Intel i7 Ivybridge on its knees.....

[FrankJH]

That's a heavy load for the CPU but a trivial one for those GPUs. The only solution is to offload decoding to the GPU.

Whether your statement is true or not, is neither here nor there unfortunately. If the software hasn't been written to hand-over decoding issues from the CPU to the GPU then it makes no difference (unfortunately for the OP)

He HAS to have a CPU based solution

 

[Humour]

That's a heavy load for the CPU but a trivial one for those GPUs. The only solution is to offload decoding to the GPU.

I hear you and I agree, it is the logical thing to do, however I do not control how the VMS provider designs their architecture or what priority they work on with their coding, I'm simply a client/integrator, one of thousands so my voice is unlikely to get heard, plus that would be a long term solution assuming the company decided to put that on their priority list.

On that basis how do you propose that "I" offload the workload to the GPU?

@ mmj - I hear you, its the same line of thought I had started with, the only difference in my case is I'm not in touch with the last 2 gens of CPU developments due to lack of time so my general reading about comparisons between the Hex and Quads did not clearly show a 40% performance difference, therefore I was questioning how much value there is in making the change, as 10% increase at a premium price is just not quite justifiable. Interestingly however, since it seems plausible to run a Hex on a dual socket workstation grade motherboard the most logical and future proof upgrade is to pair a dual socket mobo with a single Hex core and sufficient memory and see the result. If it transpires that the gain is marginal at least then there is the option to add another Hex core which should provide a substantial boost compared to the single 3770, or at least that's how the theory goes lol

regards, Hum

 

[Humour]

Oops, scratch the Dual 3930k's idea, the single QPI on the 3930k puts an end to that idea.....I bet my left sack Intel made sure that was not an option to ensure Xeon sales and margins are maintained.

Hum

 

[arad85]

So... if I understand you correctly, you have 36 decode threads per computer. In which case, you need as many real cores in each computer as you can get your hands on.

 

[thenewoc]

Would Virtu MVP be worth a look at?

 

[joeyjojo]

I hear you and I agree, it is the logical thing to do, however I do not control how the VMS provider designs their architecture or what priority they work on with their coding

I find it surprising that they offer no support for GPU decoding - it's been a standard for at least 5 years now.

I'd be curious how they achieved it in the system described in this article with 100 streams. The third page says:

...decoding HD H.264 video streams on a PC was highly CPU and graphics intensive. "The requirement to decode upwards of 14 streams on a given PC was achieved by massive upgrades to the PC's CPU and GPU [graphics processing unit], followed by significant tuning."

So it could be that asking a single box to decode 36 streams is impossible without hardware acceleration. Rather than sinking a lot of money into a few top-end "exotic" workstations, you're probably better off getting a couple more high-end workstations or servers and dividing the load as the Calgary people presumably did.

Of course, another solution may be to downsample the streams at the source - from 1080 to 720 might do the job and retain enough detail.

 

[ryder]

Of course, another solution may be to downsample the streams at the source - from 1080 to 720 might do the job and retain enough detail.

This was my first thought, hence the first question

 

[mmj_uk]

Oops, scratch the Dual 3930k's idea, the single QPI on the 3930k puts an end to that idea.....I bet my left sack Intel made sure that was not an option to ensure Xeon sales and margins are maintained.

Yeah you need a Xeon for dual but if a single 3770K is only barely struggling then a single 3930K should be more than enough I would think.

 

[IceWind]

You know.... you might actually be better with an AMD 8 core chip here... or an operton 12 core if you can get one cheap.

 

[JonJ678]

Sell the poorly chosen gpu. Buy a normal gpu. Buy two more systems. The load per system is thus halved.

Nothing else will make financial sense. Don't buy a 3k+ dual socket intel system in order to keep using a 500 quid card, that's throwing good money after bad.

Edit: 6 core sb-e may work, but doesn't exactly lend itself to future expansion

 

[Humour]

Sell the poorly chosen gpu. Buy a normal gpu. Buy two more systems. The load per system is thus halved.

Nothing else will make financial sense. Don't buy a 3k+ dual socket intel system in order to keep using a 500 quid card, that's throwing good money after bad.

Edit: 6 core sb-e may work, but doesn't exactly lend itself to future expansion

@ JonJ678 - What makes you say, "poorly chosen gpu"? I'm not buying anything yet, just considering options. Adding more computers adds to rack usage (all rack mounted), power consumption (UPS load), network ports, etc. etc. so whilst its a consideration, its not the preferred choice as there are hidden costs with that option. Also the 3770's are working at 60% load under normal conditions with peaks up to 90% in certain conditions with 36 streams loaded on the workspace, so my consideration and discussion here is already looking forward to the future, I stated in the first post that I'm NOT looking for a FIX.

@ joeyjojo - I read that article, thanks for the link, I would hazard an educated guess that the Calgary client's budget is a factor of x10 larger than our client, who is just stepping out of the Analogue world and into the Digital. My guess is that either that they scaled the wall with more hardware to drive it, perhaps 2:1 ratio on screens to workstation, or using server grade Xeon dual socket subsystem for more decoding power, or yet another option, custom engineered hardware with dedicated hardware decoders, these by the way are in the range of 20-40K per box and boast multiple Xilinx FPGA processors with far more logic gates and processing power than your commercial grade CPU, yes even a Xeon. One other clear difference between the two systems is that the Calgary client's system is based on 1.3MP/720p cameras, our client's camera system is 4CIF/D1 resolution, so there is a distinct difference in in pixels between the two systems, that's why I am able to decode 36 streams from the one workstation.

There is however something that was highlighted in the article that I will investigate when I return to site, and that is the network side because our system is also set to multicast, so there may be a hidden issue there as well.

@ IceWind - I'm not convinced an AMD solution will work. AMD chips, even with extra cores consistently perform lower than Intel chips with half the number of cores in computational related benchmarks, so adding extra cores only works if the two were even in computational performance to start with.

best regards from Doha, Qatar. Its bloody hot here, 44C during the day

Hum

 

[Cromulent]

You say that your software is running inside a VM. Are you sure that the software is making use of the hardware H264 decoding available on the GPU? If running in a VM then the software might not be able to make full use of the GPU hardware capabilities (this really depends on what hypervisor you are using).

Edit: I just checked out the specs for the GPU and it might not offer hardware H264 decoding (perhaps it does and I just missed something). If I were you I'd make sure I had a GPU with built in hardware H264 encoding / decoding since that would improve performance dramatically.

 

[Humour]

Cromulent, only the servers are running on the virtual layer. They process the inbound streams and re-route the outbound streams.

Its the client machines decoding the streams the clients wish to view.

The Hypervisor is VMware vSphere, but that's a red herring anyway with reference to your conclusion.

 

[Cromulent]

Cromulent, only the servers are running on the virtual layer. They process the inbound streams and re-route the outbound streams.

Its the client machines decoding the streams the clients wish to view.

The Hypervisor is VMware vSphere, but that's a red herring anyway with reference to your conclusion.

Just a thought.

 

[IceWind]

a lot of real and synthetic benchmarks do show that the AMD cpus are performing better than their intel equivalents at heavily threaded applications.

Intel do indeed have the crown for single threaded performance but when it comes to multitasking the AMD cpus do shine because of their more physical cores

 

[gavind]

Intel do indeed have the crown for single threaded performance but when it comes to multitasking the AMD cpus do shine because of their more physical cores

Correct. And when using this on servers, AMD is clearly the first choice here. That's multitasking multiplied by a million let's say for web servers as an example.

 

[imginy]

You guys are crazy thinking amd will be better than a 3930k