabilities to the application. Our Overall Score (Chart #1) weights each of our major categories (2D, 3D, and Tracking) evenly, so we do see some movement on the Overall Score. However, we think it is more valuable to note the lack of distinction between the cards in 2D and Tracking, and move on to 3D.
In our 3D tests (Chart #2), the 6000 Blackwell Workstation Edition is 23% faster than the 6000 Ada and 63% faster than the A6000. Most users won’t find this performance jump worth the cost, especially since they are paying for the additional professional GPU features they probably don’t need. The VRAM capacity could come into play—both multi-frame rendering and 3D workflows can start requiring higher VRAM capacities for best performance—but even the A6000’s 48 GB is likely well sufficient for the majority of users.
Video Editing / Motion Graphics: DaVinci Resolve Studio
As we mentioned earlier, DaVinci Resolve is the only video editing application we are testing in this review. In addition to the general compute improvements, which we expect to see utilized in the GPU Effects portions of our testing, NVIDIA Blackwell also supports hardware acceleration for H.264 and HEVC 4:2:2 10-bit; we should see this in increased performance in the LongGOP tests.
Starting with the Overall score (Chart #1), we see a healthy performance uplift over the last-gen RTX 6000 Ada of 21%. The media engines carry a good chunk of this gain, with LongGOP tests (Chart #2) improving by 43% over the 6000 Ada and 114% over the A6000. GPU Effects (Chart #3) are even more impressive on the 6000 Blackwell Workstation Edition, leading the 6000 Ada by 78% and the A6000 by 167%. Finally, in our AI tests (Chart #4), the new GPU offers much more modest gains of 20% and 54%.
Overall, the RTX PRO 6000 Blackwell offers solid performance improvements in DaVinci Resolve over the last-generation 6000 Ada. At an average of about 20%, though potentially much higher in certain workflows, this just about matches the price increase. We’d encourage anyone looking at this card to also look at our review of the GeForce 50-series, as the professional features may not be worth the extra price tag in this case.
This could change quite a bit with the Max-Q version of the 6000 Blackwell, however, as the GPU effects in particular can scale with multiple cards. The price tag will certainly be hefty, but for some high-end users, the cost is secondary to the potential performance gains.
Topaz Video AI
Image
Topaz Video AI is constantly releasing new versions, making it somewhat difficult to keep up with it’s most recent changes. However, when the first Blackwell cards (the consumer GeForce cards) were released, we found that they were no faster than their past-gen counterparts. Luckily, it seems as if this issue has been sorted, at least for the Professional cards.
The 6000 Blackwell Workstation Edition is 30% faster than the 6000 Ada and 84% faster than the A6000. If you use Topaz Video AI professionally, the Blackwell card may be a good option, especially depending on how (or if) Topaz Labs brings their new Starlight model to local machines, hopefully outside of the Mini version.
Game Dev / Virtual Production: Unreal Engine
Our Unreal Engine benchmark renders a variety of real-time scenes, which feature a mix of features (such as Nanite and hardware RT) at three different standard resolutions. We combine all of those into one overall score, but to highlight some areas of strength and weaknesses of the new GPUI, we have pulled out two of those individual test scenes to examine more closely.
Starting with the Overall geometric mean FPS (Chart #1), we see that the RTX PRO Blackwell card is a nice 34% faster than the 6000 Ada and 110% faster than the A6000’s 69 FPS. The second chart is the mean FPS in our “GameDev” test scene, which enables Nanite and disables hardware RT, at the 1080P resolution. Here, the performance difference is rather small, at only 20% higher than the 6000 Ada. The overall rasterized performance uplift is smaller than in some other areas, especially at lower resolutions. Conversely, our third chart is a run of a test scene with Nanite disabled, RT on, and at 4K. Here, the 6000 Blackwell is 78% faster than the 6000 Ada; we can see that the improved RT cores make a huge difference here, alongside performance scaling as resolution increases.
GPU Rendering: Blender & V-Ray
Moving on to offline GPU rendering, we tested with two different applications; we hope to include more in the future, now that application support for Blackwell cards is improving. This is perhaps the best area (other than AI) for this class of professional cards, which offer huge amounts of compute and VRAM largely unattainable elsewhere. While professional cards also see use in scientific, medical, and engineering use cases, those use cases frequently target slightly lower-end cards.
In Blender, the 6000 Blackwell Workstation Edition is 50% faster than the 6000 Ada, and three times as fast as the A6000. V-Ray’s RTX mode (Chart #3) has nearly identical results, trouncing any previous cards. Our second chart, V-Ray’s CUDA mode, manages to offer even larger performance uplifts, with the 6000 Blackwell card 55% faster than the 6000 Ada and 250% faster than the A6000.
The biggest question of this card for GPU renderers is multi-GPU support. A pair of 6000 Adas will likely be more performant than a single 6000 Blackwell Workstation Edition, for the same power draw. At current pricing, the dual-GPU option would likely be about 10% more expensive. Dual Blackwell 6000 Workstation cards are theoretically possible, but will stretch most desktop PCs’ power supplies and cooling. NVIDIA has announced a Max-Q variant, which reduces power draw and adapts the form factor to be multi-GPU friendly, so we are also interested in testing those in the future.
How good is the NVIDIA RTX PRO 6000 Blackwell Workstation Edition for Content Creation?
Overall, the NVIDIA RTX PRO 6000 Blackwell Workstation Edition is a highly-impressive, highly-performant card which offers unparalleled VRAM capacity, compute, and hardware acceleration. We are particularly impressed by the memory subsystem, with the huge VRAM bandwidth and the 96 GB of GDDR7. However, it comes with a huge price tag and thermal draw to match. This GPU is not for everyone.
In Lightroom Classic, the new Blackwell card performs essentially the same as the last two generations of top-end professional cards from NVIDIA. It does show some difference in the photo export test, but a card of this class is excessive for Lightroom users. Similarly, this card is 23% faster than the 6000 Ada in After Effects, but it still offers a somewhat dubious value for users of that application due to its high cost compared to consumer GPUs with similar performance.
For DaVinci Resolve, we get to see the impressive new compute cores and media engines at work, with 43% performance improvements in LongGOP and 78% improvements in GPU Effects. Overall, the card pushes performance in this application, and the additional VRAM is a nice add. However, we think it only justifies its value if the VRAM is necessary, for some reason, or the other “professional” features (such as certified drivers) are a must. Topaz Video AI is similar, though with a 30% average increase in performance over the 6000 Ada.
Unreal Engine, a real-time renderer, shows the 6000 Blackwell card leading the 6000 Ada by 34%. This can increase up to 78%, depending on the scene configuration, but also drop to 20% or lower. Offline renderers like V-Ray and Blender offer both the best performance improvements and the most likely use cases for this card’s huge VRAM pool. V-Ray RTX and Blender Optix show the new RTX PRO card to be 50% faster than the last-gen Ada card, and the V-Ray CUDA test pushes this lead to 55%. There may be cases where two last-gen cards are preferable, but that will depend on cost.
In the end, the RTX PRO 6000 Blackwell Workstation Edition is a very impressive GPU, and for those who need the high performance or VRAM capacities, it is likely to be a very attractive card. Like all Pro-level cards, however, it is not meant for everyone, and is targeted (and priced) for professional users where reliability is of the utmost importance and the performance gains pay for themselves over time.
Looking for a content creation workstation?
We build computers tailor-made for your workflow.
Don’t know where to start?
We can help!
Get in touch with one of our technical consultants today.