X-Message-Number: 26089
From: 
Date: Tue, 26 Apr 2005 03:03:03 EDT
Subject: Uploading technology (1.i.0) Electronics next generation.

Uploading technology (1.i.0)
 
There I take my crystal ball to look into the technological future. Because 
Uploading as I conceive it use rather ordinary electronics components, the 
subject is limited here to that technology in the comming years.
 
On the computer side, we see now the multiprocessor chip, up to nine for some 
game station. I don't think it will be pursued beyond two or three years 

because this would ask for far too much code optimization in all softwares. This

is only a transition phase until the real next step, the high electron mobility
transistor processor. AMD is said to work on the subject and Intel will do it 
too. HEMT clock can run at 12 GHz on discrete, large components, the same 

technology on very small transistors in a microprocessor would get to the 50 GHz
mark. When HEMT processors will be introduced, the clock frequency will be in 
the 5 GHz range at most, but it will have the possibility, with very few 

add-on to be multiplied by 10. The foundry will have only to tunr the faucet or

nearly so. The initial cost is not marginal yet: A new industrial plant must be
built for that technology, that is why we have the intermediate step of the 
multicore processor.
 
What about FPGA? The market is more limited here and comes from big buyers. 
Prices are not very interesting for the foundry, so they use some production 

tools used before on the microprocessor factory. That is why current FPGAs have
a maximum clock frequency in the 350 MHz range, this is the Pentium 2 

technology. It is now used in the FPGA domain. Because the current P4 technology
at 2 
- 3 GHz will be down graded by the new HEMT, it will be used as the state of 
art process in the FPGA domain. So this is not big gamble to predict a x10 
clock speed for FPGA in the comming years. This will comes with 100 million 
transistors chips, or nearly one million logical cells.
 
On the ASIC side, a big order could use the HEMT technology and a custom made 
TRAC could run at 10 GHz with one million or more operational amplifier 

cells. It could simulate ten uploading quality neurons at the same time and use 
the 
time sharing to run 10 millions such blocks, that is, its simulation power 
would be in the 100 millions neurons range. One brain on 100 chips...
 
Using FPGA at 3 GHz and 1 million logical cells, 50 neurons could be 

simulated at the same time and one million such package could be implemented in 
time 

sharing. This give 50 millions neurons/chip without using custom-made elements.
One brain on 200 chips.
 
This will be the technology in 4 - 5 years.  Assume the FPGA unit price is in 
the $1,000 and other elements cost approximatively the same amount, the brain 
price would be $400,000 in components alone.
 
Yvan Bozzonetti.


 Content-Type: text/html; charset="US-ASCII"

[ AUTOMATICALLY SKIPPING HTML ENCODING! ] 

Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=26089