My PC Rebuild — Have at it!

[Stephen Mann]

In the Rendertests, the dual Xeons blew everything else in the smoke.

SSD’s do have a write cycle limit, but PC Magazine did some tests and you will be upgrading a long, long time before that limit starts to bite you. But, yes, keep the Swapfile, Temp and Documents on other drives.

Green drives. NOOOO!! Not for an editing PC. The last thing you want is for the drive to shut down (green) just as the NLE wants data, and errors out because your media disappeared.

X79? That is so last year. I just built a new editing PC with the Gigabyte Z87 mobo, an i7-4770K, 32Gb of RAM – all new components for a total of $1300.

I’ve had nothing but problems when I ran Radeon display adapters. But if it works for you…

Steve Mann
MannMade Digital Video
http://www.mmdv.com

[Dave Haynie]

[Stephen Mann] “SSD’s do have a write cycle limit, but PC Magazine did some tests and you will be upgrading a long, long time before that limit starts to bite you.”

As long as you treat them right, that is. I do expect to see the drive replaced long before it’s getting old — unlike HDDs, solid state drives really can follow Moore’s Law. It’s also very likely we’ll be seeing alternative memory technologies in the next five years, which will address the longevity issues.

As for the WD Green drives, they work fine in editing systems. They really don’t randomly shut down like people think. They’re even supported in the Drobo RAID. And keeping cool, they last pretty much forever… I had enough of Barracudas dying early due to overheating.

[Stephen Mann] “X79? That is so last year. I just built a new editing PC with the Gigabyte Z87 mobo, an i7-4770K, 32Gb of RAM”

The Z87 is limited to LGA1155 CPUs = only two memory channels and only 16 PCI Express lanes. X79 supports the Sandy Bridge E chips on Socket 2011, four memory lanes and 40 PCI Express channels. Haswell E CPUs and system chips will certainly come out, eventually (8-core CPU, X99 chipset, DDR4 memory, etc. announced for “late 2014”), but I was looking for performance today.

[Stephen Mann] “I’ve had nothing but problems when I ran Radeon display adapters”

I actually bought both when Vegas went to full GPU support, both around $300. The Radeon was faster with Vegas, and the nVidia had bugs in their OpenCL implementation. I’m sure that was a temporary thing, but never had any problems (I was using nVidia before that). I figure I’ll wait on the GPU until there’s some compelling reason to upgrade — given the CPU, it’s unlikely today’s Vegas will see much help from the GPU, unlike with the much slower AMD system.

-Dave

[Rick Shorrock]

Either way, that’s a sweet system. I used to edit on an AMD dual-core system…slowly! The CPU was slowly overheating, so I bought an AMD APU (CPU and GPU combined)- based kit. It’s got the AMD A10-5800K Trinity 3.8Ghz APU with DirectX 11 Graphic AMD Radeon HD 7660D, MSI FM2-A75MA-35 Micro ATX motherboard, 16Gb DDR3 RAM and LITE-ON DVD burner. And the kit came with an ultra high gloss Micro-ATX case with 400W power supply. I spent less than 350 dollars for this kit, with free shipping from NewEgg. And it appears that Vegas approves of the APU! If I render my timeline (on Vegas Pro 12) with OpenCl or CUDA assistance, the timeline renders significantly quickly!

[Dave Haynie]

Ouch! I had a dual-core AMD system ages ago. In the early day of HD, it was taking something like 28 hours rendering for about hour’s worth of HD video. I knew something was up when I got a new, slower clocked Intel dual-core laptop that rendered the same thing in 24 hours. I moved to quad and then hex core after that on the desktop.

The AMD hex core wasn’t bad when it came out, for the money, but you could still do better with Intel…. of course, all without going to multiple socket. That’s an option, but a very pricey one.

The A10 is a good value; I got my son an A10 powered laptop for Christmas last year (for grad school — he’s got a i7 desktop for gaming), and it was surprisingly good for the money.

The concern I’ve always had for any of the new AMD architecture is specific to media processing — the Bulldozer/Piledriver architecture is based on core modules. Each module basically contains two integer cores and one FPU… so while your A10 has four real CPU cores, there are stalls in accessing the FPU hardware once your code gets too FPU intensive. So their high-end processors run up to 8 cores for desktop, 16 cores for server… but only half the expected floating point hardware. Initially, the octal-core chips often lost going head-to-head with the older hex cores, like my 1090T. They got past that, but mostly by going to insane speeds (single thread burst on these goes to 5GHz) and faster memory… but still no quad-bus memory.

The APU idea, on the other hand, is a pretty good one, especially for mobile. If you look at some of the A10 reviews, AMD actually beats i5s and even i7s on graphically-intensive benchmarks and Intel’s GPU. Being on-chip, there’s much less communications latency between CPU and GPU.

One of the problems with GPU computing is that, for any given GPU, there’s a fixed overhead of compiling the OpenCL, then sending compute jobs to the GPU. Increasing the amount of work you can send to the GPU each time, increasing the GPU speed, and decreasing the latency can all make your GPU help. On the other hand, if that overhead plus computation would take less time on a really fast CPU … and as long as you don’t have other work for that CPU you can pipeline (one of the main problems in rendering engines these days, I believe), then you’d be better off just using the CPU. But with that relatively weak CPU (speaking in terms of workstation-class computing power) and wicked fast interface, the GPU should be a good thing. And as I mentioned above, I’ve had very good success with the PCIe-based AMD GPUs, at least as of a few years ago. These things do tend to bounce between AMD and nVidia, who’s got the crown this week, I can’t say.

-Dave

[Rick Shorrock]

Wow…you apparently know your stuff. I had been reading that AMD was going to start combining CPU with a GPU, but had never thought of buying one till I saw the deal on NewEgg’s site. But it seems to work well for me with editing, so with the price being so low, I’m pretty happy.

[Dave Osbun]

Xeons are designed mainly for server use, but high-end workstation use as well. In regards to render times, Xeon workstations will easily outperform any Intel i7 processor.

If the budget permits, and you do this for a living, Xeon CPUs and compatible motherboards are the way to go.

Dave

[Dave Haynie]

[Dave Osbun] “Xeons are designed mainly for server use, but high-end workstation use as well.”

No. Xeons are designed entirely for server use. They’re applicable as well to certain kinds of workstation-class work, but I rekon Intel doesn’t spend ten minutes worrying about workstation-class performance when designing the next Xeon. Just too small a market, and it’s inherently going to be served by the intersection of server and “enthusiast” processors.

[Dave Osbun] “In regards to render times, Xeon workstations will easily outperform any Intel i7 processor.”

I don’t deny that there exists a dual socket Xeon workstation that’ll outperform the top i7 on renders. But you’re paying dearly for it. Take the Xeon E5-2687W 8 core, runs around $1900. Per core, this is slower than the i7-3930K 6 core (about $500). Sure, I could get a little more performance out of pure SMP with that chip in my main board, and it’s one of the Xeons (intended for blade servers) that works with commodity DDR3, rather than slower, way more expensive server registered for FBDIMM RAM (which is also more reliable, buffered to allow huge memories… but I’m at 64GB, enough for now).

And yeah, the Xeon has 20MB L3 cache, versus the i7’s 12MB L3 cache… but you know, media applications don’t bloody care about all that L3 cache. In a server, you may have hundreds of relatively minor threads, each running different code. In the main use of a media PC, you have one bit of code, optimized to fit in the L2 if not L1 cache of a commodity processor, running the same on each core… and data that’s just there for a visit, not really needing much cache at all… stream in, stream out. So sure, stack up two of these and you’re going faster.

But you’re probably paying 4x-6x as much for that 2x-ish speedup, going into server market pricing. That’s what I was getting at above. If you look at the PC market, there’s a price-performance curve, and there’s usually a pretty distinct “knee” in that curve. Below the knee, 2x as much money gets you 2x the performance. Above the knee, performance per dollar drops off exponentially.

And then there’s the reality of the market. It may not always be so, but it’s true so far that the top performance hardware doesn’t retain value. Intel’s top enthusiast processor… the top socket 2011 i7 these days, the i7-3960K, runs around $1000. When that’s replaced by one of the i7-49xx series forthcoming, it’ll drop to $500-ish pretty quickly, as long as Intel’s upping the bar still. This happens about every six months. So when you look at Xeon, sure, if you can really justify the expense of a $4000+ system every 2 years or so, go for it. But with a much longer upgrade cycle, you could go to the higher volume enthusiast chips, and have yearly upgrades for the same money.

For that matter, you can build a 64-core Opteron system from “regular parts” (any number of Supermicro main boards, Opteron 638x/637x series)… but each core is less than half the performance on an i7 core. So yeah, you can make a much faster PC, but you’re paying exponentially more for that performance.

[Dave Osbun] “If the budget permits, and you do this for a living, Xeon CPUs and compatible motherboards are the way to go.”

I actually make my living designing computer systems… embedded stuff, these days. I did desktops back in the early days of computer video… in fact, some of those kind of defined computer video back in the 80s. I mean, sure, if you can justify it for business and take a 179 deduction on the PC in a given year, that might be the right way to do it. Particularly if rendering is your only goal and your income really comes down to dollars per hour rendered.. it’s easy to see that $4000 pay for itself. But not everyone has that kind of budget.

I also have other needs: I use my system for video, photography, music, and CAD. Video drives CPU performance and HDD size, photography drives memory size, and the CAD stuff just hangs on for the ride.

-Dave

[Dave Haynie]

Hey Stephen… just what would you recommend in nVidia for ultimate performance these days? I took a look at current nVidia and AMD GPUs, just to get a sense, and nothing seems a dramatic improvement over the HD6970 I have today. Having the new system nearly all together now, did make me think on the GPU thing.

I’m not looking for “pro” cards… I have a Quadro at work. Probably a big win if you’re doing mechanical CAD, but kind of pointless otherwise, particularly at $2K or so. And yeah, there seem to be intentional slow-downs in some OpenGL ops on the GeForce cards, to make the Quadro seem much better for very specific CAD work. But for video and particularly OpenCL, I don’t see a win here in Quadro, particularly for the money.

Also decided to make the data drive RAID5. I have a controller, if the on-board isn’t sufficiently fast. For most of the things I’m doing, particular video rendering, I value read performance and reliability, don’t care so much about write performance (can always render to the SSD, an external drive, a project drive, or even an SD card, if this was a real issue…).

-Dave