Replies to post #13915 on Advanced Micro Devices Inc (AMD)

[dougSF30]

YB, he meant it was fundamentally the better 64-bit architecture for heating large rooms. Although the 32-bit Prescwatt will soon give it a run for the money.

Doug

[wbmw]

YB, Re: in the average, each of those 128-bit [IPF] instructions is well populated with nop's.

According to HP, they've been able to reduce the number of NOPs down to less than 30%, and that still leaves plenty of room for optimizations. Filling that 30% with additional instructions only add up to pure performance.

Also, something that has not yet been implemented into an IPF processor (but which may very well be in a future iteration) is low power states that can be entered whenever a bundle encounters such NOPs. The CPU can easily shut off power to certain execution units, or partition the chip so spread off heat to other areas. The possibilities are huge.

I think the bundled instruction format is unique to IPF architecture, and it could end up being a huge asset in the future.

Re: The only reson d'etre of this IPF instruction set was to get the extreme performance. So far I see that design requires 6 Mb of cache to show the top score and Intel is still shy about the Deerfield score. Given equal speed, IPF is much worse than x86. It must be much faster to be at least a little bit better.

So far, each iteration of the IPF architecture has been a huge improvement over the previous iteration. One cannot ignore that kind of trend, especially when we can already expect multicore Alpha inspired designs in the future.

The Deerfield part is actually a pretty good performer, given the power TDP of 55W. The integer performance is mediocre by today's PC standards, but systems based on Deerfield scale very well, and large datasets are handled better than x86. Floating point is especially impressive, making Deerfield still competitive with AMD's top-of-the-line Athlon 64-FX (which, incidentally, puts their own server processors to shame <G>).

Every time people try to be critical of IPF, their remarks are shown up by the performance that Intel has been able to get out of their design. Industry adoption is slow, but it is finally picking up a lot of speed. I think the results from these next few quarters are going to be surprising to some of those critics.

At any rate, it's worth keeping an eye on.