On Sun, 2007-06-24 at 00:15 -0700, Carl Lowenstein wrote: > On 6/23/07, Andrew Lentvorski <[EMAIL PROTECTED]> wrote: > > > > The second part of the problem is that the trace length between the > > processor and the memory is somewhere in the range of 15-30cm. This is > > on the order of 1/2 to 1ns even at the speed of light (and the electrons > > do not move at that speed in copper). Even if you had a RAM latency of > > 0, sending the command and getting the response is limited to about 1ns > > in latency. That means that we have about a 2-3 to 1 CPU to memory > > mismatch that can never be fixed other than by parking memory right on > > the CPU. > > Reality check. I haven't seen a computer in which the processor and > memory were 6 to 12 inches apart (15 to 30 cm) since late PDP11 days > (mid 1980's). > > Another reality check. The speed of electromagnetic wave propagation > in copper is considerably faster than the speed of the electrons. >
This takes me back. The MPEG-2 decoders I worked on at TV/COMM in 1995-1996 had 14 VideoRISC processors on a 13"x9" 12 layer PCB. Each had 2MB of DRAM which contained a portion of the video frame being processed. All the processors had to talk to each other as well as transfer data from memory to either each other or to the host (on a daughter board). The propagation delay was approximately 1nS from host interface to the last VideoRISC due to the extremely long traces for the system bus, which was a daisy-chain. Cycle time for the processors was 2nS. This is the main reason the product failed and Scientific Atlanta wound up owning the digital TV and satellite equipment market. If the bus was designed with shorter paths, such that the host interface connected to the VideoRISCs in a star configuration, the board propagation would have been at most .5nS, leaving more time (1.5nS) for IC propagation delays, signal settle time, etc. On another related note. At NASNI we had a Harris mainframe that had processor and memory more than 12 inches apart. (Note that this thing did not have a single CPU as the processor, but was a bit slice machine. For having a 2uS cycle time, it was extremely fast with IO throughput and response that would put many of today's PCs to shame.) We also had an old mini with 8" floppies that used core memory that could not be replaced. (Damn, I can't remember the make and model of that thing). Talk about slow memory access! This was in 1990-1994. PGA -- Paul G. Allen BSIT/SE Owner/Sr. Engineer Random Logic Consulting www.randomlogic.com -- [email protected] http://www.kernel-panic.org/cgi-bin/mailman/listinfo/kplug-lpsg
