|
|
Thank you Carey for sharing your insight. That was a good read
Adam Mercado Owner & Creative Idiot Influxx Media Production 714.928.9896
Moving Images. For Business A few comments:
Puget Systems REALLY gets it when it comes to addressing the needs of our market. They have a model where they do a ton of research and provide excellent support for their tested, configured system designs. They share a lot of their data and findings on their website, which benefits the community as a whole. They are uniquely positioned to do this type of testing and I’m very thankful they share. For the vast majority of people who do NOT have interest in building their own systems, but nevertheless are realizing that a custom workstation is truly the best path forward for their needs, I can’t think of a better choice.
Now onto the geek stuff, (which Puget already knows):
A - Building PC: I’m inclined towards PC building too, and have built many workstation machines. I built my first PC around 2001 mostly for gaming and offloaded encoding/rendering but I largely stuck with Mac for post production work until about 2012 when I made the complete jump to primary post production on Windows workstations. Teddy and I had a good chat on the phone back then just to gut check my system architecture before pulling the trigger on a $6k box build. I’ve built and rebuilt many since.
B-Performance: It used to be that synthetic benchmarks were quite useful in determining performance as it applied to video post work. You pretty much always needed the highest performance hardware you could get and there wasn’t a whole lot of differentiation about tasks, as it all sort of lifted with the general performance of the machine. You could look at relative performance and extrapolate your value or expected performance for the apps and work you did and make a decision about value for money/ROI/Cost-benefit with some simple math.
Those days are over. That won’t work anymore. The trend for nearly a decade is that PC performance gains have forked in multiple directions: CPUs have added cores, and clock speed increases have crept higher but very slowly. Due to thermals, you are often left with a choice of more cores vs more clock speed. GPUs have introduced massive parallelism tuned for specific compute loads but data has to be moved from memory/fast CPU access over the bus to the GPU and back so that is a factor in determining which types of “loads” are best suited to GPU compute. The problem is that synthetic benchmarks can be designed to torture test a specific type of computing load that shows maximized performance of a particular component but it has never been less relevant to apply those to your expectations in the apps you ACTUALLY USE to do post production work.
Therefore, the ONLY benchmarks that are relevant are the ones that are done by testing the apps you ACTUALLY USE to determine a particular component or configuration’s value to you. Here’s some apps that I use and what I have found they need from a machine.
For example here’s some differing compute loads and what types of PC hardware/config will serve it well.
Storage (applies to all): Everything wants fast storage and IO. Splitting up storage between media and boot and cache, each optimized is good. Storage/cacheing has become an issue too and managing the flow of data and using high performance caching devices separate from both system drive and media storage is key to optimizing. As you look at the applications below
AE: In daily use, largely still single threaded operations with dashes of multithreading throughout, and a splash of GPU compute here and there, particularly on plugins/effects. Today, this means favoring higher clock speeds over core count on the CPU. Get as much RAM as you can justify because AE and RAM previews will use it. Generally AE-only artists don’t need tons of data storage, but what they have should be fast. Often they are best served by something like a 9900k system and a modest GPU from NVidia. Note that one of the challenges of optimizing AE for multithreading is that not every type of compute task lends itself well to parallelism. Sometimes you have to wait for the result of one computation to use the result of that to start the next computation. This same performance optimization issue also affects other similar apps like Fusion.
Premiere: A little more balanced between CPU cores and clock speed and GPU. Having said that Premiere still favors clock speed over massive core count, can take advantage of Quicksync on newer Intel CPUs BTW, and uses GPU but does not rely heavily on it and primarily processes color and effects on GPU (with some exceptions). Premiere doesn’t require quite as much RAM as AE, but I guess it doesn’t hurt. Having said that Premiere seems to need an awful lot of compute hardware thrown at it to smoothly handle media at 4k. By comparison FCPX (which I don’t use because it’s Mac only) is way more optimized and gets better performance from lesser hardware. I think there’s a lot of room for improvement for Premiere here but it probably means a rewrite.
Encoding: h264 or now h265 encoding is massively parallel and really shows off what a lot of cores can do. Currently CPU dominated we’re starting to see this moving to GPU options as well. If all you did all day was encode stuff, go for core count and fast IO. However, for most of us, encoding occupies only a few minutes per day compared to creative work.
Resolve: GPU dominated (for color page operations anyway). Here, CPU performance matters but a mid-line CPU is probably adequate but what you want are GPU compute units. The best units you can afford, plus more of them. 2 or 3 GPUs scale really well with Resolve, but in my testing I’ve found it best to keep them matched. I did some mismatched testing with different gen GPUs and it’s pretty variable and inefficient. Not that every little bit doesn’t help…but one card ends up waiting on the other to finish a task. PCI-E lanes required for multi-GPU support push the user to go with higher level platforms like X299.
Cinema 4D: When rendering, massive CPU core count is your friend. But when modeling, key framing, animating etc a lot fo that is more single threaded but obviously previews can be run OpenGL (etc) on the GPU. Other renderers are GPU-dominant. 3D is not my forte so I’m going to let others cover this.
But I said all of that as a preface to make this point: You can choose any one activity or app and build a kick-butt machine tuned specifically to the requirements of that activity, and walk away with maximized ROI. An AE-only machine is going to look quite different than a Resolve-only machine, and again quite different from a C4D optimized machine. But often we need to engage in a mix of activities and the choice about whether to try to build the ultimate “do it all” workstation ($$$$) vs separating the tasks is something we all have to look at in light of our own mix of work.
Anyway, when I sat down I had no idea I’d write ll that but alas, the coffee was good and plentiful. Plus, when you really get into this stuff there are just a lot of considerations to get your head around. By no means did I write a definitive guide here (is there such a thing?) and understand why Chris Zwar’s is over 10,000 words. I look forward to it.
Carey Dissmore
Puget really impressed me with the depth and detail of their hardware analysis for very targeted demographics. It's nice to have someone really understand what the difference is between editing, motion graphics and GPU rendering. Sure thing Chris, feel free to reach me at teddygage@gmail.com. Goes for anyone- always happy to consult or chat hardware or software. Have also been doing a fair amount of VFX supervising recently for a couple TV shows. Hi all,
I started researching performance last year, with a view to writing up an article and some thoughts for the ProVideo Coalition. Currently I'm about 50% of the way through my plan and I'm over 10,000 words so it's a bit slow going. Even so, I was wanting to reach out to list members who are especially interested and knowledgable on performance related matters.
Teddy & Stephen were both on my list of people to approach for input and this thread just reinforces that! Are you two able let me know the best way to contact you off-list please? And if there's anyone else who has strong knowledge or strong opinions on performance/ components then let me know too.
Cheers,
-Chris
Those are good thoughts Stephen and I'll keep that in mind going forward. I am planning to use a Blackmagic card for video output but the monitors I have to use right now are 8bit HD. I'm starting to get into 4K VFX for films on my machine but most of the rest of the team is building HD promos/mograph content. My only pet peeve with the BM cards is when you have a 4K screen it won't downscale through the card and just goes black.
> In regards to 10bit output an option not being discussed is using a Blackmagic or Aja PCI card or device. These will give you the option of 10bit YUV or RGB output (monitor dependent). 4K capable versions of these are available for as
little as $200. I mentioned external card but you are not budgeting for the 4K 10-bit display that you will need to see the output. I have a $199 BM MiniMonitor 4K and it’s great going out to a 50” HDR TV).
But that does not get you 4K 10-bit on your desktop display or apps that don’t support that card. You have to have a second monitor where as Quadro allows you to have 4K 10-bit on your main desktop monitor. If you only have room for on 4K 10-bit, a card plus additional monitor is not an option. And even if you do, my setup I would rather always see 10-bit as much as possible in all my work and catch issues. As video moves to 10-bit and HDR in 4K, even for low-end, 10-bit is going to be a bigger deal.
If you’re talking about a budget system Quadro cards aren’t even in the conversation (5k was mentioned). If one of the main uses of the machine will be Redshift you’re better off buying the fastest GTX or RTX card(s) you can afford. The
RTX cards may end up being a better choice in the near future, if Nvidia’s and Redshift’s recent announcements are to be believed.
In regards to 10bit output an option not being discussed is using a Blackmagic or Aja PCI card or device. These will give you the option of 10bit YUV or RGB output (monitor dependent). 4K capable versions of these are available for as little
as $200.
That link is for 10-bit DirectX output which is for games and Direct X apps – not relevant to this discussion. Adobe and most other pro post apps don’t use or support DirectX. And just because a driver or monitor flags 10-bit does not mean
there is actually 10-bits happening.
10bit display support is all over the place on mac. MBP 2016 is driving a 24” external HDMI monitor in 10bit will driving the native display in 8bit. recent model imacs with 5K displays running in 10bit.
>>> Unless you NEED 10 bit color output (unlikely) do not get a Quadro. They are intended for industrial CAD / CAM workflows and are 1/10th as fast for rendering CUDA Redshift. They are intended for use cases where you need extreme floating
point precision, like designing critical infrastructure. Nobody is using Quadros for Redshift / Octane.
I disagree that 10-bit is “unlikely”. It’s more of a ignorance of benefits that an unlikely scenario. I get a lot of pro colorist and post house delivering stuff with banding, too much shadow noise and color/gamma problems that have obviously
been graded in 8-bit on typical 1000:1 panels. See it all the time.
Same as I see people mixing on headphones or uncalibrated speakers and charge pro prices for the service. The defense is always the same “looks/sounds great here!”.
GeForce cards cannot display video in Premiere or images in Photoshop in 10-bit, period or most any other pro app. You have to have a monitor card going to a separate, second monitor. That card or box + monitor for a 4K setup is not cheap.
Quadro cards offer native 10-bit OS level support with Quadros and Mac is now fully supported.
Most people think they do don’t need 10-bit output but end up delivering work that looks bad in 10-bit, 12-bit and HDR environments or displays with wider than the 8-bit (and 6-bit+dithering panels out there).
If you’re delivering 10-bit and while it used to be rare, these days almost everything I deliver is 10 bit or higher.
So I just don’t take people mastering final projects for professional deliver very seriously that don’t have a fully 10-bit or higher workflow.
Quadro plus a good affordable HDR 4K 10-bit monitor is the same or less money than a GeForce + 10-bit video output plus dedicated 4K display and offers better drivers, more stability and more flexibility (like Photoshop work).
Thanks for the input. It has confirmed my intuition to stay away from the Quadro and Xeon for this build. Looking at either 2x 2080's or 1x 2080ti now. Don't have funds for 2 of those. :-(
No need for 10bit output here. Mostly making HD TV promos on this machine.
okay I'm seeing a lot of mixed advice. The problem here is that there is a major real-world consequence of getting the wrong type of processor these days, and it is REALLY important to buy the right system for YOUR workflow
- Unless you NEED 10 bit color output (unlikely) do not get a quadro. They are intended for industrial CAD / CAM workflows and are 1/10th as fast for rendering CUDA Redshift. They are intended for use cases where you need extreme floating
point precision, like designing critical infrastructure. Nobody is using Quadros for Redshift / Octane.
- The same goes for Xeons. Do not get a xeon system. Many builders try to sell you on xeons because they can make more markup on it. They are extremely expensive and offer little to no performance benefits vs. i7 or i9. This is basically
due to A) lack of overclocking ability and B) way more expensive ECC RAM required (not necessary). The large number of cores means each core is slower than the equivalent i7/i9. AE and most tasks in Cinema are still single-threaded. The exception is if you
plan on doing a lot of CPU rendering in Arnold (which is going GPU soon anyway), or high end simulation work in houdini etc.
- Basically - the more cores you have, the slower each core is, this is based on thermal and manufacturing limitations. AMD TR chips are good but the same applies - only necessary for pure multithreaded CPU tasks.
- Currently, for future-proofing, the only GPU I'd look at right now is the GTX 2080ti - but try to find the "blower" models and not the "open fan" designs, as those tend to just push hot air around inside your case.
I'm not going into a specific parts list but from what you're saying I'd be looking at a system with an i9 chip with 8-12 cores, high megahertz boost / OC abvility (4.2-5 Ghz) very good all in one liquid cooling (a 360mm radiator is essential),
and one or two GTX 2080ti's. And yes, RAM prices have blown up and stayed there due to manufacturing shortages, supposedly.
I am currently rocking a machine I built with the core i7 7900X, 4x 1080ti, 128 GB RAM for similar workflows and it does great. I will say avoid Asus motherboards for now, they can have issues
_____________________________ Teddy Gage President Shotgun Inc. 360 VR | MOGRAPH | VFX BKLYN NY 11215
|
|
