Messages in this thread | | | Date | Sun, 6 Sep 1998 14:05:31 +1200 | From | Chris Wedgwood <> | Subject | Re: AMD SMP |
| |
On Fri, Sep 04, 1998 at 07:02:12PM -0700, Gary Smith wrote:
> Now the processor moves more and does more specialized functions on data > than needs to be done. A good processor should be simple. They call it > RISC. I have always been and Intel and Windows person (still am) but I > think that the entire concept that Intel has to deal with (backward > compatibility) should be dumped and an entirely new processor be invented.
Like the Alpha?
> They have the technology to achieve 1ghz. It's been around for some time.
Like the Alpha?
> But the current limitation on the Intel and Motorola architecture does not > allow for it. I sat that we should design a new chip entirely with 3 things > in mind:
> Speed (achievable) > SMP/ASMP/TANDEM processing (currently here)
Like the Alpha?
> Heat dissipation (possible if chip size reduced ie removal of unneeded > instructions)
Heat dissipation is a tricky one... if you up the clock speed, increased stray capacitance means you need higher currents, smaller sizes or both, so even though we can now may much more efficient transistors and gates and 10 years ago, we make them work very much harder too.
Also, adding instructions won't necessarily increase the overall gate count very much unless the functionality they require is radically new and requires new functional units or major change to existing ones.
A non-insignificant percentage of the gates found on modern CPUs are L1 cache, write-back buffers, pipelining and the magic that makes this all work. Arguably, scaling those features - actually helps performance.
Moving from a two-way associative cache to four-way (larger) L1 cache probably added more transistors to MMX pentiums than MMX itself did, and those do indeed improve performance (along with fixed branch prediction).
> As a programmer, you can do everything that the MMX processor does > with code and if the chip was 2 times faster you code would beat > the MMX (30%) internal ability. If you removed MMX you could > achieve that 2x speed.
MMX is a different story. IMO, Its a damn nice idea, badly implemented (says me, because I can't use it, but clever people like mingo I'm sure can). It probably only exists because Intel didn't want third-party vendors shipping motherboards with DSPs...
MMX2 or whatever its called (Kamatai?) looks like they design is a tad more sane and might even be genuinely useful for games and other stuff.
I think most other architectures like PPC, Alpha, Sparc and MIPS (?) have various media extensions to them.
> Remember, you could put a RISC chip into a microwave and would work > but why would it need MMX.
Because its simpler to build building blocks that allow large parallel comuptations (like matrix ops) than make 8+ instuction units all work in parallel via register renaming and othe black magic?
-Chris
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/faq.html
| |