How about workstations?

Submitted by Dan Lufkin on Tue, 08/09/2005 - 01:32.

I received a Dell catalog today and they have some good prices on pretty powerful workstations with dual processors, fast memory and lots of storage capacity. Has anyone run DNS on a really high-end machine?

Dan Lufkin's blog
Login or register to post comments

Nope -- DNS is Hardwired to Run on P2's!

Submitted by BruceCyr on Tue, 08/09/2005 - 03:06.

Of course they have. IBM ran VV on a Deep Blue machine in PC simulation mode, and reportedly it ran a little bit faster.

Dual CPUs ain't going to make much difference in how any SR program runs UNLESS its written to take advantage of multiple processors. No current commercial SR product is written that way SFAIK (although reportedly Sphinx 4 is multithreaded).

If you run a lot of other active processes while you're using DNS, dual CPUs might make a difference, but only because DNS might have something like exclusive use of a single CPU. If the dual processors run at less than the speed of a faster single CPU and you aren't actively multi-tasking, then the dual CPU rig might slower be than the single CPU rig.

Massive memory isn't going to make a difference either UNLESS you are actively multi-tasking and using all available memory because, according to Chuck Runquist, DNS is hardwired to use a maximum of 256MB.

If you expect to see a huge improvement in how DNS runs as a result of spending big bucks on hard silicon, you might be disappointed. Some improvement, "Maybe"; big improvement, "Not likely". Only the spender/user can determine whether the expenditure is worth it.

OTOH, its feels cool to have the fastest rig on the block -- for as long as that condition holds true.

Bruce

My fantasy computer

Submitted by frog_jr on Tue, 08/09/2005 - 04:52.

I had hoped to purchase a platypus drive or something like that and install DNS on that. But fiscal reality seemed to interrupt once again. Then I decided to try purchasing 4 gig of high speed RAM and creating a virtual ram drive and installing to that, and as of yet I only have 2 gig, with no immediate plans of getting the other 2. I don't know, but that's the only way I can imagine significantly speeding up SR. And I don't know if that would work either. You would have to put everything on the ram drive, really. Probably very slow boot time, but maybe well worth the wait? I don't know.

RAMDisk

Submitted by ScottW on Tue, 08/09/2005 - 05:49.

Don't bother with a RAMdisk. It won't help.

I'm very excited however about dual core processors. Although Bruce is certainly correct that a dual core processor will do very little to improve the responsiveness of DragonNaturallySpeaking under ideal conditions (when your computer isn't doing anything else) it'll help when your computer is busy doing other tasks.

Since I started with speech recognition I thought that recognition should be moved to another processor. After all, these days we don't expect the CPU to draw graphics, encode or decode our audio signals, listen to the network cable, deal the keyboard, or mouse, or manage our storage system. All of these I/O tasks are delegated to dedicated processors, as should speech recognition.

-- Scott W
Speech resources at SpeechWiki
Products at Applied Recognition

Using Dual CPUs

Submitted by BruceCyr on Tue, 08/09/2005 - 15:06.

People have done this for the very reason you cite using "old fashioned" multi-CPU mainboards, including (as I recall) an IBM SR researcher (Hubert Crepy?) who commented on the old VoiceGroup list a number of years back. I'm sure others have done the same thing.

As you and I note, any benefit depends on how you work -- heavy multi-tasking users are most likely to see some benefit, e.g., people processing graphics or multimedia files, or running complex engineering or financial simulations programs, compiling lengthy programs (Linux?), etc., in the background.

Maybe you can comment on whether its currently (or likely to become) possible to dedicate a specific processor to one or more tasks?

The important thing that I want to emphasize for the typical user is that any actual benefit is likely to be disproportionately small relative to the expense.

In other words, the moral for most of us is something like: "Don't get your hopes up", or "Deal with it".

Eventually we have to hope that either ScanSoft/Nuance or MS rewrites their SR product to accommodate multi-threading. At this point, my money is on the latter to do it first, because it seems to have more resources and a greater interest in the prospects.

We probably agree that the main reason why there aren't dedicated SR subsystems like those for graphics, audio, network, keyboard or mouse is that SR is such a minute sliver of a niche market, whereas those other I/O subsystems affect at least a majority of users.

Therefore its clear that the best way to promote SR is get the US government to pass a law mandating the universal use of SR. This would surely be as enlightened and earth-shaking a policy as prohibition Smiling

Bruce

Using Dual CPUs

Submitted by Chuck Runquist on Wed, 08/10/2005 - 03:31.

The nature of SR does not lend itself to multithreading, or to be more technically correct - piping. The reason is simply because SR transcription has to be conducted in a single pipe. I have talked with the programmers on this subject and they all agree that it is not possible to make use of either Hyperthreading or dual core processors in SR. Its the nature of the beast that makes it impossible to engage these in the service of SR.

However, all of these more recent advances do and will have an overall performance improvement factor, which in turn will improve the performance of DNS.

In short, transcription all has to occur in RAM and in one place. The process cannot be split up into multiple segments. The algorithms don't and can't work that way. The processes simply cannot take direct advantage of either Hyperthreading or dual core processors because the whole shooting match is loaded in RAM.

So, I concur. HT and DC may improve performance overall, but, like RAM, they can and will have no direct affect on SR. If you are going to buy a Dual Core system, buy it for overall performance, not for SR, and don't bank on any significant improvement in SR performance unless you heavily tax your system with a lot of software and background tasks that are slowing DNS performance down now.

Chuck Runquist

Your Fantasy Computer and Hard Core Reality

Submitted by BruceCyr on Tue, 08/09/2005 - 14:39.

There's a guy who's big on RamDisks here:

http://overclockers.com/tips00813/

and he cites an actual test at AnandTech:

http://www.anandtech.com/storage/showdoc.aspx?i=2480&p=1

The emphasis is on using the RamDisk for booting into Windows, and the results are underwhelming for some obvious reasons that apply even more strongly to DNS:

The role of disk reading/writing with DNS is minimal assuming you have sufficient RAM to avoid Windows paging. Basically, you read DNS from disk when you start it and the whole thing goes into 256MB or less, depending on things like your vocabulary, usage, etc. Then you may write your userfiles back to disk at the end of the session. That's what? 20 seconds max of disk operations? Versus 4-8 hours of dictation a day. Way less than 1% of you time using DNS is occupied by disk operations. This is the explanation for the dismal results noted in ScottW's link below.

So RamDisk is more or less irrelevant to DNS' operations, unless you are continually shutting it down and starting it up throughout the day.

I'd max out my system RAM first, although few people need more than 1GB. Then I'd get a high quality microphone that suits my way of working ($50-$500). Then I guess I'd bank the rest of my money and wait for some fantasy gizmo worth spending it on.

Bruce

Good to know

Submitted by frog_jr on Tue, 08/09/2005 - 15:03.

Well, I'm glad I didn't buy that Platypus then. And I just purchased 2 gig of some of the best 400 speed ram available for my next best computer (I always have to stay comfortably behind the cutting edge of new hardware, for money reasons) which, BTW, will be a Shuttle XPC, which is one I might recommend for SR. The CPU cooler design and general cooling setup is the unique thing about the XPC's (as well as ease of assembly.) Never seen anything else like it, although one would think somebody else would have done it. It is *very* quiet, and not too pricey.

Shuttle XPC and Heat Pipe Technology

Submitted by BruceCyr on Tue, 08/09/2005 - 15:32.

Yeah, I know that Tom Westheimer has been putting together Shuttle rigs for a couple years now, and other users like Susan Craig are big fans and use them as "portable" rigs!

However, the use of heat pipe cooling technology is scarcely unique to Shuttle. It's been widely used by silent PC users for several years now as well as by overclocking enthusiasts. Mainstream manufacturers like Apple and Dell have incorporated it into desktop designs, and of course it was pioneered in notebook systems in the PC realm.

Bruce

Re: Shuttle XPC and Heat Pipe Technology

Submitted by frog_jr on Wed, 08/10/2005 - 05:00.

BruceCyr wrote:

the use of heat pipe cooling technology is scarcely unique to Shuttle ... Mainstream manufacturers like Apple and Dell have incorporated it into desktop designs, and of course it was pioneered in notebook systems in the PC realm.

No, it is not heat pipe technology that makes them unique. It is how it is implemented. Others may have done something like this; Apple is usually pretty ingenious in their case design. Dell, too. But I have seen plenty of "heat pipe" heatsinks, and I have yet to see a heat pipe cpu cooler that places the "heat exchange" of the cooler so that the case fan blows air through them after it has already gone through the case. Subsequently, the CPU heatsink is not heating up the air inside the computer case. (See: http://eu.shuttle.com/en/resourceimage.aspx?rid=675 -- the fan is also the exhaust fan for the computer case.) I am beginning to sound like an advert here, sorry. No other heat pipe unit I have seen so far has applied this simple solution. Heatpipes, yes. I have a Thermalright XP 90 here on my workbench right now, some (Thermalright) say it is the best.

Presently, I am recording meetings with an IRiver IFP-895, bringing them home and transcribing them using an Infinity pedal and Express Scribe. (Tried recording with a "professional solution," a Marantz PMD 660, and found it sorely wanting.) I was dictating into VV for 2 years (initially due to a pain in my elbow from typing) and now have migrated to DNS Preferred. My computer is a Gigabyte P4 with 1 g of 400 speed DDR ram, and a 2.66 ghz Northwood cpu. It has 3 cooling fans, including one that I drilled a round hole in the top of the case for, and it is the noisiest thing in this room (esp. since I put that hole in the top of the case, heh.) As soon as the XPC arrives, I will have my first machine with dual channel ram. I am also hoping this makes a significant difference in performance. Any (affordable) suggestions? Or not so affordable?

Cooling Suggestions

Submitted by BruceCyr on Wed, 08/10/2005 - 14:48.

Thanks, I forgot the nuance of the Shuttle using the case exhaust fan as the heat pipe cooler. However, I'm confident that silent PCers and overclockers have already done it. Then there are those, like Zalman, Apple, and super gamebox builders with names like Voodoo{?), who dispense with fans altogether and use the case as a heatsink to draw heat from the heat pipe -- silent, effective and very expensive.

I've run water cooling for nearly four years now and am very satisfied with the results. There are excellent and affordable (i.e., $100-$400) add-on commercial units available, including fanless but awkward tower cooling solutions from Zalman and ThermalTake.

If I were doing a commercial solution, I would think primarily of the Koolance EXOS unit or maybe the SwiftNet kit based on the WhiteWater block that my custom system uses. I use a Hydor L30 pump and reservoir and radiator from DangerDen, with a Zalman rheostat controller for my Panasonic fan so that I can vary the air flow/wind noise to suit conditions. My setup is geared more towards performance than quietness, based on my preferences, but it just as easily be tuned toward silent operation for gear that's run within normal parameters.

I'm not sure how any of this compares with the Shuttle because I haven't tried it, but I would think that the fanless tower units mentioned above would be quieter. Its a question of whether your environment/work habits are conducive to the use of the tower coolers.

Bruce

Here are some links for the items mentioned above:

Zalman's fanless watercooling tower system:

http://www.zalman.co.kr/usa/product/view.asp?idx=160&code=021

Zalman's fanless heatpipe box:

http://www.silentpcreview.com/Sections+index-req-printpage-artid-161.htm...

A review of new, smaller, and apparently cheaper version of the unit above, which is not yet documented on Zalman's site:

http://www.hexus.net/content/reviews/review.php?dXJsX3Jldmlld19JRD0xNDA3...

A review of a Zalman's fanless heatpipe gamer rig sold by another manfacturer:

http://www.silentpcreview.com/Sections+index-req-printpage-artid-161.htm...

You will see links for ThermalTake's fanless tower watercooler and more traditional watercoolers:

http://thermaltake.com/watercooling/bigwaterse.htm