P4 3.6 or Dual Xeons?

[Kevin Shaw]

Eddie: if you look around the internet at various user reports, it becomes apparent that dual Xeons are very effective for things like rendering when used with software which is dual processor savvy. People with good dual processor computers report being able to do two-pass MPEG2 encoding in faster than real time (i.e. less than one hour to encode an hour of video). With dual core processors you may be able to get close to a similar level of performance without having to spend so much on a fancy workstation motherboard, and perhaps with less heat generated than with two single-core Xeons. And eventually we'll have the option to use two dual-core processors, which should accelerate things even further.

[Tim Bucklin]

Originally Posted by Glenn Chan

Tim,
Have you run benchmarks to validate your statements?

Absolutely. Sandra alone has shown higher performance (memory bandwidth) on dual opterons versus my 3.2E P4 and even the new Dual-Core Smithfield Extreme Edition (that's a dual-core, hyper-threaded 3.2GHz, 800MHz FSB)

Some of them I doubt:
Aggressive timings: I have tested this with Vegas and the sad truth is memory timings does nothing for Vegas.
http://www.dvinfo.net/conf/showthrea...id=18841<br />
It may make a difference for MPEG2 encoding? At best, it would be a few percent. In any case, low latency memory is not worth paying money for. Brand name or the cheapest RAM you can find will work. You may want to test RAM with memtest to account for poorer quality control among cheaper RAM.

The above mentioned Smithfield system has 1 GB (2x512) DDR2, 667 in dual-channel configuration. My personal 3.2E has PC4000 OCZ Gold Rev. 2 RAM with very aggressive timings. Sandra shows a 5.7GB/s bandwidth on my machine versus that of merely 4.7GB/s on the P4D system. My board's an 875i chipset, and I believe the Dual-Core EE is on a 955, but I'm not 100% sure. Our dual opteron systems are consistently over 6GB (using PC3200 ECC).

For DV, a RAID is unnecessary and may hurt performance (or not help it at all). Non-RAID is a faster configuration if you read off one drive and write to the other. In many scenarios for *DV*, hard drive speed is not an issue.
In special cases it is: i.e. laptops getting dropped frames, drives that are running out of space and/or are fragmented

DV requires nowhere near the bandwidth as HDV, especially when you're dealing with 1080. We have a number of testing sytsems, and systems with RAID arrays far outperform those with standalone drives. When you're trying to play three streams at once, one drive can't handle it.

DV and HDV are different ballgames, and a few percent here or there is an improvement, gets your rendering done faster, and makes real-time playback even smoother.

We've tested a number of systems over and over again -- Athlon 64 systems, P4 HT systems, Dual Xeons, and Dual Opterons -- and for the most demanding HDV projects, the Dual Opterons just perform better. The tests that we've run involve our software with PPro, and Sandra.

EDIT: Obviously, if you are sticking with DV most of this doesn't apply as the improvement gained by more sophisticated hardware is minimal compared to when dealing with HDV. Remember though, you generally pay for what you get.

[Glenn Chan]

Hi Tim,
If you look in my link:
http://www.dvinfo.net/conf/showthrea...threadid=18841
You'll see I have run my own benchmarks to measure things. A ~50% change in memory timings or memory bandwidth make less than a few percent difference in render performance. It's so small that it's hard to measure, but nonetheless less than a few percent.
*Memory bandwidth was measured using memtest86. It gives very similar results to one of the Sandra memory tests... the unbuffered ones I believe.

I think my benchmark is a much closer approximation to real world results than Sandra. You may see on many hardware review sites (xbitlabs.com, anandtech.com, etc.) that memory bandwidth correlates extremely weakly to actual performance because it typically is not the bottleneck. Memory bandwidth may have been a bottleneck in the RAMBUS days, where it may have made a meaningful difference in performance.

The above mentioned Smithfield system has 1 GB (2x512) DDR2, 667 in dual-channel configuration. My personal 3.2E has PC4000 OCZ Gold Rev. 2 RAM with very aggressive timings. Sandra shows a 5.7GB/s bandwidth on my machine versus that of merely 4.7GB/s on the P4D system. My board's an 875i chipset, and I believe the Dual-Core EE is on a 955, but I'm not 100% sure. Our dual opteron systems are consistently over 6GB (using PC3200 ECC).

You can change the memory timings on your RAM if you go inside the BIOS (*warning: making your RAM run faster, by lowering the timings, may/will cause instability). You will likely find that they make, at best, a few percent difference on performance and a milk (<~10%) difference to memory bandwidth.
I have run my own benchmarks, which you can see in the link above.
Hardware review sites also report similar things:
http://www.xbitlabs.com/articles/mem...ay/p4-mem.html

2- Memory bandwidth is only one factor is determine why one processor is faster than another. Quick way of checking this is to drastically alter memory bandwidth (i.e. changing the FSB/RAM divider)- my own tests show less than a few percent difference in performance.
There are a large number of factors which affect processor performance- two of them for example are pipeline length [which explains why Prescott core Pentiums are usually slower than Northwood cores, even though Prescotts have more cache and look better on paper], and floating point performance [a major reason why AMD is very good at DAW performance]. Because of these factors there is a wide variation in performance. In my opinion, the best thing you can do is to find benchmarks for your specific scenario (main factors are program you run, and which processors you can afford).

We've tested a number of systems over and over again -- Athlon 64 systems, P4 HT systems, Dual Xeons, and Dual Opterons -- and for the most demanding HDV projects, the Dual Opterons just perform better. The tests that we've run involve our software with PPro, and Sandra.

You might be right there. I have no way of telling whether that's true because I don't have those systems to play with, and there are few relevant benchmarks.

Vegas may be a little different than Premiere. rendertest.veg results for Vegas shows Intel and AMD processors running neck to neck (Intel perhaps a little faster by a few-several percent). Those results are old though.

Results I've seen for Premiere Pro show AMD processors ~20% faster than an "equivalent" Pentium.

[George Ellis]

Again, remember we are talking about Pinnacle Liquid Edition 6 here. Pinnacle does not list AMD processors on the recommended or minimum list, although they have tested some of them. For 5.x, they did list various nForce2/3 boards as compatible, and I do not expect any issues with any AMD products. I do know that SSE2 compatibility is assumed on 5.x and you had to turn it off on AMDs that did not have it. 6.x seems to have gotten a lot less picky about hardware, but does like TI IEEE 1394 and ASIO2 support. Do get a M-Audio or Creative Labs Audigy 2 ZS or 4 card instead of using onboard Realtek chips. Onboard has been a source of OOS issues.

They are a couple of guys building X2 and 275 systems. But, since they did not have duals to compare to, and we do not have a standard benchmark, there really is not a way to compare.

Also, learn towards ATI for a Pinnacle system. nVidia is more often prone to driver issues with LE. LE prefers DX9 performance over OpenGL performance. Memory is king for HDV. I would recommend a X800 with 256MB for better seamless video. Not sure an 850 would buy much more. But boy, I sure want to try the new R520 boards.