Computer upgrade time

[Randall Leong]

Thanks for the encouragement.

My render was a semi-failure, however: Although it completed the render in just over 16 hours, the result was an unacceptably flickery video. It seems as if DVDA does not properly handle lossless compressed or uncompressed AVI in any color space other than 4:4:4 RGB: When I stepped through the AVI file, some of the frames were blank. Thus, if I want to preserve the original YV12 (4:2:0) color space, I would have to somehow transcode it to a lossy compressed MPEG-2 or AVC type format with a minimum of conversions before I import it to DVDA. Unfortunately, Vegas and most other NLEs convert YV12 content to RGB as the video is being imported into the NLE's timeline, then re-convert the color space to whatever is specified in the final format. This can adversely affect color reproduction quality.

[Randall Leong]

Originally Posted by Craig Coston

I know a few people have said that they like to run at stock speeds for stability, but I really feel that you are underserving yourself if you do not overclock the 920. I have been running at 4.2ghz (on a Noctua D12 cooler) and have not had ANY problems in Premiere CS4, and the performance jump is absolutely huge.

If you go to http://ppbm4.com/Benchmark.html you can see the performance difference I got in Premiere compared to a stock 2.66 speed on that chip. My machine (ranked second, the wildfly one) benched higher than a dual Xeon W5590 system, which is pretty cool. I too was in the "better safe than sorry" boat in regards to overclocking on past processors, but Intel really has stepped up the stability of their processors when overclocked, and it's almost expected that the 920 will be overclocked.

Good luck with it, and I hope you enjoy the new system. I'm having a blast with mine!

Right now I'm still on the stock boxed processor cooling. Unfortunately, with the stock cooler, I could not get the processor to operate reliably at 3.8GHz at stock voltage (1.2V); Windows crashed on boot-up at those settings. I had to bump up the core voltage to Intel's recommended maximum of 1.35V before I could boot into Windows reliably, and then I stopped my overclock at 4.0GHz. I do not want to risk destroying a $290 (USD) processor, especially in the hotter months of summer.

I am currently running the 64-bit version of Prime95 on that system at 4.0 GHz. If it passes the Prime95 stress testing for several hours, then I might still have more overclocking headroom. But if it crashes in less than an hour, then I have taken the overclock a bit too far.

UPDATE #1: Looks like I had taken the overclock a little too far on that stock cooler. Prime95 testing showed half the workers stopped working at 4.0GHz. I am retesting the system at a more conservative 3.8GHz, and report my results later.

UPDATE #2: At 3.8GHz with 1.35V VCore, it was stable for an hour (no errors; however, I stopped the test after an hour due to heat issues, as mentioned below) in Prime95 - but the core temps as measured in RealTemp were right at the maximum. So I backed it off a bit, and then experimented with overclocking at the stock 1.2V VCore. At that VCore setting I could only muster about 3.5GHz before Prime95 reported errors.

Based on these results, assuming that I have the turbo switched on I feel safe with an overclock to only 3.44GHz at the higher 1.35V VCore. At the lower 1.2V VCore, I might have to settle for an overclock to only 3.2GHz for best reliability. (And these results are with the stock boxed Intel CPU cooler.) If I decide to use 3.5GHz @ 1.2V or 3.66GHz @ 1.35V, I would have to disable the Turbo feature for best stability.

The conclusion in my case is it's the stock Intel boxed CPU cooler that's holding it back. But these overclocking results were actually quite good without having to spend a single penny more for an optional high-performance cooler.

[Craig Coston]

Have you tried bumping up to 1.25 or something incrementally instead of heading right to 1.35? I don't have much experience with that Intel board. Almost everyone has been throwing the 920 onto the Asus P6T platform (one variation or another).

Have you seen the overclocking the 920 walkthrough that Harm Millaard found and posted?

Adobe Forums: Overclocking the i7, a beginners guide

[Randall Leong]

Originally Posted by Craig Coston

Have you tried bumping up to 1.25 or something incrementally instead of heading right to 1.35? I don't have much experience with that Intel board. Almost everyone has been throwing the 920 onto the Asus P6T platform (one variation or another).

Have you seen the overclocking the 920 walkthrough that Harm Millaard found and posted?

Adobe Forums: Overclocking the i7, a beginners guide

I read that. As it is (at 3.5GHz and 1.20V with the stock Intel boxed CPU cooler), I'm already getting temps in the mid-70s C at 100% load - and any raising of the core voltage (VCore) would raise the temps further even at lower core clocks. Intel recommends a maximum TJunction (core temperature) of 70°C under continuous conditions.

I am also checking the memory settings. It seems like I had the QPI/Uncore voltage (VTT) also set at 1.35V (as the XMP memory profile dictated). I will reset everything back to stock for about half an hour until the CPU has cooled down sufficiently, and then redo my overclocks.

UPDATE: Upon further checking of that guide it seems like 3.5GHz is about the highest overclocked speed that I can safely go to with the stock Intel cooler. This is because the core temps at that speed get to just above 70°C at 100% load. (Speeds higher than 3.5GHz with the stock cooler are stable under Prime95, but causes the core temps to rise to potentially harmful-to-components levels - above 80°C - under anywhere near a full load.) In other words, the Intel board is very capable but the stock CPU cooler is holding it back (in the case of my particular system).

Gosh, I need a better CPU cooler before I can overclock any further than 3.5GHz...

As an addendum, the DX58SO is currently the only "consumer" Intel-branded motherboard for Socket 1366. The other Intel-branded Socket 1366 motherboards are sold for workstation and small-server use and may provide incomplete or no tweaking capabilities. The DX58SO is part of the "Extreme Series" of Intel-branded motherboards, which also includes a standard ATX and a microATX motherboard for Socket 1156 with the company's P55 chipset (however, I have been unable to locate either of those other two motherboards which tipped the scales towards 1366 for me). Intel also distributes various lower lines of motherboards including a microATX and a standard ATX Socket 1156 P55 motherboard in the Media Series. Unfortunately, none of the Media Series, Classic Series or Essential Series lines of Intel-branded motherboards provide any overclocking capabilities at all whatsoever, and most motherboards in those three lines provide minimal (if any) memory or voltage tweaks.

And the biggest reason why I picked Intel over Asus or GigaByte is the LAN. The boards from Asus and GigaByte people have talked about all use a Realtek LAN chip for their primary LAN ports (and I have had nothing but bad luck with Realtek LAN controllers in the past, such as very slow transfer speeds and recalcitrant wired connection reliability, which made me very leery about buying anything with a Realtek LAN controller to this day). The lower-end Intel-branded motherboards also use Realtek LAN controllers - but the Media and Extreme Series lines of Intel boards use the company's own (IMHO superior) LAN controllers. (Speaking of Asus, the last board I purchased from that brand had an Attansic L1 LAN controller chip, which had compatibility issues of its own, and also has especially poor driver support since Attansic was bought out by Atheros back in 2007.)

[Craig Coston]

If you decide to drop money on an aftermarket cooler, my strong suggestion is the Noctua NH-D14. It's absolutely amazing. Make sure to use their provided thermal paste, as it is better than Arctic Silver's offerings when paired with this particular cooler (my own tested experience). Just make sure you have enough room in your case for this one. I have everything in a 4U height rackmount and had to dremel the top of my case to get it to fit right.

[Randall Leong]

Originally Posted by Craig Coston

I know a few people have said that they like to run at stock speeds for stability, but I really feel that you are underserving yourself if you do not overclock the 920. I have been running at 4.2ghz (on a Noctua D12 cooler) and have not had ANY problems in Premiere CS4, and the performance jump is absolutely huge.

If you go to http://ppbm4.com/Benchmark.html you can see the performance difference I got in Premiere compared to a stock 2.66 speed on that chip. My machine (ranked second, the wildfly one) benched higher than a dual Xeon W5590 system, which is pretty cool. I too was in the "better safe than sorry" boat in regards to overclocking on past processors, but Intel really has stepped up the stability of their processors when overclocked, and it's almost expected that the 920 will be overclocked.

Good luck with it, and I hope you enjoy the new system. I'm having a blast with mine!

Craig,

I did some additional experimenting with my system. It turned out that an overclock to 3.5GHz with the Vcore and VTT voltages only turned up slightly to 1.2V from the stock 1.15V and the FSB:DRAM ratio remaining at 2:8 (this results in my memory now running at an effective 1400MHz instead of the stock 1066MHz) is the best compromise between performance and stability/heat (with the stock cooler). I did short Prime95 runs while also running TrueTemp with all three combinations - stock CPU speed and stock 1066 memory speed and timings at stock voltages, stock CPU speed but using XMP-1600 memory settings and the aforementioned overclock mentioned above. It turned out that the overclocked-to-3.5GHz config ran only 2°C higher on average at 100% load (in terms of the average of the temperatures of all four CPU cores) than the stock setup with stock memory! In fact, running a stock-speed CPU with overclocked XMP memory actually produced higher temps than the overclocked setup. 3.6GHz booted into Windows at these same voltage settings, but Prime95 testing throttled back one core - this meant that I would have needed to raise the Vcore and VTT a little higher than the 1.2V that I used at 3.5GHz in order to get things stable at 3.6GHz (which, in turn, would have raised the CPU core temps even higher, necessitating the use of a higher-performance CPU cooler for such a level of overclocking).

As for the actual Blu-Ray re-encoding from XDCAM EX MXF to AVC in DVD Architect, re-encoding 2 hours worth of content took nearly 4 hours longer with the stock-speed setup than with the overclocked setup - 18+ hours versus just over 14 hours. (The stock-speed performance is nearly as slow as my old Q9450 with 4GB of DDR2-800 memory could muster; that older setup now resides in my auxiliary system which is currently in use for some of my standard-definition video work.) And yet the total CPU utilization barely reached 25% even with a stock 920.

I also looked at the table of rendering performance with the different systems. A configuration that's closest to that of my old Q9450 setup did respectably well in the tests (a Q9550 on a non-overclockable Intel-brand DDR2 motherboard with a G45 chipset that's similar in performance to the P35 chipset Intel-brand motherboard which my Q9450 resides on), in fact beating a few of the stock i7-920 configurations. And an Asus motherboard is no guarantee of superior performance; other components such as multi-disk RAID setups also factor into the overall performance picture. Most (though not all) of the very top-performing systems used AMD/ATi HD4870 graphics cards (I have an HD4850, one step below the 4870, in my overclocked 920 system).

[Craig Coston]

Have you ran a speed test on your drives? You should be getting much higher than 25% utilization during encoding I would believe. I am not a DVD Architect user though, so I don't know how that software utilizes the processor.

As far as Asus/superior performance subject, I believe that Asus delivers a product that is very easy to configure to get the maximum performance and they are very reliable. Yes, a motherboard does not make or break system performance. Bottlenecks can occur in many different areas, drives being of critical concern.

And video cards: I've built an i7 965 machine with the 4870, and it was good. I like my GeForce 285 2GB better though =) Especially when I factor in the Mercury Playback Engine by Adobe, which will not be supporting ATI graphics cards.

[Randall Leong]

Maybe this is why I was not seeing as much of an improvement of a performance as one would expect: Almost all of the other systems on that list have at least one additional drive outside of the internal Windows boot drive. (Sadly, consumer external hard drive kits are ill-suited to being the project disk due to their not only being slower than internal drives in external drive enclosures, but also due to their power management features which often cannot be disabled at all.) My times were based on just using the same Windows boot drive for both my project and the system files. Adding enough drives to make a substantial difference would have cost me far more money than I can currently afford - at least $1,200 worth of drives plus another $200 worth of add-in controllers. And $1,400 is a lot to spend just to get a relatively small amount of performance gain with everything else.

Also, the Vegas' and DVDA's Sony AVC rendering engine is relatively old (even in the most recently released versions, Vegas 9.0c and DVDA 5.0b): It still cannot take full advantage of the new features in the i7 processors; after all, it was designed during a time when the fastest consumer processor available was the Core2 Extreme QX6850. This is why a stock-clocked i7-920 came out as slow as a stock-clocked Core2 Quad Q9450 in the rendering tests (both processors run at a stock 2.66GHz). After all, Vegas and DVDA are better at MPEG-2 encoding than AVC encoding (performance-wise, even though lower-bitrate HD MPEG-2 renders produce lower image quality than AVC renders of half the given MPEG-2 bitrate). Plus, even the 64-bit versions of Vegas are still not fully 64-bit; in fact, all of the plug-ins which come with Vegas are still 32-bit.

That said, I would like to have at least one or two other drives added to my system. Unfortunately, the Intel ICH10R's built-in RAID feature requires that all hard drives connected to the SATA ports be RAIDed together - it's all RAID or no RAID. Although the improvement in DVDA recompression times are in line with what a mere increase in the processor's clock speed expected, the times are still quite a bit longer than it should have been due to the use of a single hard drive for everything.

By the way, you can still run CS5 on a system with a graphics card which does not support the Mercury Playback engine. (This includes all ATi graphics cards and all older and lower-end NVIDIA graphics cards.) It just means that the Mercury engine will not run with an unsupported graphics card, and the system will thus default to the slower standard Adobe playback engine. In addition, Adobe stated that 12GB is the minimum amount of RAM required in order to run CS5 with the Mercury Playback Engine; it will not run at all with less than that much RAM. And as Craig pointed out in another thread the Mercury Playback Engine will only work with an NVIDIA graphics card with hardware-based CUDA acceleration (so far, only the very high end of the GeForce GTX2xx line plus higher-end versions of the latest edition of the Quadro FX line support hardware-based CUDA acceleration although all GeForce and Quadro FX cards based on a G80 series or later GPU support software emulation of CUDA).

And given my current equipment even after the i7-920 upgrade, I will have to spend at least another $1,500 worth of hardware upgrades plus another $800 or so on Premiere CS5 just to even run CS5 in its full glory. This is because my motherboard effectively has only three DIMM slots; thus, I would need three 4GB memory modules in order to run CS5. (Technically, the motherboard has four DIMM slots - but using all four slots will force part of the memory to run in single-channel mode instead of triple-channel.)

[Craig Coston]

In your bios set your SATA controller to non-RAID mode. Add two identical drives (even 500 GB drives are fine, and you could get two of them for just over $100). Then use Windows to create a RAID 0. It will work wonders for you.

[Randall Leong]

Craig,

I did not even think about that. The separate software RAID will speed up whatever projects that I have running through my current NLEs.

Does it matter which RPM speed the two drives spin at?

[Craig Coston]

I would ideally try to use two identical drives, running a minimum of 7200 RPM.

Newegg has Seagate 750 GB drives for $59.99 with free shipping. Two of those would be fine (as long as you don't run them as part of a RAID 5... they have issues with that). If you can afford a third drive, that makes performance even better.

[Randall Leong]

Originally Posted by Craig Coston

I would ideally try to use two identical drives, running a minimum of 7200 RPM.

Newegg has Seagate 750 GB drives for $59.99 with free shipping. Two of those would be fine (as long as you don't run them as part of a RAID 5... they have issues with that). If you can afford a third drive, that makes performance even better.

That version of the Seagate (the 750GB 7200.12) is fine by me. But I found that the generation of the drive being used is almost as equally important as the spindle speed of the drive. After all, some resellers sell 320GB and smaller SATA drives cheap - but most such drives tend to be of significantly older designs which deliver less than half the average sequential throughput as the newest hard drive designs. In this particular case, striping two such older drives would still result in slower sequential speeds than simply using a single drive of the newest design.

I've just checked Newegg again, and that 750GB drive actually went up in price to $79.99 each (still with free shipping). The 1TB version of this drive is the same drive as this one; however, the smaller-capacity 750TB drive actually leaves one side of one of its two platters unutilized (the 1TB version utilizes both sides of both platters). The 500GB version has the same basic physical performance as the two larger drives in the series (because this version is a single-platter design), but with only half the cache memory. (The smaller cache amount - 16MB versus 32MB in the larger 2-platter versions - is perfectly adequate for such a smaller-capacity drive.) Forget about the harder-to-find less-than-500GB versions of this drive; they leave part of their single platters unutilized. Such partially-unutilized drives do have quicker access times, but may have slower maximum sequential read and write speeds than the drives which fully utilize at least one side of at least one of the platters: Depending on the manufacturer, some such drives utilize only the inner portion of the platters and thus the result is a crippled maximum sequential performance like I described earlier in this sentence. Others may use only the outer portion, which increases minimum sequential performance at the expense of usable capacity.

In the interim, I experimented with a 1.5TB external hard drive kit connected to eSATA. With that drive I found that the type of controller powering the eSATA port(s) to be a noticeable factor; in fact, that drive connected to my older Q9450's single eSATA port to be actually faster in sequential read and write performance than that same drive connected to an eSATA port of my current i7-920 motherboard. Upon further investigation I found that the eSATA port on my Q9450 system was actually connected directly via a cable running from the backplate to one of the internal eSATA ports of that system's Intel DP35DP motherboard, which uses the Intel ICH9R "South bridge". Hence, the eSATA drive connected to that older system used AHCI (and could have also been RAIDed together with other add-in hard drives had I chosen to use that function). The disadvantage to this setup is that I lost one of the internal SATA ports to eSATA, limiting the maximum number of internal SATA drives supported from six to five. But on my i7 system, the eSATA ports are actually run off of a separate Marvell eSATA controller chip on the motherboard, which supports neither AHCI nor RAID at all. As a result, with just a single drive connected to the Marvell eSATA port, the maximum sequential write speed is only about half as much as what the older system could muster (primarily because the Marvell controller is permanently stuck in the IDE mode) while the sequential read speed is also slightly lower than the older system. Hence, my quest for a true performance RAID add-in system for video editing/rendering work.

UPDATE: I did receive a pair of the Seagate 1TB 7200.12 drives this morning (I had actually ordered the drives before I asked which drives to get because I was a little disappointed in the somewhat slow rendering performance without the extra disks). I just installed and formatted the two drives with Windows 7's software RAID feature (the Professional, Enterprise and Ultimate versions also support RAID 1 in addition to RAID 0, while the Home Premium is limited to RAID 0). My initial testing with this software RAID 0 auxiliary disk setup (copying files from my system's main hard drive to the RAID array and secure-erasing large files using the freeware Eraser program indicates that my system's main hard drive had been transferring as fast as it could only to be beaten so far by the RAID array.) I will re-test my rendering later to see if I achieve a marked change. My results with the external eSATA drive connected to my i7-920 system's eSATA ports indicated slightly slower rendering times than simply using my system's hard drive for everything.

[Peter Telian]

*mis-posted*

[Randall Leong]

Originally Posted by Randall Leong

I will re-test my rendering later to see if I achieve a marked change.

Initially, I did not notice much of a difference between the RAID and the single drive - until I read the Videoguys' guide. They actually recommended putting the source media files and the rendered files on separate drives. This means that if the media files are on the system drive, the RAID 0 array should be used for your output. Conversely, if your source media files are on the RAID 0 array, the rendered output should be set to a different hard drive such as the system drive.

By the way, I have just updated my eSATA external hard drive tests on file copy transfers. It turned out that the Marvell eSATA controller on my system's motherboard actually supported RAID (in the form of the chipset manufacturer's software RAID) in addition to IDE mode. (With only two ports, the Marvell controller supports only RAID 0 or RAID 1.) This meant that the controller had been handicapped by the default Microsoft IDE driver, which limited write performance to about half of what the external hard drive kit was capable of in the first place. So, I went to the motherboard manufacturer's Web site, and downloaded and installed the correct Marvell driver. AHCI is automatically enabled when the driver is installed. The result is a doubling of the write throughput. Still, the external hard drive performed no better than my system (C:) drive in the read and write tests. And the Marvell eSATA controller's performance is good news to those who plan to do a RAID 0 setup using two SATA internal hard drives mounted inside eSATA external enclosures.

[Randall Leong]

Originally Posted by Craig Coston

Have you ran a speed test on your drives? You should be getting much higher than 25% utilization during encoding I would believe.

Even with the RAID 0 array as my source drive and a separate disk as my render-to drive, I've averaged around 30-something percent utilization, with a max of around 40 percent. This is likely because the current 64-bit version of Vegas does not make full use of the system's 64-bit capabilities.

I experimented again with rendering MXF files into AVC in Vegas. I used four different processor and memory speed settings: stock speed on both CPU and memory, stock-speed CPU with memory overclocked to 1600, CPU overclocked to 3.2 GHz with memory running at 1600 and CPU overclocked to 3.5 GHz with memory running at 1400. The result is that the overclock to 3.5 GHz is fastest, rendering a 16-second MXF file to AVC in about 72 seconds. The overclock to 3.2 GHz took 78 seconds to do the same work, while the stock-speed CPU with overclocked memory and stock-speed memory took 89 seconds and 92 seconds, respectively. What's more is that overclocking the memory alone produced a very minimal improvement in rendering performance compared to overclocking the processor.

In this respect, you are correct that overclocking the i7 processor can (and does) produce significantly improved rendering performance.