Published on September 17, 2007
How do we evaluate computer architectures?: How do we evaluate computer architectures? Think of 5 characteristics that differentiate computers? Can some processors compute things that others can’t? How do we evaluate computer architectures?: How do we evaluate computer architectures? Think of 5 characteristics that differentiate computers? Single-Cycle Performance: Single-Cycle Performance Last time we saw a MIPS single-cycle datapath and control unit. Today, we’ll explore factors that contribute to a processor’s execution time, and specifically at the performance of the single-cycle machine. Next time, we’ll explore how to improve on the single cycle machine’s performance using pipelining. Three Components of CPU Performance: Three Components of CPU Performance Cycles Per Instruction CPU timeX,P = Instructions executedP * CPIX,P * Clock cycle timeX Instructions Executed: Instructions executed: We are not interested in the static instruction count, or how many lines of code are in a program. Instead we care about the dynamic instruction count, or how many instructions are actually executed when the program runs. There are three lines of code below, but the number of instructions executed would be 2001. li $a0, 1000 Ostrich: sub $a0, $a0, 1 bne $a0, $0, Ostrich Instructions Executed CPI: The average number of clock cycles per instruction, or CPI, is a function of the machine and program. The CPI depends on the actual instructions appearing in the program—a floating-point intensive application might have a higher CPI than an integer-based program. It also depends on the CPU implementation. For example, a Pentium can execute the same instructions as an older 80486, but faster. In CS231, we assumed each instruction took one cycle, so we had CPI = 1. The CPI can be andgt;1 due to memory stalls and slow instructions. The CPI can be andlt;1 on machines that execute more than 1 instruction per cycle (superscalar). CPI Clock cycle time: One 'cycle' is the minimum time it takes the CPU to do any work. The clock cycle time or clock period is just the length of a cycle. The clock rate, or frequency, is the reciprocal of the cycle time. Generally, a higher frequency is better. Some examples illustrate some typical frequencies. A 500MHz processor has a cycle time of 2ns. A 2GHz (2000MHz) CPU has a cycle time of just 0.5ns (500ps). Clock cycle time Execution time, again: CPU timeX,P = Instructions executedP * CPIX,P * Clock cycle timeX The easiest way to remember this is match up the units: Make things faster by making any component smaller!! Often easy to reduce one component by increasing another Execution time, again Example 1: ISA-compatible processors: Let’s compare the performances two x86-based processors. An 800MHz AMD Duron, with a CPI of 1.2 for an MP3 compressor. A 1GHz Pentium III with a CPI of 1.5 for the same program. Compatible processors implement identical instruction sets and will use the same executable files, with the same number of instructions. But they implement the ISA differently, which leads to different CPIs. CPU timeAMD,P = InstructionsP * CPIAMD,P * Cycle timeAMD = = CPU timeP3,P = InstructionsP * CPIP3,P * Cycle timeP3 = = Example 1: ISA-compatible processors Example 2: Comparing across ISAs: Example 2: Comparing across ISAs Intel’s Itanium (IA-64) ISA is designed facilitate executing multiple instructions per cycle. If an Itanium processor achieves an average CPI of .3 (3 instructions per cycle), how much faster is it than a Pentium4 (which uses the x86 ISA) with an average CPI of 1? Itanium is three times faster Itanium is one third as fast Not enough information The single-cycle design from last time: The single-cycle design from last time A control unit (not shown) generates all the control signals from the instruction’s 'op' and 'func' fields. The example add from last time: The example add from last time Consider the instruction add $s4, $t1, $t2. Assume $t1 and $t2 initially contain 1 and 2 respectively. Executing this instruction involves several steps. The instruction word is read from the instruction memory, and the program counter is incremented by 4. The sources $t1 and $t2 are read from the register file. The values 1 and 2 are added by the ALU. The result (3) is stored back into $s4 in the register file. How the add goes through the datapath: 10100 I [15 - 11] How the add goes through the datapath 4 I [25 - 21] 01001 I [20 - 16] 01010 RegWrite 00...01 00...10 00...11 PC+4 Performance of Single-cycle Design: Performance of Single-cycle Design CPU timeX,P = Instructions executedP * CPIX,P * Clock cycle timeX Edge-triggered state elements: Edge-triggered state elements In an instruction like add $t1, $t1, $t2, how do we know $t1 is not updated until after its original value is read? We’ll assume that our state elements are positive edge triggered, and are updated only on the positive edge of a clock signal. The register file and data memory have explicit write control signals, RegWrite and MemWrite. These units can be written to only if the control signal is asserted and there is a positive clock edge. In a single-cycle machine the PC is updated on each clock cycle, so we don’t bother to give it an explicit write control signal. The datapath and the clock: The datapath and the clock On a positive clock edge, the PC is updated with a new address. A new instruction can then be loaded from memory. The control unit sets the datapath signals appropriately so that registers are read, ALU output is generated, data memory is read or written, and branch target addresses are computed. Several things happen on the next positive clock edge. The register file is updated for arithmetic or lw instructions. Data memory is written for a sw instruction. The PC is updated to point to the next instruction. In a single-cycle datapath everything in Step 2 must complete within one clock cycle, before the next positive clock edge. How long is that clock cycle? The slowest instruction...: The slowest instruction... If all instructions must complete within one clock cycle, then the cycle time has to be large enough to accommodate the slowest instruction. For example, lw $t0, –4($sp) needs 8ns, assuming the delays shown here. 2 ns 2 ns 2 ns 1 ns 0 ns 0 ns 0 ns 0 ns ...determines the clock cycle time: ...determines the clock cycle time If we make the cycle time 8ns then every instruction will take 8ns, even if they don’t need that much time. For example, the instruction add $s4, $t1, $t2 really needs just __ns. How bad is this?: How bad is this? With these same component delays, a sw instruction would need 7ns, and beq would need just 5ns. Let’s consider the gcc instruction mix from p. 189 of the textbook. With a single-cycle datapath, each instruction would require 8ns. But if we could execute instructions as fast as possible, the average time per instruction for gcc would be: (48% x 6ns) + (22% x 8ns) + (11% x 7ns) + (19% x 5ns) = 6.36ns The single-cycle datapath is about 1.26 times slower! It gets worse...: It gets worse... We’ve made very optimistic assumptions about memory latency: Main memory accesses on modern machines is andgt;50ns. For comparison, an ALU on the Pentium4 takes ~0.3ns. Our worst case cycle (loads/stores) includes 2 memory accesses A modern single cycle implementation would be stuck at andlt;10Mhz. Caches will improve common case access time, not worst case. Tying frequency to worst case path violates first law of performance!! Summary: Summary Performance is one of the most important criteria in judging systems. Here we’ll focus on Execution time. Our main performance equation explains how performance depends on several factors related to both hardware and software. CPU timeX,P = Instructions executedP * CPIX,P * Clock cycle timeX It can be hard to measure these factors in real life, but this is a useful guide for comparing systems and designs. A single-cycle CPU has two main disadvantages. The cycle time is limited by the worst case latency. It isn’t efficiently using its hardware. Next time, we’ll see how this can be rectified with pipelining.