Steve, I don't disagree with you at all. The technical reasoning is very clear. What's not clear is when the heat/energy hurdle will be overcome-- or why it's taking a while to get there. In the meantime, the mini cluster solution is the only way to go for the moment.

Likewise, I also agree with the software quandary: hardware can do *more*, in layman's terms, than the software can. Photoshop has been my personal benchmark app, particularly on the Mac platform-- so much has fallen in line behind PS developments that one can almost set their watch with software trends as PS takes its stand.

You also hit on something about which I've mused elsewhere-- and that if any company would get apps to 64-bit it will be Apple. If it does no good to have a powerful computer when software won't take advantage of it, it follows that a company like Apple would follow their hardware pioneering with fully optimized versions of their own software.

Any allusions to the oddness of all this has less to do with it being shrouded in mystery and more to do with how awkward a situation it is with hardware being currently a bit ahead, so to speak, of the software.

That will change soon enough, and perhaps Steve Jobs will have much more to say about this next week since his last keynote placed 64-bit processing as a headline with no real mention of its pro apps taking advantage of this just yet.

64-bit aside, just getting 32-bit apps completely recoded for Universal Binary and optimized for the increasing sizes of CPU clusters will be a most intersting development to watch.

@Another User said:

'm not sure what the point would be if an 8 processor core machine equipped with 16 or 32GB of RAM would be if it could only run a "handful of instances of VI's."

And if the 16GB of RAM continues to cost $4000.

But I have to repeat that running both plug-ins and stand-alone instances of instruments is the shiznit. G5s may be slightly underpowered compared to the latest Intel machines, but I can't believe how much I'm able to get out of a single 2 x 2.5 with 8GB of RAM in it.

@Another User said:

G5s may be slightly underpowered compared to the latest Intel machines, but I can't believe how much I'm able to get out of a single 2 x 2.5 with 8GB of RAM in it.

Nick-- can you post your instances and sample count? Or, if you've already done this, can you point me to a thread? I have the same CPU and RAM as you, but I'm a little underwhelmed at what I'm able to get loaded in at once.

Nick:

Unless history is no guide, RAM prices for ECC FB RAM, now extraordinarily high, are likely to come down. I'm old enough to remember when 16MB chips were priced at $1000 + per chip.

"...can you post your instances and sample count?"

I don't know if Nick ever posted these specifically. Search for the following thread for the most detailed overview I can recall: "Vienna Instruments Plugin + Standalone and Logic??" I tried to copy and paste the link, but it wouldn't take (operator error, probably).

Nick, I have a question for you too. With your RAM maxed out, what kind of CPU hit do you take on a 2 x 2.5 when the sequencer is running with nothing playing back? As I mentioned in another thread, just loading the RAM seemed to require a lot of CPU while running, even when the SPL was passing over no regions. So upping my memory was a mixed blessing: I could have more samples on line, but I could playback even less than at a lower RAM config. I'm on a 2 x 1.8, so maybe things get exponentially better at 2 x 2.5.

I'll post when I get a chance to load it all up again. What I can say is that it was playable - the CPU wasn't gagging.

@Another User said:

I'm old enough to remember when 16MB chips were priced at $1000 + per chip.

Hey you young whippersnapper, I'm old enough to have paid $400 to upgrade my Mac Plus from 1MB to 4MB. And that was after paying $650 for the 30MB hard drive that went with that "insanely great" system. [:P]

Ix VSL Stand-alone + Logic

http://homepage.mac.com/virtualinstruments/.Pictures/1X%20VSL%20stand-alone%20+%20Logic.png">

Logic, 3 Viennas, K2, other stuff

Nick--

You are da MAN!! Thanks!

J

"...the CPU wasn't gagging."

By comparison, my CPU meter shows 35% use on the left (first core) and about 20% on the right (second core) after I hit play with nothing performing. And that's with only one GB in VSL server and one-half GB in Logic, and a remaining 3.19 GB reported as "free" in Activity Monitor.

When I loaded RAM closer to Nick's numbers, I was seeing a 50 to 60% hit in *both* CPU's as soon as the transport was running without any music. It was a surprising loss of CPU headroom just because the RAM was loaded, irrespective of playback limits, bus speed, disk speed, etc. I was paying a CPU penalty just for getting more sounds online. Within a handful of cross-faded layered tracks, or after maybe ten velocity-switched tracks, Core Audio overloads would stop playback.

I think in 12 to 18 months, this will all be moot.

Yes, I'm looking forward to hardware and software improvements myself. An Intel tower or two are on my must-haves for 2007.

I'm also eager to see how Leopard impacts Cube updates...

@JWL said:

...if the CPU won't increase in speed substantially, then Apple opts to put more of them in one machine. Very odd in a way.

But other things will have to improve to get the best out of VIs-- hard drive speeds and transfer rates as a standard would be my first desire for improvements. Bus bandwidth would also be among the most important improvements.

While I'll never expect that one computer will run much more than a handful of instances of VIs needed for a full orchestral realization, it remains a little strange that more can't be done on one computer than is currently possible.

The "multiplying" of CPUs seems pretty much unavoidable to me -- unless they come up with a fundamentally different paradigm for processor design. There are physical limits to how far the current approach can go (power consumption vs heat dissipation), which we first saw when the early 90-nanometer chips hit the shelves. It was a bit of a fiasco, really...

But the RAM thing is still a strange and interesting problem to me. I think you're absolutely right about hard drive and bus speeds being the primarily focus for the work we do. Some of the advances in flash-based drives, like Samsung released back in the summer (I haven't looked into this lately... any improvements?), are quite promising. IMO, we're kind of looking for the wrong solution by wanting all our samples to be loaded into RAM. The speed of RAM today is necessary for *calculations*, but is basically overkill for playing back samples. Yes, when you've got loads of samples to play in that time it's a somewhat different story, but that could also be dealt with using a flash-based HD on an appropriately designed bus (don't know the numbers, but PCIe is probably already capable). But if you keep in mind the fact that all the samplers we've been using for the last few years have been "streaming" samplers anyway, then it's quite clear that it's only the first few milliseconds of playback where there's a major issue. With a device that could deliver in those first milliseconds (i.e., the next-to-zero seek time of a flash-based HD) we should be able to stream to our heart's content. Anyway, I don't think loading more and more into RAM is the only, or the best, way of improving sample-based work.

The other solution, which I mention whenever I get the chance, is to merge the sequencer and the sampler at the lowest level possible. I keep pestering everybody with a VSL-designed and created notation-based sequencer because I genuinely think that will be the best solution. The fact is that, except for those times when we're literally *playing in* a new part, we absolutely *do not* need all those samples to be loaded in RAM! For the most part, the incredible speed that RAM is capable of is wasted on storing samples for music that is played back in precisely the same way everytime we hit the "play" button. If VSL authored a sequencer, even if it wasn't notation-based, they could address the preloading of samples based on the information contained in the sequencer track. In the simplest possible model, they could create a system whereby, once a passage is recorded, a sample list is created, with timestamps for sample preloading. The preload timestamps would be placed an appropriate period before each note, say 50 milliseconds (overkill, probably) and the sample head would be buffered in that window. I built a system like this in MaxMSP and Java which actually worked quite well (though, not being a sequencer itself, it needed to analyze midi files to build its sample list), and those are definitely not the best-suited languages for this kind of work! (In a more sophisticated version, the timestamp could be calculated dynamically, in order to more efficiently distribute the workload in busy passages.) But of course, if you don't know what note's coming, then you can't buffer the sample to play it, so unless VST is given a view on the entire sequence (does VST 3 support this?), then we won't see this *extremely* simple solution implemented, since plugins remain completely blind to the future.

Anyway, the point is that unless we plan on having 50 keyboard players on stage, all playing a single VI each, *live*, then there's absolutely no need to have all those samples playing from RAM on each and every pass.

J.

...of course, i realize that the "solution" I proposed above is basically the same as "freezing" tracks, and that then the real issue would be reloading the samples for live playback/recording. But that's where a flash-based drive would come in, making the buffering of samples for live playback *much* faster.

Obviously, I don't have a final solution. I'm just dreaming and hoping, like everyone else...

J.

i'm considering the hybrid-drives (legacy rotating magnetic disks combined with flash) as a dead end even before it is released. flash-memory has a limited number of write-cycles (80.000 - 150.000) and will become unreliable after a certain number of cycles. very nifty algorithms had to be used to spread the write-operations evenly across all available memory cells.

more interesting would be millipede-storage (originally announced for past summer) http://en.wikipedia.org/wiki/IBM_Millipede offering data throughput in the gigabyte range - an ideal medium for read-only data like sample libraries.

we will have to wait how latency values for millipede will be in real life, because latency of flash-drives are great for little buffer-sizes, whereas the size of storage is too limited (currently 8 GB)
christian

Yeah, the millipede looks great! I think you pointed me to that article once before... (It sounds familiar anyway.) I didn't mean to suggest that flash-based drives were necessarily the best way to go, just that something other than rotating magnetic discs would probably pose a way forward. But it's nice to have some details! Thanks.

J.

Plowman, if I remember right I had a sequence playing when I did one of those tests. (This was a few months ago now.)

I wonder why your machine isn't behaving the same way. And to tell you the truth, I'm not as excited about 64-bit access as I once was, because I'm able to access all the affordable RAM I need to access right now. Maybe the Intel Mac RAM will come doown, but somehow the idea of spending $250/MB (vs. the $75 I paid for the 8GB in this machine) to acccess still more RAM on a single machine is uninspiring.

A 2x2.5 G5 with 8GB and a SATA card for more storage may be 60 computer years old (1 year = 20 computer years), but it's still a very, very serious DAW. I bought this machine in April 2005; historically machines have been feeling very old by the time they get 1-3/4 years old, but not this time. I've been upgrading machines every couple of years since the mid-'80s, so this is quite a change. Part of the difference is that we're not using just one machine anymore, of course. And then Mac World is coming next week, and NAMM the week after, so it's also possible that this will sound silly. However, the point remains.

In any case, I like running stand-alone instruments outside the DAW on the same machine very much. To me it's not a disadvantage compared to having everything inside the DAW.

I have a dual 2.0 with 4.5 Gb RAM and I am running out of cpu before running out of RAM, so I optimize, freeze, and then I run out of cpu. The weird thing is that the VSL server is running clock cycles more than Logic even when the tracks are ALL frozen. I have to unload the VSL instance from within logic to get the cpu clock to go down in activity monitor. Is that normal?

For me the main issue is, yes, you can do a lot with this machine but performance is a big part of it, and to be able to play in the parts and have a decently low latency and hear what you are doing, you have to have it set to 256 buffer with small process buffer, that works nicely even though it could be better - good enough to play the parts. Of course running the machine like that means you have significantly less cpu overheads. When mixing, I reload the proejct with 1024 buffer of course, when I'm finished playing in parts, so that is fine and then you can go and load reverbs etc (whereas I just use a "cheap" reverb sound for composing to save on cpu and get more sequencing done).

Even though my machine isn't fast enough to use it - what is this I'm hearing about using other instances of VI as standalone does this really work nicely?

Miklos.

@mpower88 said:

The weird thing is that the VSL server is running clock cycles more than Logic even when the tracks are ALL frozen. I have to unload the VSL instance from within logic to get the cpu clock to go down in activity monitor. Is that normal?

I'd be interested to know the answer to this question as well. In my experience, it seems as though the VIs run at full CPU whenever dsp is running -- so in some apps that means a high-CPU idle, while in others it means high-CPU any time the transport is running. They do seem to be quite efficient, but I'm curious to know whether they mute themselves when they are not processing audio (or could they be made to do so?).

J.

Okay, I guess I need to install the V.I. player on my dual 2.0 and see whether I have the same problem. You can see the Activity Monitor dumps above to prove that my 2 x 2.5 is perfectly happy working this way.

Were there any architectural changes in the machines? As far as I know the only difference in the 2 x 2.5 is that the CPUs run a little faster and it has a liquid cooling system.

Weird.

Hi Nick,

I, too, would appreciate it if you would try that test with dual 2 Gig machine. That's the machine I have and I really, really want to get sorted out which way to go with it.

I've spent several days now going over options of combinations of ways to go. Even though Mac World & NAMM are happening right away, knowing a clear assessment of my dual 2 Gig machine would certainly help with any decision to be made.

I want to thank you in advance for taking the time and effort to check this out. Also, where do you think would be the best place to look for older dual & quad G5's?

Best regards,
Jack