All hardware recommendations I could find about this topic fall down to the same criterium: render times. However, in my workflow, this is not, by any stretch of the imagination, the most important thing. Cycles is, of course, an important component of Blender, and ray tracing is crucial in this job, but Blender has so much more functionalities, that we can’t just look at it only from this angle.
Several months ago I bought this computer that I use for my 3D workflow. It is a Ryzen 5950x, with 128 GB of RAM and, at the time, it had a Nvidia 2060 with 12 GB of VRAM (I upgraded to a RTX 4090 lately, I will cover that). Initially, I specified this computer for photogrammetry workloads, especially Meshroom and Metashape. For these software, we need lots of RAM and solid multicore performance, which almost coincides with the needs of a renderer. At the time, the Ryzen 5950x was on par with Intel’s 12900K, but was cheaper and more power efficient.
The GPU performance is not that important in this case, because it accounts for just a smaller part of the job. However, since we were dealing with very high-resolution images, it was a good idea to have the most VRAM possible. Since the 2060 had more VRAM than its most expensive counterparts for a fraction of the price, and not a crucial difference in performance in those applications, I could save a buck and invest in faster storage. That proved to be a really smart move, even when I moved to Blender.
How a computer works
If you understand this concept, all will be clear. Basically, a computer has three types of memory: processor cache, RAM, and storage. The cache is much faster than RAM which is much faster than storage. The CPU actually works with the data that is stored in the cache, which, in turn, it gets from the RAM. When it needs something that is not in RAM, it gets it from the storage. The storage is more than 10 times slower than the RAM, so we can see where it is going.
But it gets worse. Imagine that the RAM is fully occupied. The operating system, then, moves something that is in the RAM to the storage (aka SSD) to free some space for new things. But, to add insult to injury, the writing speed of an SSD is even slower than the reading speed. You can see where this cycle goes. The system becomes a real mess real quick, slows down and eventually crashes.
The key is to have enough RAM so the OS doesn’t have to use the storage device to process data. The same logic goes for the GPU that, in fact, is nothing more than a computer in itself, but optimized for a lot of parallel small calculations. In both cases, all you need is that your model fits in the RAM (or VRAM, video-RAM, in the case of GPUs).
As for the cache, you don’t need to worry about that. The chip makers put the needed amount in them anyway.
If you got this section right, this article will age well, unless computer architecture changes radically.
Sometime after I built my system, I started messing up with Blender. With all the RAM and the VRAM, I was able to make really nice things, this system served me well and I could make good money from it. The 2060 was a nice shot, more often than not, I found myself being saved by the 12 GB of VRAM. The reason is that the GPU is only useful when the model fits into the memory. After all VRAM is consumed, it starts sharing the system RAM, and things start to fall apart real quick. If you look into Windows Task Manager GPU usage graphs, you can see the GPU waiting for data.
That’s why a fast GPU with little VRAM is almost useless in Blender, and you are going to be seeing yourself going for CPU rendering more often than you would like. In fact, this breaks the deal for those working with animation, because even the fastest CPU can’t handle things fast enough to render lots of frames without taking ages. So, a 3080 with 10 GB can be less useful than our humble 2060 with 12 GB.
VRAM is king. Between a fast GPU and one one with the most VRAM, always opt for the most VRAM. Best bang for the buck here is an Nvidia 2060 or a 3060 with 12 gb. If budget is restricted, you can get away in the beginning and start make some money with 8 gb, and the best value here is a 3050. If you are going for the kill, 4090 is the weapon with its 24 gb. The 3090 is a good option too, and you can find good deals now that it is superseded by the 4090. If things get more serious than that, Nvidia offers up to 48 gb in the A6000 model, but you your wallet will become very light, very fast.
However, what nobody talks about is viewport and general modeling performance. Here things get really nasty, really fast. If you are going for photo-realistic stuff, you are also going for fine details, which leads to ridiculous amounts of vertices which, in turn, take a lot of memory. And, to make things worse, this will also require memory bandwidth. This is, in my opinion, the second most important aspect of a GPU. Get one with the best VRAM bandwidth possible.
About AMD graphics cards, at the time, I would say to avoid them. The performance in rendering is ages behind Nvidia and the prices are not that great from this optic. The big deal with Nvidia is the ray tracing capabilities that Cycles can explore with OptiX. At the moment, AMD has already released its ray tracing API, called HIP, but it is not on par with Nvidia. A new and improved version is expected to be released soon in Blender 3.6 or 4.0, but there are only rumors about that.
Finally, I must make some comments about my last upgrade, which is an RTX 4090. As I said, I came from a 2060, so I made my way from the bottom. What differences could I see? First, the viewport performance is almost unaffected, since those operations don’t mostly rely on GPU. Mesh editing, modifiers, geometry nodes, etc, benefited less than I expected because they rely on single-core CPU performance and RAM. On the other hand, material preview and live rendering with cycles is mind-blowing. Even heavy models with complex procedural materials can be visualized almost in real time. This saves a lot of time.
Speaking of time, I can build extremely heavy models and render them really fast. So I can make complex animations that my humble 2060 couldn’t handle. This is another major blow to AMD GPUs, they just can’t compete here. Also worth mentioning is that, since it is ridiculously fast, I can make quick renders to tweak my models all the time. This adds up quickly to the model’s quality because I can fine-tune all the details I want.
A last recommendation is to avoid older GPUs without ray tracing. You can find 1080s for a good price, but they are not compatible with OptiX, which is almost 50% faster than CUDA, the standard Nvidia API. They may be a good idea for Metashape, Meshroom, or Reality Capture, but not for Blender, especially Cycles. The best GPUs for Blender are Nvidia from 2000, 3000, and 4000 RTX series.
The benefits of multiple GPUs will show up only in production since Cycles can spread the load across several GPUs at a time. The performance scales up almost linearly with the number of cards, which means that 2 GPUs may be almost as 2x faster than a single GPU. However, keep in mind that here is where your CPU will really bottleneck the system, but I will get to it next. The viewport performance, however, may not see such an impact.
Another thing to bear in mind is that Cycles doesn’t do memory pooling. This means that it can’t sum up all the GPUs’ VRAM. For example, if you have two 8 GB GPUs, it won’t see 16 GB, but 8 GB. Also, When using multiple GPUs, it will access the amount of the smallest VRAM. If you have a GPU with 12 GB and one with 4 GB, it will spread the load between the two, but will only see 4 GB in total. If you need more VRAM, there is only one way to go, which is higher-end models like the 3090, 4090, or A6000.
As far as I can notice in my workflow, the CPU is responsible for tasks like mesh editing, modifiers, geometry nodes, texture nodes, physics, and almost all the activities in the viewport. Try to bake some rigid body simulations and watch your CPU in the task manager. You are going to see only one core being used at a time. This trend goes for almost all operations, mesh editing and modifying included.
Things get really worse when you consider that most of those tasks are actually single-threaded. So, a high-end Ryzen 7950x or Intel 13900K won’t add much here. My recommendation, if you are going with a single GPU, is to look for single-core performance and forget about rendering performance. In this case, they are all very similar from bottom to top: an i5 is not that slower than an i9. So, I would save some money with a cheaper processor and spend it on more RAM or a better GPU. I would go as far as to say that the modeling operations in my i5 11260H laptop are not much slower than in the 5950x. It is in rendering that things start to pile up, but CPU rendering, in my opinion, is a big no-no.
Most of the reviews about the best CPUs for Blender I can find really don’t talk about that – even Puget Systems or Tech Notice Youtube channel (which is the best for content creators). They are always recommending based on rendering performance, but you should avoid rendering in the CPU as the plague. First, it is slow, really slow. A scene that took two hours in my 5950x was rendered in two minutes (!!!) by the 4090. The difference here is that I can make an animation with this scene in the 4090 but can’t in the 5950x. As a comparison, the 5950x makes something about 400 points in Blender Benchmark, while the 2060 (an entry-level GPU) is able to make more than 1500, almost 4 times faster because Blender Benchmark score scales linearly with performance.
The only advantage of CPU rendering is that it can (sort of…) handle gigantic scenes. The problem is that it gets slow as the scene size grows, which defeats the purpose. Maybe CPU rendering is worth it if you work with static scenes, but even then, I would not recommend it from a productivity standpoint.
Here is where people enter the Threadripper pit – in most cases, it would be called Walletripper, but I digress. The fastest processor will get ripped (pun intended) by the most mundane GPU in render, as quick glance at Blender Benchmark rankings may reveal instantly. And, as we saw earlier, in modeling what counts is single-core performance. Although a Threadripper is an outrageous multicore, in a single core it lays behind very basic CPUs, which means they are not great for model editing. In fact, most people would have their workflows slowed down by a Threadripper.
When it makes sense, then? Threadripper – and Xeon, Epyc, etc; are really good in three things: lots of cores, lots of ram, and lots of PCI-E lanes. You can get as many as 64 cores, 2 terabytes of RAM, and 128 PCI-E lanes. A Ryzen 9 7950x, for comparison, offers 16 cores, up to 128 gigabytes of RAM, and 24 PCI-E lanes. So, you are going to need such a CPU in a few cases: where your models take more than 48 GB (so even an A6000 wouldn’t handle it), when you have to work with several large-scale models at a time, or using more than two GPUs.
The least, in my opinion (but it can change according to your circumstance) is the most important factor. If you need raw speed, you are going for multiple GPUs. PCI-E lanes are connections that go directly to the CPU, so they are really fast because the GPU can talk directly to the CPU. Every GPU uses 16 PCI-E lanes, so it would leave only 8 free lanes in a 7950x. From those 8 free lanes, I would expect 4 to go for an NVME drive and the other 4 would probably go to USB ports or chipset. So, with only one GPU, your system would be out of direct talking to the CPU, which, in turn, would slow things down a bit – but not that much. Most PC motherboards with 2 PCI-E slots for GPUs make the connection to the second GPU via the chipset, which is a bit slower than connecting directly to the PCI-E lanes of the processor itself. Although it is not a deal breaker, it is not ideal, either, but one can get away with that.
Things get interesting when adding more GPUs. PC motherboards don’t offer more than two slots for GPUs because neither the CPU nor the chipset can handle them. That’s where Threadrippers make sense. Depending on the motherboard, they can handle up to 7 GPUs. So, if you are not going for more than 2 GPUs, a Threadripper is really a wallet ripper with a penalty in viewport and modeling performance. The case against Threadrippers is even more important from a business standpoint. Money here is key. In a business, the question to answer is always the return on investment. Will a Threadripper offer more return than an i5? In most cases, the answer is no and your business can actually go bankrupt if you take the wrong direction.
A high-end system is expensive for small businesses, so it is a matter of survival to know when you really need one. The same goes for GPUs, and that is the reason why the Nvidia 4000 series makes so much sense for business. Only the time you gain from production in a 4090 pays off quickly the amount you spent on it. It is not uncommon to pay for the card in its first job! It is a no-brainer, really. Actually, it should make more sense to go for dual A4000 or A6000s before spending on a Threadripper, even more so because they support NVLink, which enables memory pooling.
Storage speed plays a great role in Blender because more often than not, we have to deal with large files, 3 GB and over. I have some HDs that I use for long-term storage, so they must be cheap and spacious. However, when I need to open those files, it is really apparent why HDs are out of fashion. They are almost an order of magnitude slower than an NVME.
So, for the working drive, I would suggest an NVME compatible with PIC-E gen 4 x 4. Normally, motherboards can connect only one NVME drive to the processor’s PCI-E lanes, and I would use the working drive in this slot. For the operating system drive, I would plug it in the second NVME slot, if available. If not, I would use a SATA SSD, it is enough for loading system files. If you are on a budget, you can get away with a PCI gen 3 x 3 NVME that is considerably cheaper and fast enough in most cases.
For long-term storage, as I mentioned, I use HD SATA drives because they offer a lot of space for a better price. Those files, most often, are works that are finished and I will probably never touch them again, but you never know when a client will ask for something, So it is a good measure to have them stored.
My workflow here is that I keep the active projects in the NVME while working on them. After the project is finished, I move them to the HDs. If I need, I can move them back to the NVME again. The HDs are only storage shelves.
As much as possible, take no prisoners here. With less than 64 GB you will get stuck pretty quickly. I can handle large projects in my system with 128 GB with no problems, but it is common to use more than 64 GB. Also, faster RAM will affect the modeling tasks, so I would prefer DDR5 over DDR4. However, I think 128 gb DDR4 makes more sense than 64 gb DDR5. Memory amount is key here, and you can save a lot of money going DDR4 since processors and motherboards are much cheaper. Then you can spend the money on more RAM or a better GPU. In my opinion, if your budget is constrained, DDR5 may not be worth it. Also, Intel processors tend to deal better with 4 sticks of RAM than AMD’s counterparts.
This is another overlooked aspect of a rendering PC. The quality of your screen hits your work frontally. The ability to see the correct colors is key to achieving good-looking renders. If your monitor is low-end, it will tend to exaggerate colors and miss some tones. When people with better screens than yours (anyone with any Mac, basically) see your renderings, chances are that your images will look amateurish.
To avoid that, look for monitors that have 10-bit color depth and can access more than 110% of the sRGB spectrum. This is key and is a detail that is crucial to get professional results – as much as a good GPU! Resolution is of secondary importance here, so between 4k and sRGB spectrum, always go for the latter. You can get away with a 1080p screen, but you can’t with crappy color coverage.
The tip here is to look for reviews of photo editing monitors. Forget about refresh rates and all the gaming hype. For artistic work, you will need something great at displaying colors, good contrast and that can hold calibration for a long time. You may end up in situations where your renders are printed, and you may need to take proofing and all, which is devastated by a bad screen.
- VRAM is the priority. The best budget option is the 3060 12 GB. The best high-end is the A6000 48 GB. The fastest option is the RTX 4090.
- Gamers hate the 4000 series, but they are an amazing investment for those that don’t live in a basement, a.k.a. entrepreneurs.
- AMD GPUs are a no go. They are slow in ray tracing for the time being (May 2023).
- Single-core performance is the most important CPU aspect for modeling. A Ryzen 5, most of the time, can do almost the same tasks as a Ryzen 9 and better than a Threadripper.
- Get the most RAM you can. RAM speed is also important, so prefer DDR5 over DDR4.
- High-end processors, like the Threadripper, make sense only with more than 2 GPUs or gigantic models.
- A good monitor is as important as a good GPU or lots of RAM.
Entry-level PC recommendation for Blender
- Intel 13600K or Ryzen 7600x (if you are really on a budget, get a 5600x)
- at least 64 GB of RAM, DDR5 is preferred
- RTX 3060 with 12 GB of VRAM
- At least a Gen 3 x 4 NVME SSD
- The biggest monitor that can handle more than 110% of the sRGB spectrum
Best value PC recommendation for Blender
- Intel i7 13700k
- 128 gb DDR5
- Nvidia RTX 4090
- 4k professional monitor for color grading