If you pay peanuts, you get monkeys. If you pay money to monkeys, you get rich monkeys.

That holds true for politicians, but also for editing systems. It does not mean that you have to spend a fortune on a good editing system, but at the same time we have to realize that video editing is very demanding in terms of the muscle required of a system to perform nicely. How much muscle is required? Well, that depends to a very large extent on the codecs used and the video editing style. Let me repeat two images I have used in the Disk Setup blog,

 

and

Whether the budget for a system is $ 2000, $ 5000 or even $ 25000 is not relevant for the topic at hand here. What it is all about is how to make sure that a certain system is 'balanced'.

OK, tell me what is a 'balanced system'?

A balanced system is a system where the capabilities of each major component closely matches the capabilities of all the other components, without overspending on one component, that is overqualified for the task thrown at it or vice versa.

You know I like to use analogies quite often to make my point. Well, an example of overspending on an overqualified component is installing ceramic Brembo disc brakes on a Toyota Aygo. The other way (underspending and under-qualified) is to install this 1,0 litre VVT engine from the Aygo in a Bugatti Veyron, far underpowered for such a car. Let's give you another analogy to emphasize this point.

First we have to make a detour to a very popular nightclub, which attracts lots of people every day. What is the main profit generating part of a successful nightclub? Correct, it is the bar. It is the Common Profit Unit, the CPU. The greater the efficiency, the less beer is lost on draughts, the more accurate  the fill rate of a glass is, the capacity of the cellar installation, the speed with which one can fill orders,  they all directly influence the bottom line. And it helps quite a bit if the bar, CPU has a Sustainable Service Entry (SSE) close by in order to restock the bar.

Of course, it makes a huge difference how the dance floor, the music, the lighting and audio-visual effects are used, because that determines how people percieve this nightclub. If the dance floor is too small it will be detrimental to business, if it is too big, people will feel abandoned, if the music does not sound good it will hamper business, because people will leave earlier, instead of ordering another drink. This is the General Public Unit, or GPU.

The floorplan is of course critical for the operation of this nightclub. If people have to walk great distances to get  from the dance floor (GPU) to the bar (CPU) for orders, if the physical space limits the numbers of visitors, that all has a negative effect on the bottom line. It always boils down to More Entries, More Orders and Reduced Yawns, in short MEMORY.

But first you have to get the customers in and keep them inside as long as possible. All these customers waiting to get into the nightclub are your Dear And Trusted Accounts, your DATA. They form a long waiting line in front of the door, where a doorman arranges Sequentially Allowed Turnstyle Access (SATA).  Obviously, the more doors you have for access, the shorter the waiting lines and the faster your customers, DATA can reach the bar, CPU to order a drink and then mingle among the other guests. MEMORY.

However, people come to the nightclub in order to have a relaxing and enjoyable evening, so from time to time they want to use Dedicated In-house Seating Comfort, DISC to get their breadth after a long dance, enjoy their drinks in comfort and wait for the next piece of nice music to dance again.  With obesitas rampant, the seats, DISC, need to be large and roomy, not small portable seats that offer no comfort at all. Now, that nightclub offers a Complete Area for Serious Enjoyment, CASE, and the bigger the area, the more room there is for different seating places, DISCS.

If the clientele often comprises VIP's and celebrities that require special attention and service, it migh be a good idea to have a Reserved Area with Immaculate Decorum, RAID, to cater for these special guests.

Now, the thing to keep in mind for a nightclub to be successful and profitable, they have to keep things in balance, keep things in perspective and carefully consider what logistical requirements there are, depending on the location and clientele.

If the club is located in a small county in a rural area, the demands are far different than for a club at the corner of 48th St. and 5th Avenue in NY. Different crowds of people, different numbers of VIP's and celebrities, different square foot prices for the size of the club, different requirements for the bar, different drinking habits, beer versus cocktails, different wine tastes and different prices for drinks, different audio-visual requirements, different kind of music and different seating arrangements.

There is not a single recipe to make the nightclub successful and profitable, but with some careful analysis and common sense, it is not very hard to come up with the  correct business plan for this club.

So where is the similarity with video editing? All over the place...

With some imagination - and we are all creative people around here - you can easily recognize the similarities between the logistics of a nightclub and a NLE system for editing. It is all about balance...

Basically, these are the major components used in an editing system:

 

This looks pretty simple, like spokes on a wheel. Note that if a raid controller is not a practical solution, the wheel will simply have one less spoke. All the components have around the same size, they are around equidistant from the centre and they all look in balance, but you can see what happens when some components are out of whack.

If one has a nicely balanced system with a fast CPU, lots of memory, a fast CUDA capable video card and a good disk setup, it is quite easy to completely destroy the capabilities of such a system by not heeding the following warning:

Now, let's continue with 'balanced' systems. Make the following changes from the 'balanced' system above: Leave out a Raid Controller, use a minimal Disk Setup and install an overqualified Video Card and the picture looks like this:

 

 

It is no longer balanced, the components are no longer equidistant from the centre. The wheel is no longer round and does not turn easily. Have you ever tried riding a bicycle where the spokes have different lengths? This means you have overspent on the Video Card, which is idle most of the time, waiting for the disks to play catch up. This is pretty easy to understand in theory, but what does it mean in an individual case?

Let's start with the most obvious component, that raises the bulk of questions, the video card. Note that hardware accelerated video cards only help accelerate partly, it does not help in a lot of cases. See Cuda Mercury Playback Engine. Third party effects like RedGiants Magic Bullet or Colorista do not benefit at all from hardware acceleration.

What about the amount of VRAM?

As a rough guide, how much VRAM should be installed on the video card, think along these lines:

  • For SD material, 1 GB is enough
  • For HD material, go for 1 - 2 GB
  • For 3K or 4K, go for 2 - 4 GB
  • For 5K+, go for 4 - 6 GB

Keep in mind that the requirements can vary significantly, depending on the complexity of the timeline. If it happens that the VRAM is depleted (not enough memory), hardware MPE is automatically turned off rather ungracefully and stays off for the rest of the encoding. If you use two video cards in CC, each card uses its own memory and cannot use the memory on the second card. The encoding process works like this: Frame 1 rendered on card 1, frame 2 rendered on card 2, frame 3 rendered on card 1, etc. If either card runs out of memory, again hardware MPE is turned off. Simply put, two video cards with 1 GB each can not be treated as a single card with 2 GB, so the risk of turning off hardware MPE automatically is bigger with these two cards than with a single card.

Explanation

It may require some explanations, why these results are shown as they are, because there are a number of implicit considerations about balanced systems in this table. Let me try to explain them.

Like any bicycle rider, you want your wheels to be completely round, all the spokes to have the same length and to be balanced. The same with a computer system. If you happen to have an AMD CPU, you are out of luck. The lacking SSE support make them unfit for editing with PR. No matter how much memory is installed, how fast your disk I/O set-up is, they are just too slow to edit effectively, even in the 8-core variant. If the Intel CPU lacks hyper-threading, the limited number of logical cores are a real bottleneck in processing speed. Such systems are not worth to have more than 16 GB of memory installed, since it brings no performance gain at all. All systems with such processors can not effectively use a dedicated raid controller due to the limited number of PCIe lanes, so transfer rates are largely irrelevant on such platforms.

When one moves up to a quad core with hyper-threading, memory gets more important, but also the transfer rates of the disk I/O system. However, when talking about a Sandy Bridge, Ivy Bridge or Haswell CPU, there is no practical possibility to use a raid controller, so higher transfer rates can only be achieved with raid0 configurations. For hexa cores these higher transfer rates are quite normal when using large parity raid arrays on a dedicated controller.

The general rule in the table above is that when CPU performance rises AND memory increases AND disk performance goes up, then a better video card is also warranted, to keep everything in balance. It makes no sense to improve only one component if the rest of the components is not up to the task.

Note that there are no Quadro cards included in this overview for the simple reason that all Quadro cards are way overpriced and underperforming in comparison to the GeForce cards. The ONLY reason one may choose a Quadro card over a GeForce card is when 10 bit output to a costly 10 bit monitor is required. See the graph below. Based on August 2013 prices in Europe excluding VAT and only comparing to GeForce cards that perform equal or better than the Quadro card in question. The Green bars show the relative performance index of the GeForce card when compared to the Quadro card. So for example, the GTX 760 is 11% faster than the K5000, despite being almost 1/8 of the price. The Red bars show how much more the Quadro cards cost compared to similar performing GeForce cards.

Based on current € prices in Europe excluding VAT, a typical well balanced single CPU system will use around the following percentages of the total budget. It does not reflect the absolute price, only the relative cost in % per category and is based on a single video card.

Note that this applies to hexa core systems using a 2011 socket motherboard. If we look at Haswell, Ivy Bridge or Sandy Bridge systems, of course there is no dedicated raid controller because of the limited PCIe lanes, so the typical well balanced system will look more like this:

 

 

 

Add comment
  • Do you have a similar viewpoint on building a larger 'rendering farm' setup -- something where the pairing of 8 GPUs to a server 'sweet spot' would be and the implications now of faster storage now such as NVMe based SSD. I might be interested in having you commission such a analysis for my 'day job' -- not the email address listed here. This would be more of a data center setup for workgroups vs an individual workstation as you have put together here.

    0 Like
  • Ver, very useful information on these pages, congratulations. I have seen in the table above that I am in the file of Intel Core 6 with HT, since I´m planning to build a PC with an i7 5820k processor. For my surprise, I have seen that the card I was going to buy (GTX 780 Ti) is far from something balanced. But, in a note, you recommend that If we use CC and lot of preview, maybe the best option is to buy 2 cards. So, do you recommend to buy, for example, 2 cards like GTX 650 Ti boost (2x180 euros= 360 euros, with 768 cuda cores each) instead of, for example GTX 680 (around 400 euros, with 1536 cuda cores)? The performance is going to be better? Thanks. I´m very interested in previewing performance of my system.

    0 Like
  • Guest - Mark VanDruff

    Which is the better CPU for the warrior system, the Intel Core i7-4960X Ivy Bridge-E 6-Core 3.6GHz or the Intel Core i7-5960X Haswell-E 8-Core 3.0GHz LGA 2011-v3?

    0 Like
  • Guest - DDoggass

    Does this also apply to animation with After Effects and 3ds max? Thanks

    0 Like
  • Guest - RRadler

    AMD CPUs lacking SSE? That statement is outdated. AMD supports SSE since K10 and the newest CPUs have also SSE 4.2 support.

    0 Like
  • That statement is still accurate today. AMD has never implemented the full 4.2 instruction set, only a part of it and what they did include is badly implemented.

    Lack (verb): be without or deficient in.

    0 Like
  • This is very timely for me because I am now looking into upgrading from a Sandy Bridge system. I think I am pretty well balanced and great for HDV, but now that I'm working with DNxHD and DSLR footage I can almost feel the strain. I have read some items recently that dispute what you have to say about the NVidia Quadro video boards that Adobe pushes. I will be checking in to learn more.

    from Henderson, NV, USA
    0 Like
  • Maybe this comment is not for this page, but don't know where to post it.
    Recently I browse through PPBM6 results and saw something that I don't understand. On the 9th page there is quite an average home system called HELGES-DATOR with mpg2 and h.264 rendering times significantly faster than any other so much better hardware. Overall result isn't spectacular, but so fast rendering is the thing that confuses me. How it's possible that such "old" system is working so fast?

    0 Like
  • I have asked him to rerun the test, because of the unbelievable results.

    0 Like
Powered by Komento