Discussion in 'CellPerformance@B3D' started by Shifty Geezer, May 14, 2009.
Were the previous tests done by grandmaster using a Gigabit connection, etc?
There are special de- and encoders for MPEG2 and H.264, so the four SPEs are used for upscaling, filtering etc.
It's 2.2 times faster compared to a mid-range laptop CPU (single pass), although he predicts that QuadCore Desktop PC may be faster. He didn't measure 2-pass since CE-10 doesn't support one. What is the power consumption of a QuadCore ?
The direct gigabit network and RAID 0 requirements may explain the bottleneck people encountered earlier.
The marketing literature claimed level 5 and 5.1 support. Isn't level 5 closer to 150Mbps ? [EDIT: Okay, you mentioned Hi10P profile for level 4.2]
At that higher level, will the relative performance difference -- if any -- be even wider ?
The same material compares CE-10 to pure hardware encoders. That's probably the reference they are measuring against (due to the limited options in CE-10 at the moment).
EDIT: Also found a full software, real-time H.264 encoder by Sony:
It runs on QuadCore PC, maxed out at 40Mbps, and can be distributed to 4 workstations for the highest picture quality. Seems to support lot's of nice features.
I wonder if FixStar can adopt the same approach: Distribute the workload to 4 or more PS3s (at a fraction of the cost and power consumption) to obtain higher picture quality.
Why would you need RAID0?
My 1 year old WD 640GB SATA 2 HDD does read/write speeds of *60-80+MB/480-640+Mbit* or write and read on the same disk at the same time *35-40MB/280-320Mbit*. CPU usage barely registers.
*Practical usage measurement during copy/read or both operations in Vista of folders etc.
May be from excessive seeks and contention between disk operations during the encoding process (if the encoder divides its work into 4 slices, the HDD will get a lot of overlapping read operations in different "regions").
Your measurement above indicates that doing overlapped disk operations (read and write together) lowers the performance significantly. It probably also depends on what kind of HDD tech was used in other people's tests.
review of the encoder compared to x264 on core i7
I believe said article is written by grandmaster.
The CE-10 FAQ page specifies a gigabit network with RAID 0 array. The user won't get the highest performance using a 100Mb LAN.
How big are the files that are going the be encoded? Small enough to put into a ramdisk on the PC?
Well it's not going to make the encoding better or make it double pass I reckon, so it won't matter much as it stands.
It's a v1.1 product. As long as the foundation is solid, they will improve:
August 2009 ver1.2 Encoder core upgrade
September 2009 ver1.3 Encoder core upgrade
October 2009 ver1.4 Encoder core upgrade
However if the tester/reviewer does not set up the use environment properly, it will be problematic in each and every case (because the bottleneck looks like the network and disk access). I am not saying CE-10 is a great product. I am saying it has its position. The 2-pass encoding may not be the objective for them at this point because they want to sell on the speed (save 30% time base on software tricks alone). If they can improve the quality further, then it will become more valuable. Otherwise, it'd just fold, or they can introduce 2-pass later (e.g., after they are done with the basic encoder core).
In a startup business like this, it is not uncommon to establish a beachhead based on a trait first (In this case, speed) even though other areas may fall behind. If they can do up to 150Mbps, it may also mean they can go into the fringe of digital cinema domain (e.g., 4k x 2k @ 30fps). If you look at how their web pages evolve, they have broadened their service to product customization too. As long as they can get enough business based on the platform, we should be able to see better stuff coming out.
As it stands right now, yes, the basic product lacks in encoder quality, but they'd be able to use the professional/integration services to fund further development of the core. This is also part of the reason I don't think grandmaster is the target audience. They can't sell CE-10 for $199 or even $1999 and hope to survive (plus requirements like gigabit network and RAID 0 array hinder casual adoption). They are after bigger fish with more complex integration and custom needs.
Also, source code for CE-10: