If you go look for some proper research papers on the subject of GPGPU computing, you'll spot graphs displaying the speed-up compared to CPU or multi-CPU setups also taking into account cost and power consumption.
Granted that the problem at hand is appropriate for parallel computing and if the guy writing the code knows his stuff, employing the GPU can bring a *huge* improvement over a CPU or CPU-based multi-processing code in terms of time, cost and power.
General purpose GPU computing is the future. In a few years regular software will seamlessly take advantage of our GPUs for non-3D tasks (in applications where the processing can be executed faster on the GPU, of course).
It's a new field (although closely related to parallel computing) and there is lack of expertise and plenty of stuff that haven't been run on GPUs (or there is plenty of room for optimization). So if you have a kid/nephew or someone who's getting into computers, let them know.
Last edited:
TheShaman,
You are very correct that the important thing here is to optimize the code to actually take advantage of the multi-thread and multi-core processing we have already available today. My CAD program is a high end piece of software and I don't think that they have optimized what they have in this regards at all yet. Rendering is the slowest process but also some small changes to a 3D design can seems to take forever for the processors to handle. I know that my next computer will have a GPU as a requirement as that just seems to be the best way to go at this time. With ever increasing multiprocessors and multicores perhaps that won't always be necessary if you are only running one task at a time. I still have an old server case with multiprocessor board just to run some old legacy hardware. It is amazing how large that thing is, but it would hold many hard drives and dual processors was really something at the time! Parallel Zeon processors seemed like such a feat years ago, today any i5 or i7 will run circles around those older chipsets!
You are very correct that the important thing here is to optimize the code to actually take advantage of the multi-thread and multi-core processing we have already available today. My CAD program is a high end piece of software and I don't think that they have optimized what they have in this regards at all yet. Rendering is the slowest process but also some small changes to a 3D design can seems to take forever for the processors to handle. I know that my next computer will have a GPU as a requirement as that just seems to be the best way to go at this time. With ever increasing multiprocessors and multicores perhaps that won't always be necessary if you are only running one task at a time. I still have an old server case with multiprocessor board just to run some old legacy hardware. It is amazing how large that thing is, but it would hold many hard drives and dual processors was really something at the time! Parallel Zeon processors seemed like such a feat years ago, today any i5 or i7 will run circles around those older chipsets!
Our entire network of Linux boxes is configured as a compute farm and at night you could have dozens of machines running a transient analysis in parallel. Things that take days can take hours. The partitioning is special software but it does exist.
Scott,
Shush with the Linux reference!!!!! If my brother sees that I will have to get an earful about the awful Microsoft monster and why I still have anything running in that OS. I do have a dual boot system with Linux and now Windows 7, no way am I going for that Windows 8. He would just love to network all the machines in my house to be running Linux, I started that monster about 20 years ago when I gave him a book on Linux. He has been in that world for that long now, OpenSuse is his distro of choice. It must be nice to work for a company that knows how to get things done, I hated fighting with the IT department over some of the dumb things that they use to do.
Shush with the Linux reference!!!!! If my brother sees that I will have to get an earful about the awful Microsoft monster and why I still have anything running in that OS. I do have a dual boot system with Linux and now Windows 7, no way am I going for that Windows 8. He would just love to network all the machines in my house to be running Linux, I started that monster about 20 years ago when I gave him a book on Linux. He has been in that world for that long now, OpenSuse is his distro of choice. It must be nice to work for a company that knows how to get things done, I hated fighting with the IT department over some of the dumb things that they use to do.
Dilbert periodically captures some of the essence of typical IT departments to great effect.
When I consulted on site at Harman I dealt with a woman, actually a friend outside of work, and had to try to negotiate with her for a faster machine. She said "But Brad! I only have a 500MHz machine at home!" I said "But ____, what do you use it for?" Of course it was nothing compute-intensive. I pointed out that increases in speed for me translated directly into savings in time and more productivity for the company, but it just didn't cut it. Many times the argument is "well if you get that, everyone will want one" and so forth. Madness.
Seems HKardon had the wrong person for the job/position for you to negotiate with. It happens a lot.
back to the numbers -- international computer ranking is done twice a year. As of June 2012, LLNL's Sequoia is the number 1 ranked as world's fastest supercomputer.
[China's Tianhe-1A is now ranked 5th.]
The problem for the physicist is the resolution of the data -- as they are always ahead of what can be built! One new supercomputer was faster and they thought more accurate but after it was up and running those important details contained in the last few places seemed like they were put there by a random number generator. they couldnt get the same answere twice. But it was fast. So, they had to go back to using the 'old' one. Most of the details of what is being looked at are in the last few digits.... they can never get enough info it seems to tell them exactly what is happening in their physics models. The super high performance computers are still not good enough. Such is the power of the human mind. -RNMarsh
back to the numbers -- international computer ranking is done twice a year. As of June 2012, LLNL's Sequoia is the number 1 ranked as world's fastest supercomputer.
[China's Tianhe-1A is now ranked 5th.]
The problem for the physicist is the resolution of the data -- as they are always ahead of what can be built! One new supercomputer was faster and they thought more accurate but after it was up and running those important details contained in the last few places seemed like they were put there by a random number generator. they couldnt get the same answere twice. But it was fast. So, they had to go back to using the 'old' one. Most of the details of what is being looked at are in the last few digits.... they can never get enough info it seems to tell them exactly what is happening in their physics models. The super high performance computers are still not good enough. Such is the power of the human mind. -RNMarsh
Last edited:
Finished my boards, separate input and output stages that fold into a 1"x1" stack. I stuck with TO-92 outputs and J310 cascodes for now and found some SOT23-6 duals NPN/PNP, PNP/PNP, and NPN/NPN I'm not ready for those tiny packages yet.
Filled in with some SOT23 to DIP-6 adaptors for matched pairs.
Filled in with some SOT23 to DIP-6 adaptors for matched pairs.
Attachments
Last edited:
After I verify functionality I will gladly post the files. Anyone can then change packages if need be and order some.
For those that missed the reference to "Dilbert". Dilbert is a long run newspaper comic series about a design engineering office gone bad.
The official Dilbert website with Scott Adams' color comic strips, animation, mashups and more!
The official Dilbert website with Scott Adams' color comic strips, animation, mashups and more!
Teeny packages, somewhat off-topic
There are smaller ones but they are 0.5 ohm on resistance and correspondingly high capacitance.
So I turn to discretes. Rohm makes dual MOS parts in their VMT6 package, about 1.2mm x 1.2mm. With two packages per line, N and P, each dual connected back-to-back, and complementary drive signals, in essence rolling one's own dual bilateral switch, it can be done, and with acceptably low capacitance. The SOT23-6 parts that are SPDT are singles, hence two needed for the two lines. To remind, SOT23 has about a 3mm x 3mm footprint.
One could at this point ask if there couldn't just be another port on the uC and a separate I2C interface thus. The answer is, maybe, if other line functions might be consolidated.
Working today (paying work, nice way to start the new year even if it isn't audio) trying to find the best way to open-circuit the I2C I/O of a part in a watch, so it can be powered down to save battery life but share the bus with an EEPROM. There are many analog switches out there now for low voltage operation, but the smallest packages I've found with the appropriate low capacitance are huge: SOT23-6I stuck with TO-92 outputs and J310 cascodes for now and found some SOT23-6 duals NPN/PNP, PNP/PNP, and NPN/NPN I'm not ready for those tiny packages yet.
There are smaller ones but they are 0.5 ohm on resistance and correspondingly high capacitance.So I turn to discretes. Rohm makes dual MOS parts in their VMT6 package, about 1.2mm x 1.2mm
. With two packages per line, N and P, each dual connected back-to-back, and complementary drive signals, in essence rolling one's own dual bilateral switch, it can be done, and with acceptably low capacitance. The SOT23-6 parts that are SPDT are singles, hence two needed for the two lines. To remind, SOT23 has about a 3mm x 3mm footprint. One could at this point ask if there couldn't just be another port on the uC and a separate I2C interface thus. The answer is, maybe, if other line functions might be consolidated.
bcarso,
What type of watch are you working on that would have short battery life these days? Can't fathom that. I still have the original Acutron red faced led watch I had as a kid, that had some sort of micro-switch that all you did was move your wrist quickly and it would light up the display. Learned my first mechanical lessons fixing pocket watches as a kid.
What type of watch are you working on that would have short battery life these days? Can't fathom that. I still have the original Acutron red faced led watch I had as a kid, that had some sort of micro-switch that all you did was move your wrist quickly and it would light up the display. Learned my first mechanical lessons fixing pocket watches as a kid.
It's a watch with many functions accessible with a touchscreen. The standby current equates to a long battery life, but if you start using some of the functions, particularly the LED backlight, or stopwatch, or... things deteriorate significantly. I have been hovering around the project without being involved from the outset, just called in when the client is desperate. Involvement at the beginning might have made a lot more sense, but you know how these things go.
Finished my boards, separate input and output stages that fold into a 1"x1" stack. I stuck with TO-92 outputs and J310 cascodes for now and found some SOT23-6 duals NPN/PNP, PNP/PNP, and NPN/NPN I'm not ready for those tiny packages yet.
Filled in with some SOT23 to DIP-6 adaptors for matched pairs.
Hello Scott
What PCB board layout software do you use.
So the foke-lore on the street says. But, the laptop cpu architecture isnt optimized for number crunching and the word size is a lot smaller etc etc. Larger word size means more can be done in each cycle. Number crunching is what the super's do best as they have a lot of it to do. Also, the use of memory is much faster and a lot more of it. etc etc etc. [ I was at LLNL when the first Cray went in and subsequent ones etc.] Thx-RNM
Last edited:
Stuck with ExpressPCB all free but you can't transfer the results to another CAD package.