Date: August 4, 2019
Author(s): Rob Williams
We’ve taken a look at AMD’s latest Radeons and NVIDIA’s latest GeForces in gaming, so it’s time to turn our attention to creators. We’re pitting the RX 5700 series against NVIDIA’s RTX SUPER, with a spattering of other models added in for a fuller look at what the proviz performance picture looks like today.
Having tackled gaming performance on AMD’s RX 5700 and NVIDIA’s RTX SUPER graphics cards over the past couple of weeks, we can now turn our attention towards the creative side of the market. Admittedly, this article should have been posted weeks ago, but we had more than one wrench thrown into our plans, some of which we’ll talk about throughout the article.
It used to be that creative work done on a PC would imply you were using a workstation-class graphics card, but that’s not anywhere close to being the case today. Even AMD and NVIDIA have softened up on their respective workstation card pushes, and it makes sense. Radeon Pros and Quadros have reasons for their heftier price premiums, but most of the world can’t stomach them.
There are a couple of reasons why you’d still want to consider a Radeon Pro or Quadro over a Radeon or GeForce, but nowadays, that largely has to do with the software in question. If you’re using a high-end CAD suite that focuses its support around workstation-class GPUs, you should probably just suck it up and go that same route to avoid as much potential headache as possible. But for hobbyists or students, gaming cards can fill the void more often than not.
A common selling-point for workstation cards for quite some time has been support for 30-bit color in OpenGL. Once again proving that we’re never quite sure what to expect, NVIDIA just opened up that feature to its current GeForce and TITAN line when the Studio driver is used. We’re still digging for the final word on this, but AMD’s Radeon gaming cards had 30-bit support enabled in OpenGL since around 2014, however, applications like Photoshop and Premiere Pro would not enable it unless a workstation card was detected. Whether this changes going forward, we don’t know.
Let’s spend a minute taking a look at the cards being tested. For the sake of keeping things simpler here, we’re including only the GPUs being tested today in these tables. You can look at fuller lists in some of our other recent GPU content.
|AMD’s Radeon Gaming GPU Lineup|
|Cores||Base MHz||Peak FP32||Memory||Bandwidth||TDP||SRP|
|WX 8200||3584||1200||10.8 TFLOPS||8 GB 8||512 GB/s||230W||$999|
|Radeon VII||3840||1400||13.8 TFLOPS||16 GB 4||1 TB/s||300W||$699|
|RX 5700 XT||2560||1605||9.75 TFLOPS||8 GB 1||448 GB/s||225W||$399|
|RX 5700||2304||1465||7.95 TFLOPS||8 GB 1||448 GB/s||180W||$349|
|RX 590||2304||1576||7.1 TFLOPS||8 GB 3||256 GB/s||225 W||$199|
|Notes||1 GDDR6; 2 GDDR5X; 3 GDDR5; 4 HBM2|
5 GDDR6 (ECC); 6 GDDR5X (ECC); 7 GDDR5 (ECC); 8 HBM2 (ECC)
|NVIDIA’s GeForce Gaming GPU Lineup|
|Cores||Base MHz||Peak FP32||Memory||Bandwidth||TDP||SRP|
|TITAN RTX||4608||1770||16.3 TFLOPS||24GB 1||672 GB/s||280W||$2,499|
|RTX 4000||2304||1005||7.1 TFLOPS||8 GB 5||416 GB/s||160W||$900|
|RTX 2080S||3072||1650||11.1 TFLOPS||8GB 1||496 GB/s||250W||$699|
|RTX 2070S||2560||1605||9.1 TFLOPS||8GB 1||448 GB/s||215W||$499|
|RTX 2060S||2176||1470||7.2 TFLOPS||8GB 1||448 GB/s||175W||$399|
|GTX 1660 Ti||1536||1500||5.5 TFLOPS||6GB 1||288 GB/s||120W||$279|
|GTX 1080 Ti||3584||1480||11.3 TFLOPS||11GB 2||484 GB/s||250W||EOL|
|Notes||1 GDDR6; 2 GDDR5X; 3 GDDR5; 4 HBM2|
5 GDDR6 (ECC); 6 GDDR5X (ECC); 7 GDDR5 (ECC); 8 HBM2 (ECC)
Regardless of the performance picture, we’re recommending anyone looking for a new graphics card for creative work to at least go with an 8GB option. You’ll be doing yourself a disservice by using less; maybe not right this minute, but surely before you’ll be looking to upgrade. We have a suspicion that the big reason NVIDIA didn’t increase the 2080 SUPER’s framebuffer is so that it didn’t eat more into the 2080 Ti’s market positioning, but it’s too bad nonetheless. Perhaps next-gen, we’ll see higher than 8GB become as standard as that is right now.
Chances are good that anyone looking at gaming cards are not going to be too concerned over error correction memory. It’s our understanding that in professional visualization, ECC is used primarily by the medical and energy markets, and not so much in design. That said, the option is available to those with higher-end workstation cards, but never on gaming equivalents.
Because NVIDIA offers 30-bit color on GeForce and TITAN now, the exclusive features list of Quadro slims a bit. As with the RPro cards, Quadros still carry a price premium largely because of the highly tuned drivers, and also performance optimizations for select high-end suites. At least with the case of NVIDIA, Quadro cards also offer far more NVENC streams than GeForce (20+ vs. 2), but a feature like that isn’t exactly relevant to the ProViz audience (it’s more relevant to Plex hosts).
In the workstation vs. gaming argument, the best point to make for RPro and Quadro is premium support. If you’re a gamer running into an issue, you’re not likely to be treated with priority, whereas that’s implied with workstation GPUs. If your work is critical, you need to weigh these factors.
On the following pages, the results of our workstation GPU test gauntlet will be seen. The tests chosen cover a wide range of scenarios, from rendering to compute, and includes the use of both synthetic benchmarks and tests with real-world applications from the likes of Adobe and Autodesk.
Twelve GPUs were tested suite-wide, but for our CUDA-only page, the non-SUPER 2060/2070/2080 and 2080 Ti were added to help bolster the results a bit more. Time didn’t allow us to run those through the rest of the suite, though the results on this CUDA page will largely represent scaling seen elsewhere.
Unfortunately, this article isn’t as standard fare as we’d like. We ran into a number of software-related issues that prevents the usual list of results from being seen. In particular, AMD’s Navi cards didn’t work in MAGIX Vegas Pro or SiSoftware’s Sandra, something we’re sure will be rectified before long. It’s doubly unfortunate that our AMD Radeon ProRender testing also had to be dropped due to a bug (which we talk about on the next page), as that would have been a third renderer for AMD cards, whereas NVIDIA has the bonus of four additional ones on the CUDA-only page.
Alas, nothing can be perfect. We’ll retest in due time, but for now, we still have many useful results to pore over. That all covered, the specs of our test rig are seen below:
|Techgage Workstation Test System|
|Processor||Intel Core i9-9980XE (18-core; 3.0GHz)|
|Motherboard||ASUS ROG STRIX X299-E GAMING|
|Memory||G.SKILL Flare X (F4-3200C14-8GFX)|
4x8GB; DDR4-3200 14-14-14
|Graphics||AMD Radeon VII (16GB, Adrenaline 19.7.1)|
AMD Radeon RX 5700 XT (8GB, Adrenaline 19.7.2)
AMD Radeon RX 5700 (8GB, Adrenaline 19.7.2)
AMD Radeon RX 590 (8GB, Adrenaline 19.7.1)
AMD Radeon Pro WX 8200 (8GB, Enterprise 19.Q2.1)
NVIDIA TITAN RTX (24GB, GeForce 431.36)
NVIDIA GeForce RTX 2080 SUPER (8GB, GeForce 431.56)
NVIDIA GeForce RTX 2070 SUPER (8GB, GeForce 431.36)
NVIDIA GeForce RTX 2060 SUPER (8GB, GeForce 431.36)
NVIDIA GeForce GTX 1080 Ti (11GB, GeForce 431.56)
NVIDIA GeForce GTX 1660 Ti (6GB, GeForce 431.56)
NVIDIA Quadro RTX 4000 (8GB, Quadro 431.02)
|Storage||Kingston KC1000 960GB M.2 SSD|
|Power Supply||Corsair 80 Plus Gold AX1200|
|Chassis||Corsair Carbide 600C Inverted Full-Tower|
|Cooling||NZXT Kraken X62 AIO Liquid Cooler|
|Et cetera||Windows 10 Pro build 18362 (1903)|
Our benchmark results are categorized and spread across the following four pages. On page 2, we’re taking a look at neutral renderers with the help of Blender and LuxMark. On page 3, we’re heading down the CUDA-only path with V-Ray, Arnold, Redshift, and Octane.
Page 4 is home to our encoding tests, which are handled by Adobe’s Premiere Pro, MAGIX’s Vegas Pro, and a new entrant in our GPU tests: Agisoft’s Metashape. Metashape is a photogrammetry tool, but until we add more such tools to our suite (such as Reality Capture, coming soon), it’s being tossed into the encoding section. Finally, page 5 handles viewport performance across a range of popular suites.
And with all of that covered, let’s get on with it:
In March, we took a deep-dive look at Blender 2.80 beta performance, and now that the 2.80 final has been released (as of last week), we can compare how things have changed since then. Ultimately, performance has in fact improved a bit over the past few months, at least for GPU rendering. All of the same projects rendered seconds quicker with the 2.80 final vs. the older beta build, which is great to see.
Interestingly, the changes seemed to give the Radeon VII a sweet boost in the Classroom test. It once came behind a few NVIDIA cards, but the tides have changed. In the BMW render, the TITAN RTX manages to hold onto its top spot. At the really high-end, some odd-ish results can be seen, like the 2070S outpacing the 2080S, albeit by a single second. At this point in time, the BMW project isn’t exactly representative of current Blender work, but it somehow sees massive improvement on the TITAN RTX.
When looking at Navi’s Classroom results, it almost feels like AMD performance has been given some real polish. The $400 RX 5700 XT places well ahead of the RTX 2060 SUPER in that test, but oddly, the roles completely reverse in the BMW test. With the heavily GPU-focused Eevee renderer, NVIDIA’s Turing seems to bring about some massive performance benefits, with all of the RTX cards leading the pack. The Radeon VII’s strength in the Cycles renders didn’t carry over to Eevee too well, with it sitting close to the bottom.
For those interested in Blender viewport performance, you can check out the respective performance on page 5. For CPU performance, our recent Ryzen 3000-series look should do the trick.
Please note that we’d ordinarily have Radeon ProRender performance here, but the most current plugin version (even as of the time of writing) for 3ds Max (2.5) does not agree with our tested projects. Our issues have been similar for both CPU and GPU rendering, where a few iterations would execute without issue, but at some point, the render process seems to break, and no further rendering is taking place – despite the pass numbers continuing to iterate.
We’ve been in contact with AMD regarding this issue, and we’re hoping to conquer it soon. AMD released a number of ProRender updates at SIGGRAPH, with new plugin versions having hit Blender and Maya. Once the new version for 3ds Max drops, we’ll give it a test. If we still run into issues, we’ll work to procure more current ProRender projects that will cooperate with the current plugin better.
If you are a ProRender user, and are interested in contributing a project for use in our benchmarking, please leave a comment!
A benchmark like LuxMark proves that we can’t just run a single test and feel confident in our results, because different renders behave differently on different hardware. In the LuxBall test, the Radeon VII dominates, which is something we’ve seen ever since the card’s existed. AMD even talked about this specific result during its CES keynote, but the more complex Hotel scene was left alone – and we can see why. The VII takes out the TITAN RTX in LuxBall, but the roles reverse in Hotel.
The Navi cards perform fine in LuxBall as well, but for some reason sit at the bottom of the Hotel render chart – even behind the obviously much slower-and-older Polaris-based RX 590. We’d wager driver optimizations could improve this in time.
All-around, NVIDIA’s SUPER cards perform very well here, while AMD is clearly hit-or-miss.
This page is dedicated to NVIDIA CUDA-based renderers, so Radeon fans will want to move onto the next page in order to get back to relevant results. Some of these renderers are known to be contemplating AMD and Intel GPU support in the future, and if/when it happens, we’ll adjust our testing accordingly.
Note that because these results would have been a bit barren given the lack of AMD cards, we tested the non-SUPER 2060/2070/2080 cards, as well as the 2080 Ti, to help flesh things out a bit better.
The middle of the pack might perform the same here, but we’re still able to glean some interesting results from these graphs. It’s interesting that in the simpler Flowers scene, the 1080 Ti settles low on the table, but in the more complex Teaset scene, it redeems itself. Meanwhile, NVIDIA’s Turing architecture again seems to offer some nice speed benefits in V-Ray.
If you’ve been debating picking up an affordable GPU that doesn’t sacrifice anything too severely, the match-up between the 1660 Ti and RTX 2060 is going to be a common one. Based on all of these results here, we’d wager that the RTX 2060 SUPER would be a bit of a no-brainer, if the 1660 Ti was the other option. The non-SUPER 2060 is also a strong performer in comparison, but for graphics work, the 8GB framebuffer of the SUPER is more attractive.
In recent months, Chaos Group overhauled its standalone V-Ray benchmark tool, reflecting an upgrade to the 4.X plugin version. With this move came a dropping of AMD support, which may mean the company has shelved its plans for better Radeon support for the time-being – a bit unfortunate as the 3.X version of the benchmark supported AMD via OpenCL (albeit with sporadic results). It could be that Chaos Group is looking for alternative methods for adding non-CUDA support. With Intel’s dedicated GPU coming, renderer companies are not likely to continue ignoring NVIDIA’s competition.
For the most part, the results here scale similarly to our real-world tests, with the middle of the pack performing largely the same, but each end of the stack providing a big delta in performance between it and the next step up or down. If you have a current mid-range NVIDIA GPU, you should feel satisfied. If you have deeper pockets and are able to splurge, some real gains can be seen by the top-end cards.
In Redshift, all of the cards scale about just as we’d expect, with none of the oddities that we’ve seen in some previous tests. Our chosen project for rendering is not exactly demanding, so it could be that with a more complex one, we’d see more interesting scaling. We might also see more interesting scaling if there were more GPUs in general included here.
All things considered, one card that’s performed quite well overall has been the 1080 Ti, so if you are currently rocking one of those still, your need to upgrade won’t be as severe as many others.
Octane clearly loves NVIDIA’s Turing architecture, with the new RTX 2060 SUPER beating out the 1080 Ti, and the TITAN RTX leaping far ahead of the rest at the top-end. The only reason the TITAN RTX isn’t allowed to completely dominate is because the 2080 Ti was introduced into these tests.
And to think, that’s with “RTX Off”. Let’s turn it on:
If scores accurately correlate to render times, then having “RTX On” is going to be a no-brainer. That also assumes that the end result is in fact final frame quality. We’ve been told by OTOY in the past that final frame renders with RTX On is in fact in the cards, so we hope to see that become a reality when the supported plugin comes out of beta.
Rendering a single frame for benchmarking’s sake is one thing, but in normal workflows, rendering is done often – and constantly if you’re using a live rendering mode. The faster your GPU, the less you have to wait. Clearly, you don’t want to go too low-end, but it’s nice to know that a ~$400 GPU is all it takes to deliver great performance in Octane.
We wasted no time jumping on Arnold GPU after it released in April, and at the time, we chalked some of the odd performance scaling up to the plugin being in beta. Well, fast-forward to today, where we’ve had a couple of updates since, we’re still seeing some oddities.
In the automotive render, scaling is overall quite fair, and expected, with the midfield performing similarly, and both ends of the pile showcasing large deltas. With our Sophie scene, the situation is quite different. Some of the GPUs simply won’t render the scene at all, or would take more than an hour to do it. This seems specific to certain GPUs, based on our earlier testing, but we wouldn’t expect it to be this way forever. Not that we can begin to understand why a bug would plague specific GPUs.
To Arnold’s benefit, the GPU renderer is in beta, and is in no way meant for production work, so there’s still time to see these kinds of bugs ironed out. Admittedly though, we didn’t expect to continue to see such odd behavior all this time after the beta plugin launch, but we’ll continue to monitor new releases and see if things improve with that particular project (which, for the record, was given to us by Autodesk for our Arnold GPU testing).
We’ve covered a few times before that our Adobe Premiere Pro tests will improve in time, but other test work has taken precedence (a DaVinci Resolve test will come before an updated PP test). For now, all we have are a couple of basic encode tests, which don’t show amazing scaling, but show scaling nonetheless.
8K is a meaty resolution, but with simple transcodes, it’s not incredibly demanding (at least, not our downloaded RED footage). When a real project is encoded, things change up a fair bit, but ultimately, the top-end of the stack performs similarly, with the Radeon VII keeping up to the NVIDIA SUPERs.
The picture for Navi in both tests isn’t the greatest, but, we are seeing big improvements over the Polaris-based RX 590. Will the same continue into Vegas? Well, it could, if it worked. Let’s discuss:
Vegas is another piece of software that we’ve revolved an entire performance piece around, but it’s admittedly one of our least-favorite tools to test, due to its unpredictable performance scaling and issues on NVIDIA hardware.
Long story short, NVIDIA users still need to add a profile for Vegas to their NVIDIA Control Panel in order to unlock full acceleration – though “full” might not be accurate, since it still feels like something is broken in the optimization chain. Thanks to all of this, AMD’s GPUs win this battle, unless we’re talking Navi.
Since our RX 5700 series cards worked for GPU acceleration in Adobe Premiere Pro, it’s unfortunate to see the cards not work in Vegas. Based on the fact that the cards will not work at all (an error is spawned as soon as encoding is started). We’d guess that this is a fix MAGIX needs to make on its end, and could relate to AMD’s new Radeon Media Engine. We’ll be monitoring the situation, especially as VP17’s launch is in our near-future.
Metashape is a brand-new addition to our GPU test suite, helping us take care of photogrammetry performance. That said, Metashape is not the only such tool out there, and we’re currently looking to expand our testing in the nearish future with Reality Capture. We’ve been testing with Metashape for a while as a CPU benchmark, but recent versions have added some GPU support, so we’re overdue on testing it.
For our GPU test, we just need to run the first two phases of the entire process. The first is to align and match photos, a process which happens so quickly, it’s hard to generate comparative data for it. The second phase, and its “Depth Maps Generation” is what we’re after for our GPU testing, as the CPU is left alone during the process.
We’re never sure what to expect from scaling when we test new software, but we’re happy to see pretty fair scaling overall here, with the TITAN RTX earning its lead, and performance scaling fairly naturally down the list. That said, AMD’s Navi cards are clearly suffering here, with the new RX 5700 XT failing to beat out the older RX 590. We’re not sure what will improve performance for Navi here, but we’ll test again when either a new Metashape build comes out, or we find out from AMD that a new driver improves the performance.
SolidWorks is considered one of the most important CAD applications in the business by AMD and NVIDIA, which is made clear by the fact that its workstation graphics cards give a bit of a boost to performance over the gaming counterparts. While not shown here, the workstation cards are required to access SW’s RealView feature, which enhances the viewport with better shading and lighting.
At 1080p, AMD’s Radeon Pro WX 8200 performed exceptionally well here, which is nice to see. At 4K, the advantage of the 8200 is replaced by advantages of the RTX 4000. It’s odd that the 8200 couldn’t keep up the same way at 4K, especially since both cards have the same framebuffer size. Ultimately, though, both of those cards are performing comparatively – as they should, given their similar price tags (~$900).
As for the Navi and SUPER battle, AMD has managed to shine here, with both of its RX 5700 cards placing ahead of every one of the SUPER cards – even the 2080!
If our last-gen Quadro P5000 and P6000s were tested for this article, we’d see more SKUs sharing the top of the chart, but since those are not present, the TITAN RTX is allowed to shine, placing ridiculously far ahead of the Quadro RTX 4000. For the second time in a row with a Dassault suite, AMD’s Navi performs great overall, once again placing ahead of all three SUPERs.
SNX is one of our favorite results to talk about, because it gives us the most extreme example of Workstation GPU vs. Gaming GPU that there can be. Siemens optimizes exclusively for workstation GPUs, seemingly going out of its way to cripple performance for gaming cards. This ultimately means that if you are an SNX user, you must go Radeon Pro, Quadro, or TITAN.
That said, it must be pointed out that AMD somehow performs very well in comparison to NVIDIA here. We’ve seen that before, but with Navi, we can now see that it’s not just a fluke. Still, the performance on offer with these cards is likely hitting a borderline for usability.
For serious work (and SNX implies that), even a low-end workstation card will fare better than a high-end gaming card. For comparison, the low-end Quadro P2000 hits 167 points at 1080p in this test, more than double the RX 5700 XT.
We’re not too familiar with how the real Creo application behaves, but we at least know from this test that gaming GPUs are not treated as poorly by PTC as they are by Siemens. The faster your GPU, the better your performance – it really is that simple in this case. We don’t see any odd scaling here, and overall, things seem fair between AMD and NVIDIA. If only more software behaved like this.
AMD won its SolidWorks and CATIA battles with Navi, but NVIDIA strikes back with 3ds Max and Maya. These results actually highlight the last-gen 1080 Ti’s benefits quite well. If you bought one of those cards at launch for creative purposes, it really has proven to be a worthwhile purchase. Even with today’s RTX cards, the 1080 Ti still shines.
Yet again, the TITAN RTX has proven that its extra muscle isn’t just for show. The performance delta it holds over the next-step down is impressive. It must be said if we’re comparing the 1660 Ti and RTX 2060 SUPER again, the latter is hugely favorable, partly because of its increased performance, but especially because of its larger framebuffer.
We’ve seen from previous Blender testing that NVIDIA’s Turing is a real force to be reckoned with. At all resolutions, the Turing-based 1660 Ti doesn’t fall too far behind the top-end Pascal 1080 Ti, and the RTX pushes things ahead even more. Fortunately for AMD, its GPUs hold their own quite well, and in some cases, better than NVIDIA. The 5700 (non)-XT outperforms the 2060 SUPER at all three resolutions. The XT manages to outpace the Quadro RTX 4000 just the same.
We should stress that the LookDev mode is really intensive, and 60 FPS performance isn’t really expected. Anywhere near 30 is going to feel fluid enough, but of course, the higher, the better. But when the TITAN RTX only musters 60 FPS at 1080p, it really goes to show how hard LookDev is on hardware. The Solid and Wireframe modes will run exceptionally well on modern GPUs at 1080p and 1440p, with most still being fine at 4K. LookDev is when you want your GPU to feel real pain.
One thing this article helps prove is the point that not all workloads are built alike. It’s a disastrous idea to have someone go out and randomly pick hardware for their given task. Right before Techgage came into existence, in 2004, a friend of mine purchased an expensive AMD FirePro GPU, simply because he believed the more expensive option would win out in gaming. He soon returned it for a better gaming option at the time, which of course cost less than half as much.
With the professional visualization market, a gaming card can either act as a superb value, or a major detriment. With Siemens NX, for example, gaming cards are treated like trash, where performance is so poor, users are forced down the workstation path. Fortunately, though, even lower-end workstation cards will deliver suitable performance in many of these high-end CAD/CAE suites.
In the basic AMD vs. NVIDIA battle, we have to talk features. AMD doesn’t offer the RTX features NVIDIA does, and at the moment, that could either matter a lot to you, or nothing at all. If you’re not using a solution that takes advantage of either the RT or Tensor cores, then you’re not going to miss out on anything by going with AMD. That of course assumes that your software doesn’t require CUDA support.
We’ve seen from OTOY that RTX support can dramatically improve render times, but we don’t want to count chickens until they hatch, or in this case, until the plugin is out of beta and user feedback starts rolling in. If RTX can actually triple rendering performance in most cases, that’d be a massive win for NVIDIA. Other CUDA addicts like Redshift, Arnold, and V-Ray are said to be adding RTX support soon. Even the free and open-source Blender will be.
Unfortunately for AMD, its Navi cards didn’t fare too well across a bunch of our tests. It did perform extremely well in SolidWorks, CATIA and Blender viewport tests, but in some others, like Metashape’s depth map generation and LuxMark’s Hotel render, both 5700-series cards nonsensical fell behind the older RX 590.
With AMD launching a brand-new architecture with Navi, its launch niggles are not too much of a surprise. Meanwhile, NVIDIA’s SUPER cards are iterations, so its support is strong from the get-go. We’re sure AMD’s situation will improve in the weeks and months ahead, as more updates roll out, either via the graphics driver or the vendor’s software updates.
It’s hard to declare a winner here, given that it truly depends on your workload, but NVIDIA does look quite strong overall, especially thanks in part to the fact that AMD couldn’t even run a couple of our tests. If you’re a Blender user, though, Navi proves to be a great choice, even managing to take out SUPER in the viewport test.
If you’ve read through and are still unsure about which card you should be going with, feel free to leave a comment!
Copyright © 2005-2019 Techgage Networks Inc. - All Rights Reserved.