| |
| | External Vector Coprocessing |
 | | Before that, add-on vector units offered for the System/360 and the original System/370, such as the IBM 2938 Array Processor and the IBM 3838 Array Processor, had their own internal memories, to which data would be transferred before the start of a computation involving vectors. |
 | | This, of course, would mean that vectors not fully aligned, in terms of the width of the entire assembly of external coprocessor chips, would involve a significant performance penalty, although even that penalty would not outweigh the benefit of avoiding the limited-width data path to the CPU itself. |
 | | Since the external vector coprocessor only works with memory in its native width, the only variation of the Common floating-point format that would be used would be the one where an excess-128 binary exponent occupied the last eight bits of a 32, 64, or 128 bit long floating-point number with a sign-magnitude mantissa. |
| www.quadibloc.com /arch/ar0802.htm (2950 words) |
|