MSimon wrote:It might be useful to do multiplies on the accelerator and adds/subtracts on the main processor.
Unless they are MADDs, as those tend to run in a single cycle on GPUs....
64bit isn't that far away for GPUs, the issue won't be computational intensity, it'll be memory. 1G video ram is terribly rare. GDDR5 should have tons of bandwidth, but you won't want to do scatter/random memory access with it, and it isn't clear to me when it will actually make itself known in GPU-world.
[My understanding of NVidia's schedule as it exists now is that 64bits arrives first half 2008, while second half sees the release of some bigger guns.]
And there is always FPGAs if we have the $$$ to go custom.
In fact if it looks like this kind of a machine is a go I can see purpose built silicon to do simulations. That should get the speed up to a useful level.
If Congress turns on the $$$$$$$$ I would do an FPGA machine first as proof of concept and then go to custom silicon for a 10X speedup.
Engineering is the art of making what you want from what you can get at a profit.
AMD announced a GPU with 64-bit math in November, to be released Q1 2008. It, too, will ship with a C streams toolkit. The press release says it'll have 2 GB of onboard RAM and will run at 500 GFLOPS, a number that would equal the NVIDIA Tesla... that 64-bit processing is looking awfully nice.
MSimon wrote:If you decide to go the FPGA route let me know. I have some experience along those lines.
I would think building an FPGA to do double- or quad-precision floating point would be a rather specialized, demanding task. Are there libraries to do the basics (MDAS, say) available?
MSimon wrote:If you decide to go the FPGA route let me know. I have some experience along those lines.
I would think building an FPGA to do double- or quad-precision floating point would be a rather specialized, demanding task. Are there libraries to do the basics (MDAS, say) available?
The only special stuff is the arithmetic alu. The rest is just registers, memory, stack and external ram interface.
Engineering is the art of making what you want from what you can get at a profit.
The device is almost electrically neutral. The departure from neutrality to create a 100 KV well is only one part in a million, when you have a density of 10^12 cm^3. The departure from neutrality is so small that we found current computer codes and computers available to us were incapable of analyzing it because of the numeric noise in the calculations by a factor of a thousand.
You don't need to depart from neutrality at all to create a 100kV well, all you have to do is displace a negative charge inward with respect to a positive charge (i.e. a spherical capacitor)
The device is almost electrically neutral. The departure from neutrality to create a 100 KV well is only one part in a million, when you have a density of 10^12 cm^3. The departure from neutrality is so small that we found current computer codes and computers available to us were incapable of analyzing it because of the numeric noise in the calculations by a factor of a thousand.
You don't need to depart from neutrality at all to create a 100kV well, all you have to do is displace a negative charge inward with respect to a positive charge (i.e. a spherical capacitor)
A local departure from neutrality will require precision similar to a global departure.
Engineering is the art of making what you want from what you can get at a profit.
Would it be possible to model the distributions assuming equal charge densities, and then model (and scale) the departure from neutrality as a separate variable?
I guess it would be hard to apply in the areas where the near-neutrality assumption breaks down (like near the electron guns)...
93143 wrote:Would it be possible to model the distributions assuming equal charge densities, and then model (and scale) the departure from neutrality as a separate variable?
Yes, this is called "perturbation theory". The equation of state relating pressure to number density is the tricky part.
I guess it would be hard to apply in the areas where the near-neutrality assumption breaks down (like near the electron guns)...
You can still do that, you just isolate the space zones into different problems. This is how people solve the plasma sheath near cathode and anode arcs. Each region near the source or sink is modeled separately from the bulk plasma, and you match density and potential and current flow at the region boundaries. Usually you also have to match derivatives as well so the blending is realistic.
They have virtual machines "for rent", where you pay by-the-hour for runtime. The highest class includes 15GB RAM with a 64-bit CPU platform. The S3 storage service gives you a central dump spot for putting data (I/O from the internet to S3 costs money, I/O from the internet to EC2 instances costs money, but I/O from S3 to EC2 costs nothing)
Still need something to collect all the data, and in a precise-enough format, then enough bandwidth to upload everything... but it's an option. Compute-on-demand.