PCMech Forums - View Single Post

Toaster · 07-03-2000, 09:24 AM

Well Byte, if you think you need oodles of L2 cache then you might be using a poorly designed cache. Cache is greatly mis-understood. Most think: "More is better" and "less is bad". To some degree, this thinking is correct but after this point it`s as far from the truth as one could get.
It all boils down to *HOW* the cache is designed and used by the host CPU. Intel is notorious for using small L1/L2 cache but equally notorious for having the highest cache efficiency. One exception is the Xenon processor. This puppy boasts upto 2GB of L2 cache running at full system clock in a 4way set associative mannor with a 128bit address width. Until the release of the Duron and other CPU`s from AMD, the cache was running at 1/2 system clock and a latency exceeding a factor of 16. Even the cache bus width hasn`t changed with AMD and still remains at 64bit or less. 128/256kb of full speed cache outperforms caches many times in size because of efficiency.
A better designed cache reduces the number of CPU "do nothing loops" and keeps the CPU buzy with incomming data. For each cache "miss", the CPU runs idlely until the cache resupplies the CPU with new data. With a cache thats too small, the cache is "thrashed" (dumped and refilled needlessly) and performance suffers. With a cache thats too large and slow, the CPU outruns the cache and system performance suffers. This is where the Xenon excells with it`s huge cache running at full system clock.
The "Alpha" CPU is a totally unique design in that it`s a "RISC" based CPU and it`s cache I/O registers are 256bytes wide and are superpipelind as well as being 8way set associative in design. This makes the Alpha far more efficient then ANY CISC based (Intel, AMD, Cyrix) CPU.
RISC= Reduced Intruction Set Computing
CISC= Complex Intruction Set Computing

Also, the Alpha does the same amount of "work" in fewer CPU cycles resulting in very substantial speed gaines.
If a given CPU uses 100 cycles to perform a task and a RISC based CPU does the same work in 50 cycles, the speed increase is 2 fold.
This equates to the following:
A 1GHZ Alpha runs well over twice as fast as a 1GHZ X86 class CPU because it can do the same work in fewer cycles. The same principle applies to cache and cache speed/efficiency.
If you want to go REALLY fast, tell Intel and other CPU makers to make the break away from the X86 compatability and go RISC.
(both companies make RISC based CPU`s but not in any numbers for home use)

07-03-2000, 09:24 AM	#6
Toaster Member (13 bit) Join Date: Apr 1999 Location: Now in Phoenix, AZ. Where next? Only 8 states left to see. Posts: 4,661	Well Byte, if you think you need oodles of L2 cache then you might be using a poorly designed cache. Cache is greatly mis-understood. Most think: "More is better" and "less is bad". To some degree, this thinking is correct but after this point it`s as far from the truth as one could get. It all boils down to HOW the cache is designed and used by the host CPU. Intel is notorious for using small L1/L2 cache but equally notorious for having the highest cache efficiency. One exception is the Xenon processor. This puppy boasts upto 2GB of L2 cache running at full system clock in a 4way set associative mannor with a 128bit address width. Until the release of the Duron and other CPU`s from AMD, the cache was running at 1/2 system clock and a latency exceeding a factor of 16. Even the cache bus width hasn`t changed with AMD and still remains at 64bit or less. 128/256kb of full speed cache outperforms caches many times in size because of efficiency. A better designed cache reduces the number of CPU "do nothing loops" and keeps the CPU buzy with incomming data. For each cache "miss", the CPU runs idlely until the cache resupplies the CPU with new data. With a cache thats too small, the cache is "thrashed" (dumped and refilled needlessly) and performance suffers. With a cache thats too large and slow, the CPU outruns the cache and system performance suffers. This is where the Xenon excells with it`s huge cache running at full system clock. The "Alpha" CPU is a totally unique design in that it`s a "RISC" based CPU and it`s cache I/O registers are 256bytes wide and are superpipelind as well as being 8way set associative in design. This makes the Alpha far more efficient then ANY CISC based (Intel, AMD, Cyrix) CPU. RISC= Reduced Intruction Set Computing CISC= Complex Intruction Set Computing Also, the Alpha does the same amount of "work" in fewer CPU cycles resulting in very substantial speed gaines. If a given CPU uses 100 cycles to perform a task and a RISC based CPU does the same work in 50 cycles, the speed increase is 2 fold. This equates to the following: A 1GHZ Alpha runs well over twice as fast as a 1GHZ X86 class CPU because it can do the same work in fewer cycles. The same principle applies to cache and cache speed/efficiency. If you want to go REALLY fast, tell Intel and other CPU makers to make the break away from the X86 compatability and go RISC. (both companies make RISC based CPU`s but not in any numbers for home use)