Date: November 18, 2020
Author(s): Rob Williams
With a collection of CPUs and GPUs, we’re exploring performance in the latest version of Blender, 2.90. As before, we’re going to take a look at rendering performance with both Cycles and Eevee, as well as hybrid and OptiX rendering. For good measure, we’ll also take a look at Material Preview viewport performance.
November 18 Update: Article results updated to include AMD Radeon RX 6800 and RX 6800 XT performance.
Blender’s newest 2.90 version released two months ago, and for a couple of reasons, this performance overview has been lagged. Hassles in testing led to initial delays, and then other work came along and pushed production back further. We’ve continued to see reader request for this article, however, especially since the release of NVIDIA’s latest graphics cards – so here we are.
As we’ve done before, we’re going to be taking a look at both CPU and GPU rendering in Blender’s latest version, across both Cycles and Eevee render engines. We’ll also be taking a look at Material Preview viewport performance. As soon as 2.91 drops, we’ll jump on retesting.
It’s unfortunate that Radeon is the reason this article was originally delayed. A bug in recent Adrenalin drivers will cause Blender to crash during multiple renders on either the RX 5700 or RX 5700 XT. For that reason, we had to test those cards on the older 20.5.1 driver from this past summer.
It’s disappointing to see this issue arise, because we’re dealing with GPUs that are over one year old, and they remain AMD’s current top-end choices. Yet, for the past four months, Blender will crash during a simple BMW project render, each and every time. Unless you downgrade to the 20.5.1 (and possibly 20.6.X driver). If you opt for Radeon for Blender use, bear in mind these potential issues.
Without wasting more time, here’s a look at all of the hardware being tested for this article:
|CPUs & GPUs Tested in Blender 2.90|
|AMD Ryzen Threadripper 3990X (64-core; 2.9 GHz; $3,990)
AMD Ryzen Threadripper 3970X (32-core; 3.7 GHz; $1,999)
AMD Ryzen Threadripper 3960X (24-core; 3.8 GHz; $1,399)
AMD Ryzen 9 3950X (16-core; 3.5 GHz; $749)
AMD Ryzen 9 3900X (12-core; 3.8 GHz; $499)
AMD Ryzen 7 3700X (8-core; 3.6 GHz; $329)
AMD Ryzen 5 3600X (6-core; 3.8 GHz; $249)
AMD Ryzen 3 3300X (4-core; 3.8 GHz; $120)
Intel Core i9-10980XE (18-core, 3.0 GHz; $999)
Intel Core i9-10900K (10-core; 3.7 GHz; $499)
Intel Core i5-10600K (6-core; 3.8 GHz; $269)
|AMD Radeon RX 6800 XT (16GB; $649)
AMD Radeon RX 6800 (16GB; $579)
AMD Radeon VII (16GB; $EOL)
AMD Radeon RX 5700 XT (8GB; $399)
AMD Radeon RX 5700 (8GB; $349)
AMD Radeon RX 5600 XT (6GB; $279)
AMD Radeon RX 5500 XT (8GB; $199)
AMD Radeon RX Vega 64 (8GB; $EOL)
AMD Radeon RX 590 (8GB; $199)
NVIDIA RTX 3090 (24GB, $1,499)
NVIDIA RTX 3080 (10GB, $699)
NVIDIA RTX 3070 (8GB, $499)
NVIDIA TITAN RTX (24GB; $2,499)
NVIDIA GeForce RTX 2080 Ti (11GB; $1,199)
NVIDIA GeForce RTX 2080 SUPER (8GB, $699)
NVIDIA GeForce RTX 2070 SUPER (8GB; $499)
NVIDIA GeForce RTX 2060 SUPER (8GB; $399)
NVIDIA GeForce RTX 2060 (6GB; $349)
NVIDIA GeForce GTX 1080 Ti (11GB; $EOL)
NVIDIA GeForce GTX 1660 Ti (6GB; $279)
|Motherboard chipset drivers were updated on each platform before testing.
All chips were run with DDR4-3200 16GBx4 Corsair Vengeance.
All GPU testing was done with our Intel Core i9-10980XE workstation.
AMD Radeon Driver: Adrenalin 20.8.3 (20.5.1 for RDNA, 20.45.01.12 beta for RDNA2)
NVIDIA GeForce & TITAN Driver: GeForce 451.77
All product links in this table are affiliated, and support the website.
Because GPUs have become so powerful, the CPU is arguably not going to be an important decision for most people. If you want a CPU for rendering, you really need to opt for a model with as many cores as you can afford, since Blender does a great job of scaling up. With Cycles, however, high-end GPUs will match the performance of top-end CPUs easily.
With Cycles, we’ve seen (and will see again) that NVIDIA’s OptiX acceleration can dramatically improve performance, which is something really hurting Radeon prospects as well (if the driver hassles weren’t enough). Eevee, by contrast, only uses the GPU. If you want a fast CPU but don’t care about using it for rendering, you’ll want to prioritize clock speed.
A CPU like AMD’s $329 Ryzen 7 3700X or Intel’s $379 Core i9-10700K would make for a great choice, although if you are on a budget, the 6-core models should fit the bill just fine, as well.
From the get-go, NVIDIA’s latest GPUs have proven themselves a force to be reckoned with. NVIDIA promised that the RTX 3070 would be as fast as the 2080 Ti, and while punches are traded often on the gaming side, it’s almost always better in rendering. In each of the three render tests above, the $499 GeForce RTX 3070 outperformed even the TITAN RTX, not just 2080 Ti.
It’s always been intriguing that the Radeon VII performs so well in the Classroom render vs the other projects, but now that we’ve been able to add the Radeon RX 6800 series to the mix, we can see that some of those strengths have carried over to AMD’s newest cards. Unfortunately for team red, NVIDIA’s Ampere proves a lot stronger than the previous generation Turing, and this is before we see accelerated ray tracing with OptiX.
Here’s some Eevee performance:
It’s clear that NVIDIA doesn’t see as large a performance boost in Eevee as it does with Cycles, although its parts are powerful enough to soar to the top, anyway. Here, the RTX 3070 actually fell behind the RTX 2080 Ti, although not by too much. The RTX 3090 really stands out, pushing well ahead of the RTX 3080. It seems that Eevee really likes fast memory, and the more of it, the better.
With regards to AMD’s new RX 6800 series, we can see huge performance gains in Eevee, when compared to the last-gen RX 5700 series. The 6800 XT sits behind the last-gen 2080 Ti, but both it and the RX 6800 make a sandwich out of the RTX 3070.
CPU + GPU hybrid rendering has been good for quite some time in Blender, although since the introduction of NVIDIA’s ray tracing acceleration, it’s a little less lucrative than it used to be. Whereas the 64-core Ryzen Threadripper 3990X and 2060 SUPER rendered BMW in 39 seconds, we’ll see next that OptiX on a modest RTX 2060 can do it in the exact same time. That said, if you don’t have a newer GPU, but do have a powerful CPU, you will definitely want to take advantage of hybrid rendering.
It’s worth noting that OptiX doesn’t offer 100% feature parity with the CUDA API at the moment, but that’s unlikely to be an issue for any modern projects that are built around OptiX from the start. Speaking of OptiX, let’s tackle that performance next:
These results make it clear that NVIDIA’s OptiX API can make a huge improvement to Cycles render times. Normally, a fall from 52 to 34 seconds in a render would require an entirely different model product, but in the RTX 3070’s case, it’s a matter of enabling OptiX.
With AMD’s cards now tested, we can compare our earlier performance to here. The RX 6800 XT rendered the Classroom project in 41 seconds, which put it at the top of the chart. With OptiX engaged, the RTX 3080 settles at 31 seconds. That’s quite strong performance from the Radeon side nonetheless, although the Classroom project proves the most favorable towards it. The Controller project performance deltas are a bit wider.
As a reminder, OptiX acceleration doesn’t impact Eevee, so if you’re thinking about skimping to get a lower-end GPU because the OptiX RT is so good, bear in mind the Eevee scaling, as well as the amount of memory you’re going to have available. We’d highly encourage opting for GPUs with at least 8GB of memory nowadays.
We’re going to take a look at viewport performance on the next page, as well as try to draw up some conclusions.
With the launch of Blender 2.80, the software introduced a new look development mode called Material Preview. It loads in the textures and enables accurate shading and lighting to give you a mostly accurate representation of your scene without having to go the rendering route.
It is important to note, though, that Blender has multiple denoisers built-in, so in the event you need to test changes with a real render, you can also take advantage of Intel’s OpenImageDenoise or NVIDIA’s OptiX for AI denoising to get an even quicker representation of the scene. You can see some of this in action with OptiX in our recent RTX 3090 rendering video.
From these results, it’s clear as day that the chosen project isn’t the most scalable. That’s because it’s not overly complex, so at 1080p, the CPU quickly becomes a bottleneck with top-end GPUs. That’s even largely the case at 1440p, so we’re going to find a secondary (more complex) project to add soon in hopes we’ll see some better scaling in more demanding projects.
Nonetheless, we’re still seeing great scaling at 4K, as the GPUs have to work a lot harder than the CPU to get the image rendered to the screen. NVIDIA’s Turing and later GPUs dominate the ranks here, with the GTX 1660 Ti even managing to topple AMD’s RX 5700 XT. Luckily, AMD’s new RX 6800 series perform a lot better. With LookDev viewport performance, you can get by with 20 FPS, but naturally, the faster the frame rate, the better the overall experience will be in more complex scenes.
Which CPU or GPU you need for your Blender work will depend largely on your budget and needs. If you’re primarily working with the Eevee engine, then your best hardware is going to differ from someone working completely in Cycles. NVIDIA’s OptiX makes its RTX GPUs a no-brainer for Cycles work, but options become more varied when we’re dealing with the memory-hungry Eevee.
On the CPU front, it’s clear that faster CPUs render a project faster, but it almost doesn’t matter if you’re rendering using Cycles and have an NVIDIA RTX GPU at-the-ready. Given the speed-ups with the GPU, even without OptiX, it makes sense to focus your CPU purchase on one that offers a really fast peak clock speed. The better the single-thread performance, the snappier UIs and interactions will be. Note that we’ll be testing Zen 3 CPUs for our upcoming 2.91 performance evaluation.
It’s worth noting that there is more to Blender than the viewport and rendering, with video point tracking, physics calculations with soft body meshes, the various baking, caching and composite stages used in creating the scenes in the first place, all being largely CPU-bound with limited multi-threading capabilities. It comes back to what do you spend most of your time doing in Blender, and if that work prevents you from doing something else at the same time.
While CPU speed can sometimes impact viewport performance, that kind of thing won’t matter in Blender so much. Wireframe and solid modes are blazing fast on any CPU and GPU; it’s the Material Preview mode where additional performance would be nice, but that’s tied directly to the GPU. We ran our Controller project viewport on four of our platforms (AM4+3950X, TRX40+3990X, X299+10980XE, and Z470+10900K), and with a Quadro RTX 6000, the performance was identical on each.
As mentioned in the intro, we’re less-than-impressed with Radeon in Blender at the moment, and if we had to choose just one reason, it’s because AMD’s latest available series crashes in it with any driver from the past four months (validated on two PCs with both 5700 and 5700 XT). Fortunately, we’ve actually had the opposite experience with AMD’s new ‘Big Navi’ graphics cards. So far, Blender testing on the RX 6800 and RX 6800 XT has been rock-solid.
All of that being said, we expect the situation with Radeon on the whole to get better again. The company has been enjoying so much success lately, it simply has to lead to there being more engineers on the driver team.
Blender 2.91 is due in couple of weeks, and we won’t be wasting time jumping on testing.
If you are still left confused, or have other questions, please feel free to drop them in the comments below.
Copyright © 2005-2020 Techgage Networks Inc. - All Rights Reserved.