gsg0384@uxa.cso.uiuc.edu (08/07/89)
Hi,
I am curious about vector registers.
Q1. How much do they cost? I mean if one wants to build a low-end workstaion
with vector registers with MIPS or M88000 or Sparc chip, how much does
it price more than the one without the registers?
Q2. With what processors is it possible to implement vector architecture?
Q3. When do you expect to see vector workstations around $ 20k in the market?
Q4. What are those vector machines ( < $100k) in the market?
Hughslackey@bbn.com (Stan Lackey) (08/08/89)
In article <112400001@uxa.cso.uiuc.edu> gsg0384@uxa.cso.uiuc.edu writes: >I am curious about vector registers. > >Q1. How much do they cost? I mean if one wants to build a low-end workstaion > with vector registers with MIPS or M88000 or Sparc chip, how much does > it price more than the one without the registers? The cost of the actual registers (RAM?) is the cost of the memory. Unfortunately, this is insignificant compared to the cost of the pipelining and other control logic you need to make the registers worthwhile (for a large multichip design. The cost of the registers in a single chip attempt is probably significant.) If you would would like to add hardware to a microprocessor such that vector instructions are added to its instruction set, and the number of cycles spent to begin and end the instructions is to be minimum, you have a very challenging job to do. Some of the micros are set up to extend the instruction set, but at least the 68020 was really only defined to generate a single address at a time, not the linear (or indexed) series you need to make vector processing worthwhile. Newer uP's may be better. >Q2. With what processors is it possible to implement vector architecture? Ardent used a customized version of one of the MIPSco chips for a similar purpose. If you can get it, this may be a good starting point. Or you could just buy an Ardent, I guess. Or any other mini-super, depending upon you needs. >Q3. When do you expect to see vector workstations around $ 20k in the market? This is a problem. You really need a few million transistors to pull this off, not to mention the pipelined memory system to really give full world-class vector performance on large problems. (Note: I'm not saying it can't be done. I just wouldn't expect it in the next few years.) >Q4. What are those vector machines ( < $100k) in the market? Alliant, Ardent, and Steller come to mind. There may be more. You also might want to check out Multiflow, but I think their low end price is higher. Yes I know Multiflow doesn't do vectors. Also: The 88000 can be programmed in a way that handles vectors efficiently, taking advantage of the pipelined, scoreboarded FP units and memory access features. Not as well as a vector machine, but the price is attractive. This comment of course has absolutely nothing to do with the fact that BBN just announced an 88000-based product. :-) -Stan
keith@mips.COM (Keith Garrett) (08/11/89)
In article <43974@bbn.COM> slackey@BBN.COM (Stan Lackey) writes: >In article <112400001@uxa.cso.uiuc.edu> gsg0384@uxa.cso.uiuc.edu writes: >>I am curious about vector registers. [question about vector register cost] >>Q2. With what processors is it possible to implement vector architecture? > >Ardent used a customized version of one of the MIPSco chips for a >similar purpose. If you can get it, this may be a good starting >point. Or you could just buy an Ardent, I guess. Or any other >mini-super, depending upon you needs. > Ardent uses off-the-shelf r2000 processors, and adds their own floating-point vector co-prrocessor and graphics processor. -- Keith Garrett "This is *MY* opinion, OBVIOUSLY" UUCP: keith@mips.com or {ames,decwrl,prls}!mips!keith USPS: Mips Computer Systems,930 Arques Ave,Sunnyvale,Ca. 94086