Performance Characterization of the Pentium Pro Processor

Performance Characterization of the Pentium®
Pro Processor
Dileep Bhandarkar and Jason Ding
Intel Corporation
Santa Clara, California, USA
Abstract
In this paper, we characterize the performance of several
business and technical benchmarks on a Pentium®
Pro
processor based system. Various architectural data are
collected using a performance monitoring counter tool.
Results show that the Pentium Pro processor achieves
significantly lower cycles per instruction than the
Pentium processor due to its out of order and speculative
execution, and non-blocking cache and memory system.
Its higher clock frequency also contributes to even
higher performance.
Keywords: Pentium®
Pro processor, computer
architecture, performance evaluation, workload
characterization, out of order execution, speculative
execution, SPEC CPU95, SYSmark/NT.
1. Introduction
The Intel Pentium®
Pro processor was disclosed in
February 1995 at ISSCC [1] and began shipping later
that year. The micro-architecture implements several new
features that are not found in previous implementations
of the Intel Architecture. This paper analyzes the major
performance characteristics of several business and
technical benchmarks on a Pentium Pro processor based
system. Measurements were performed using the built-in
performance counters of the processor. Results are
presented for cycles per instruction, cache miss statistics,
branch prediction statistics, speculative execution, stall
cycles, and other micro-architecture features.
Current literature contains numerous papers that present
simulations of various machine structures. Often these
simulations do not model the entire machine accurately
or only use traces of parts of popular benchmarks. We
present measured characteristics of a recent
microprocessor to allow researchers to calibrate their
theoretical results. The paper presents a lot of raw data
and some analysis wherever possible. In a modern
superscalar out-of-order processor, it is not always
possible to derive precise cause-effect relationships.
Some of the results presented here are consistent with the
behavior of SPEC benchmarks on other architectures,
e.g., the FP benchmarks have lower Icache misses and
higher Dcache misses than the integer benchmarks.
Other measurements (branch mispredicts, micro-op
statistics, and speculative execution) provide insight into
the inner workings of the Pentium Pro processor.
2. Architectural Features of the
Pentium®
Pro Processor
The Intel Pentium Pro processor implements dynamic
execution using an out-of-order, speculative execution
engine, with register renaming of integer, floating point
and flags variables, multiprocessing bus support, and
carefully controlled memory access reordering. The flow
of Intel IA-32 Architecture instructions is predicted and
these instructions are decoded into micro-operations
(uops), or series of uops. These uops are register-
renamed, placed into an out-of-order speculative pool of
pending operations, executed in dataflow order (when
operands are ready), and retired to permanent machine
state in source program order. This is accomplished with
one general mechanism to handle unexpected
asynchronous events such as mispredicted branches,
instruction faults and traps, and external interrupts.
Dynamic execution, or the combination of branch
prediction, speculation and micro-dataflow, is the key to
its high performance.
Figure 1 shows a block diagram of the processor. The
basic operation of the microarchitecture is as described in
the ISSCC paper [1]:
1. The 512 entry Branch Target Buffer (BTB) helps the
Instruction Fetch Unit (IFU) choose an instruction
cache line for the next instruction fetch. Icache line
fetches are pipelined with a new instruction line
fetch commencing on every CPU clock cycle.
2. Three parallel decoders (ID) convert multiple Intel
Architecture instructions into multiple sets of uops
each clock cycle. Instructions that require more than
4 uops are handled by the microinstruction
sequencer.
3. The sources and destinations of up to 3 uops are
renamed every cycle to a set of 40 physical registers
by the Register Alias Table (RAT), which eliminates
register re-use artifacts, and are forwarded to the 20-
entry Reservation Station (RS) and to the 40-entry
ReOrder Buffer (ROB).
Copyright 1997 IEEE. Published in the Proceedings of the Third International Symposium on High Performance Computer Architecture, February 1-5, 1997 in
San Antonio, Texas, USA. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional
purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works,
must be obtained from the IEEE.

4. The renamed uops are queued in the RS where they
wait for their source data - this can come from
several places, including immediates, data bypassed
from just-executed uops, data present in a ROB
entry, and data residing in architectural registers
(such as EAX).
5. The queued uops are dynamically executed according
to their true data dependencies and execution unit
availability (integer, FP, address generation, etc.).
The order in which uops execute in time has no
particular relationship to the order implied by the
source program.
6. Memory operations are dispatched from the RS to
the Address Generation Unit (AGU) and to the
Memory Ordering Buffer (MOB). The MOB ensures
that the proper memory access ordering rules are
observed.
7. Once a uop has executed, and its destination data has
been produced, that result data is forwarded to
subsequent uops that need it, and the uop becomes a
candidate for "retirement".
8. Retirement hardware in the ROB uses uop
timestamps to reimpose the original program order
on the uops as their results are committed to
permanent architectural machine state in the
Retirement Register File (RRF). This retirement
process must observe not only the original program
order, it must correctly handle interrupts and faults,
and flush all or part of its state on detection of a
mispredicted branch. When a uop is retired, the ROB
writes that uop’s result into the appropriate RRF
entry and notifies the RAT of that retirement so that
subsequent register renaming can be activated. Up to
3 uops can be retired per clock cycle.
The Pentium Pro processor implements a 14-stage
pipeline capable of decoding 3 instructions per clock
cycle. The in-order front end has 8 stages. The out-of-
order core has 3 stages, and the in-order retirement logic
has 3 stages. For an integer op, say a register-to-register
add, the execute phase is just one cycle. Floating point
adds have a latency of 3 cycles, and a throughput of 1 per
cycle. FP multiply has a latency of 5 cycles and a
repetition rate of 1 every 2 cycles. Integer multiply has a
latency of 4 cycles and a throughput of 1 every cycle.
Loads have a latency of 3 cycles on a Dcache hit. FDIV is
not pipelined; it takes 17 cycles for single, 32 cycles for
double, and 37 cycles for extended precision. The
processor includes separate data and instruction L1
caches (each of which is 8KB). The instruction cache is
4-way set associative, and the data cache is dual ported,
non-blocking, 2-way set associative supporting one load
and one store operation per cycle. Both instruction and
data cache line sizes are 32 byte wide. More details of the
microarchitecture can be found elsewhere [2].
Bus Interface Unit
Instruction Fetch Unit
Branch
Target
Buffer
InstructionDecoder
Micro-
instruction
Sequencer
Register
Alias Table
ReservationStation
Memory Reorder
Buffer
Data Cache Unit
Memory Interface
Unit
Address
Generation Unit
Integer Execution
Unit
Floating Point
Execution Unit
Reorder Buffer
and
Retirement
Register File
L2 CacheExternal Bus
Figure 1 Pentium®
Pro Processor Block Diagram
The secondary cache (L2 cache), which can be either
256KB or 512KB in size, is located on a separate die (but
within the same package). The L2 cache is 4-way set
associative unified non-blocking cache for storage of both
instructions and data. It is closely coupled with a
dedicated 64-bit full clock-speed backside cache bus. The
L2 cache line is also 32 bytes wide. The L2 cache fills the
L1 cache in a full frequency 4-1-1-1 cycle transfer burst
transaction. The processor connects to I/O and memory
via a separate 64-bit bus that operates at either 60 or 66
MHz. The bus implements a pipelined demultiplexed
design with up to 8 outstanding bus transactions.
3. Performance Monitoring
Facility
The Pentium®
Pro processor implements two
performance counters[3]. Each performance counter has
an associated event select register that controls what is
counted. The counters are accessed via the RDMSR and
WRMSR instructions. Table 1 shows a partial list of
performance metrics that can be measured by selecting
the two events to be monitored.
Table 1. Pentium®
Pro Processor Counter based Performance Metrics
Performance Metric Numerator Event Denominator Event
2

Data references per instruction DATA_MEM_REFS INST_RETIRED
L1 Dcache misses per instruction DCU_LINES_IN INST_RETIRED
L1 Icache misses per instruction L2_IFETCH INST_RETIRED
ITLB misses per instruction ITLB_MISS INST_RETIRED
Istalls cycles per instruction IFU_MEM_STALL INST_RETIRED
L1 cache misses per instruction L2_RQSTS INST_RETIRED
L2 cache misses per instruction L2_LINES_IN INST_RETIRED
L2 Miss ratio L2_LINES_IN L2_RQSTS
Memory transactions per instruction BUS_TRAN_MEM INST_RETIRED
FLOPS per instruction FLOPS INST_RETIRED
UOPS per instruction UOPS_RETIRED INST_RETIRED
Speculative execution factor INST_DECODED INST_RETIRED
Branch frequency BR_INST_RETIRED INST_RETIRED
Branch mispredict ratio BR_MISS_PRED_RETIRED BR_INST_RETIRED
Branch taken ratio BR_TAKEN_RETIRED BR_INST_RETIRED
BTB miss ratio BTB_MISSES BR_INST_DECODED
Branch Speculation factor BR_INST_DECODED BR_INST_RETIRED
Resource stalls per instruction RESOURCE_STALLS INST_RETIRED
Cycles per instruction CPU_CLK_UNHALTED INST_RETIRED
Table 2. Basic Characteristics of Systems
Processor Intel Pentium®
Pro Processor Intel Pentium®
Processor
CPU Core Frequency 150 MHz 120 MHz
Bus Frequency 60 MHz 60 MHz
Data bus 64-bit 64-bit
Address bus 36-bit 32-bit
On-chip L1 cache 8 KB data, 8 KB instruction 8 KB data, 8 KB instruction
Off-chip L2 cache 4-way 256 KB 512 KB (Dell), 256 KB (Gateway)
L2 cache timing 4-1-1-1 @ 150 MHz CPU freq. 3-1-1-1 @ 60 MHz bus frequency
System Chip Set 82450GX/KX 82430FX
Memory timing
(bus cycles)
14-1-1-1 (4-way interleaving)
14-2-2-2 (2-way interleaving)
14-4-4-4 (no interleaving)
13-3-3-3 (Fast Page Mode DRAM)
13-2-2-2 (EDO DRAM)
Basic Pipeline 14 stages 5 stages
Superscalar 3-way 2-way
Execution units 5 3
Branch prediction 4-way 512 entry BTB,
4-bit history, 2 level adaptive
4-way 256 entry BTB,
2-bit history
Execution model Out of order In order
Speculative Execution Yes No
McCalpin Streams
Memory Bandwidth
140 MB/sec (4-way
interleaving) 128 MB/sec (2-
way interleaving)
97 MB/sec (no interleaving)
82 MB/sec (Gateway 2000 P120)
SYSmark/NT rating 497 (Digital Celebris∗
XL6150) 294 (Gateway 2000 P120)
SPECint95 6.08 (Intel Alder System) 3.53 (Dell Dimension XPS P120)
SPECfp95 5.42 (Intel Alder System) 2.92 (Dell Dimension XPS P120)
4. Comparing the Pentium®
and
Pentium®
Pro Processors
This section compares the basic performance
characteristics of the Pentium [4] and Pentium Pro
processors. Table 2 compares the basic characteristics of
these two processors. We chose the 120 MHz Pentium
and the 150 MHz Pentium Pro processors because both
are fabricated in the same 0.6µ technology and use a 60
MHz external bus. For the SPEC benchmarks, the
Pentium system was a Dell Dimension XPS P120 with a
3

512KB pipelined burst L2 cache, and the Pentium Pro
system was an Intel Alder system with a 150MHz
Pentium Pro CPU with 256KB L2 cache and a 4-way
interleaved memory.
0 0.5 1 1.5 2 2.5
SPECint95
go
m88ksim
gcc
compress
li
ijpeg
perl
vortex
SPECfp95
tomcatv
swim
su2cor
hydro2d
mgrid
applu
turb3d
apsi
fpppp
wave5
CPI ratio
Perf Ratio
Figure 2 Performance Comparison of Pentium®
and
Pentium®
Pro Processors on SPEC95
Figure 2 shows the SPECratios and the cycles per
instruction of the Pentium Pro processor relative to the
Pentium processor for the SPEC95 benchmark suite for
the two systems. The SPEC*
results were obtained using
Intel Reference Compiler 2.3 Beta on UnixWare*
v2.0 on
an Intel Alder system. The Pentium Pro processor
achieves CPIs 15% to 50% lower than the Pentium
processor, in spite of the fact that it uses a design style
that emphasizes a fast clock frequency. Designs that
emphasize clock frequency generally result in deeper
pipelines and longer CPI. The Pentium Pro processor
design attempts to increase frequency while reducing
CPI, without being overly focused on optimal CPI or
fastest clock [5].
The Pentium Pro processor runs at 1.6 to 2.4 times the
performance of the Pentium processor on the SPEC95
suite[6], achieving 70% higher SPECint95 and 85%
higher SPECfp95. This performance comes from a 25%
faster clock frequency and a 15 to 50% reduction in CPI
compared to the Pentium processor. The Pentium Pro
processor can issue up to 3 instructions every clock cycle,
while the Pentium processor can issue only two. The out
of order execution model of the Pentium Pro processor
also allows useful work to proceed while prior operations
are stalled, thereby lowering the CPI.
- 0.50 1.00 1.50 2.00 2.50
Word
Excel
Powerpoint
Texim
MaxEDA
SYSmark/NT
Performance Ratio
Figure 3 Performance Comparison of Pentium®
and
Pentium®
Pro Processors on SYSmark/NT
Figure 3 shows the performance of a 150 MHz Pentium
Pro processor based Digital Celebris 6150 compared to a
120 MHz Pentium processor based Gateway 2000 P120
system on the SYSmark*
for Windows NT suite from
BAPCO[7], which contains project management software
(Welcom Software Technology Texim*
Project 2.0e),
computer-aided PCB design tool (Orcad MaxEDA*
6.0)
and Microsoft Office*
applications for word processing
(Word*
6.0), presentation graphics (PowerPoint*
4.0), and
spreadsheets (Excel 5.0). Both system had a 256KB L2
cache, but the Pentium Pro processor had a faster L2
cache (4-1-1-1 timing at full CPU clock frequency) on its
dedicated L2 cache bus. The Pentium Pro processor runs
29% to 113% faster than the Pentium processor, with an
overall 54% higher SYSmark score.
These results are slightly lower than the SPEC95 results
because the desktop applications in the SYSmark
benchmark perform some I/O operations that include
wait times that do not scale with CPU performance. The
SPEC benchmarks used compilers that generate binaries
that are optimized for each target machine. The
SYSmark/NT benchmarks use old binaries that are not
optimized for the Pentium Pro processor. These
benchmarks have large working set sizes for code and
data and also contain many context switches. The
SYSmark/NT benchmarks also result in higher L2 cache
misses as shown in a later section.
5. Detailed Characterization of
SPEC CPU95 Benchmarks
This section presents a detailed characterization of
Pentium®
Pro processor running the SPEC CPU95 suite.
The performance counter measurements presented in the
rest of this paper were done on a Digital Celebris XL6200
running Microsoft Windows*
NT Workstation Version
3.51. The central processor in the Digital Celebris
XL6200 is a 200MHz Pentium Pro processor with 256KB
L2 cache. The Celebris XL6200 system that we used in
our test was configured with 128MB DRAM with 2-way
interleaving and 14-2-2-2 memory timing at 66 MHz bus
4

frequency. The SPEC benchmarks were compiled with
Intel FORTRAN and C Reference Compilers Version 2.3.
5.1 Cycles per Instruction
Figure 4 shows the cycles per instruction (CPI) for the
SPEC95 benchmark suite. Several integer benchmarks
achieve less than one cycle per instruction. The CPIs are
remarkably low for a processor that implements a 14-
stage pipeline. The low CPI is due to the overlapped out-
of-order execution that mitigates the effect of the latency
of individual operations, fast L2 cache, and adaptive two
level branch prediction scheme. The FP benchmarks have
higher CPI due to longer execution latencies and higher
L2 cache misses.
0 0.5 1 1.5 2 2.5 3 3.5 4
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Cycles per instruction
512 KB L2
256 KB L2
Figure 4 Cycles per Instruction
We measured the CPI for the processor with 512 KB L2
cache too, by replacing the CPU in the Digital Celebris
system. Only compress (28%), li (5%) su2cor (17%), apsi
(15%), and wave5 (6%) showed more than 5%
improvement in CPI with the larger cache. The L2 miss
ratio reduction was 85%, 89%, 40%, 48%, and 19%
respectively. Simulations show that the CPI would
decrease further for su2cor (25%), apsi (14%), and wave5
(5%) if the 4-way L2 cache is doubled again to 1 MB.
5.2 Instruction Decode
The Pentium Pro processor has 3 decoders that can
handle up to 3 instructions every cycle (one instruction
with up to 4 uops, and two single uop instructions)[5].
The decoder has a 6 uop queue at its output. Only 3 uops
can be renamed per cycle, so the decoder has to stall if
the queue is too full. Figure 5 shows the percentage of
cycles in which 0, 1, 2, or 3 instructions were decoded.
Benchmarks with high Icache or L2 misses show many
cycles (35% to 51% for integer, 67% to 83% for FP) in
which no instructions are decoded. During L2 misses, the
CPU can run out of other machine resources causing back
pressure on earlier pipe stages. On the integer
benchmarks 33% to 54% of the instructions are decoded
in cycles in which 3 instructions are decoded; 25% to
64% for FP benchmarks.
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
SPECint
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
SPECfp
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Percent of all Cycles
3 inst_decoded
2 inst_decoded
1 inst_decoded
0 inst_decoded
Figure 5 Instruction Decode Profile
5.3 Cache Misses
The L1 data cache can accept a new load or store every
cycle and has a latency of three cycles for loads. It can
handle as many as four simultaneously outstanding
misses. Figure 6 shows the L1 data and instruction cache
misses, and L2 cache misses. Except for gcc and
m88ksim, the L1 data misses are always higher than the
L1 instruction misses. In most cases the L1 instruction
misses are so small that they don’t even show on the
scale used in Figure 6. The integer benchmarks, in
general, show much lower L1 data cache and L2 misses
than the floating point ones (larger data sets); but higher
L1 instruction cache misses (larger code size and fewer
loops). The benchmark (wave5) with the highest L1
misses does not have the highest L2 misses. Figure 7
shows a fairly strong correlation between L2 misses and
CPI, indicating that the L2 miss latency (about 50 CPU
cycles) is not completely overlapped.
0 20 40 60 80 100 120 140 160
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Misses per Thousand Instructions
L2 Misses
L1 Instr Misses
L1 Data Misses
5

Figure 6 Cache Misses per Thousand Instructions
vortex
turb3dli
compress
fpppp
mgrid
ijpeg
perl
su2cor
hydrotomcatv
swim
applu
wave
apsi
gcc
go
m88ksim
0
0.5
1
1.5
2
2.5
3
3.5
4
0 10 20 30 40 50 60
L2 Cache Misses per Thousand Instructions
CyclesperInstruction
Figure 7 CPI versus L2 Cache Misses
5.4 TLB Misses
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Misses Per Thousand Instructions
Figure 8 ITLB Statistics
The Pentium®
Pro processor has separate TLBs for
instructions and data. The processor also has separate
TLBs for 4-Kbyte and 4-Mbyte page sizes. The ITLB for
4KB pages has 32 entries. The DTLB for 4KB pages has
64 entries. Both are 4 way set associative. The ITLB for
large pages has 4 entries, while the DTLB has 8 entries;
both are 4-way set associative. As shown in Figure 8, the
ITLB misses are well below 0.1 per thousand
instructions, except for a couple of integer benchmarks.
The DTLB misses are generally higher than ITLB
misses, but they could not be measured accurately. TLB
misses do not contribute much to the CPI, as shown later.
5.5 Memory References
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
DataReferences
0
10
20
30
40
50
60
70
80
MemoryTransactions
Data References Per Instruction
Memory Transactions Per Thousand Instructions
Figure 9 Memory Reference Statistics
Figure 9 shows the number of data references per
instruction and the number of memory transactions per
thousand instructions. On the average, both the integer
and FP benchmarks generate about 1 data reference every
two instructions. The IA-32 architecture results in more
data references than most RISC architectures because it
has fewer registers (8 vs. 32). As might be expected,
there is a strong correlation between L2 cache misses and
memory transactions. The memory transactions per
instruction are higher for the FP benchmarks due to a
higher L2 cache miss rate. Note that there can be more
than one memory transaction per L2 cache miss if a dirty
cache block has to be written back to memory.
5.6 Branch Prediction
The Pentium®
Pro processor implements a novel branch
prediction scheme, derived from the two-level adaptive
scheme of Yeh and Patt[8]. The branch target buffer
(BTB) retains both branch history information and the
predicted branch target address. The BTB contains 512
entries. If a branch is not found in the BTB, a static
prediction (backwards taken, forward not taken) is used.
There is no penalty for correctly predicted not-taken
branches. Correctly predicted taken branches incur a 1
cycle penalty. Mispredicted branches incur a penalty of
about 10-15 cycles, plus additional cycles required to
retire the mispredicted branch.
6

0 0.1 0.2 0.3 0.4 0.5
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Ratio
Branch mispredict ratio
BTB Miss Ratio
Branches per Instr
Figure 10 Branch Statistics
Figure 10 shows the frequency of branches, fraction of
branches that hit in the BTB, and the accuracy of branch
prediction. Even though the BTB miss ratio is fairly high,
the branch mispredict ratio is less than 10% for all but
one benchmark. The BTB miss ratio is high partly due to
the fact that unconditional branches are not stored in the
BTB, but are included in the total branch instruction
count. As might be expected, the integer benchmarks
contain more branches than the FP benchmarks, and they
incur a higher branch mispredict ratio (fewer loop
branches). The number of mispredicted branches range
from about 2 to 40 per thousand instructions for the
integer benchmarks, and about 0.1 to 4 for the FP
benchmarks. For most of the benchmarks, branch
mispredict stalls are not a major contributor to overall
CPI.
5.7 Speculative Execution
The Pentium®
Pro processor fetches instructions along
the predicted path and executes them until the branch is
resolved. If a branch is incorrectly predicted, the
speculated instructions down the mispredicted path are
flushed. Note that there can be other mispredicted
branches down a mispredicted branch. Figure 11 shows
the average number of instructions issued per retired
instruction for the SPEC benchmarks. There are about 13
to 37 speculated instructions per mispredicted branch.
Mispredicted branches are not recognized for about 10 to
15 cycles, and the processor can issue up to 3 instructions
per cycle. Benchmarks with higher mispredicted
branches per instruction have higher speculated
instructions.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Speculation Factor
Figure 11 Speculation Factor
5.8 Resource Stalls
Figure 12 shows the I-stream stalls and resource stalls,
measured in terms of the cycles in which the stall
conditions occur. I-stream stalls are caused mainly by I-
cache misses and ITLB misses. Resource stalls show the
number of cycles in which resources like register
renaming or reorder buffer entries, memory buffer
entries, and execution units are full; but these stalls may
be overlapped with the execution latency of previously
executing instructions. The FP benchmarks, except for
fpppp (long basic blocks), incur negligible I-stream stalls.
They do incur significantly more resource stalls than
integer benchmarks, probably due to long dependency
chains.
0 0.5 1 1.5 2 2.5 3 3.5 4
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Cycles per instruction
CPI
I-stream Stalls
Resource Stalls
Figure 12 Stall Cycles
7

5.9 Micro-Operations
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
099.go
126.gcc
130.li
134.perl
102.sw im
104.hydro2d
110.applu
141.apsi
146.w ave5
Number of micro-ops per instruction
Figure 13 Micro-operations per Instruction
The instruction fetch unit fetches 16 bytes every clock
cycle from the I-cache and delivers them to the
instruction decoder. Three parallel decoders decode this
stream of bytes and convert them into triadic uops. Most
instructions are converted directly into single uops, some
are decoded into one-to-four uops, and the complex
instructions require microcode (sequence of uops). Up to
5 uops can be issued every clock cycle to the various
execution units, and up to 3 uops can be retired every
cycle. Figure 13 shows the average number of uops
executed per instruction for each of the SPEC95
benchmarks. The range is from 1.2 to 1.7, with an
average around 1.35.
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
SPECint
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
SPECfp
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Percent of all Cycles
3 uops retired
2 uops retired
1 uop retired
0 uops retired
Figure 14 Micro-operations retirement profile
Figure 14 shows that no uops are retired in 25% to 55%
of the cycles on the integer benchmarks, and 43% to 82%
of the cycles for FP. Benchmarks with low CPI have
fewer cycles with no uops retired. Furthermore, about
65% and 80% of the uops are retired in cycles in which 3
uops are retired for the average integer and FP
benchmark respectively. This indicates that executed
uops often have to wait for uops from previous
instructions to be ready for retirement, thereby
confirming the value of out of order execution. These
younger uops build up more for FP benchmarks because
of higher cache misses and longer latencies of FP
operations.
5.10 Adding Up the Cycles
Accounting for cycles in an out-of-order machine like the
Pentium®
Pro processor is difficult due to all the
overlapped execution. It is still useful to examine the
various components of execution and stalls and compare
them to the actual cycles per instruction as shown in
Figure 15. The CPI is about 20 to 50% lower than the
individual components due to overlapped execution. The
figure also shows resource stall cycles in which some
resource such as execution unit or buffer entry is not
available. Execution can proceed during a resource stall
cycle in some other part of the machine. Since more than
one uop can be dispatched in a cycle, the figure does not
account for execution parallelism. Micro-ops seem to
dominate in most integer benchmarks. Resource stalls,
and L2 misses contribute the most to the CPI in the FP
benchmarks. Branch mispredicts are not a major factor.
0
1
2
3
4
5
6
7
8
099.go
124.m88ksim
126.gcc
129.compress
130.li
132.ijpeg
134.perl
147.vortex
101.tomcatv
102.swim
103.su2cor
104.hydro2d
107.mgrid
110.applu
125.turb3d
141.apsi
145.fpppp
146.wave5
Cycles
-
1
2
3
4
5
6
7
8
UOPS
Resource Stalls
Branch mispredicts
ITLB Misses
DTLB Misses
L2 Misses
L1 Inst Misses
L1 Data Misses
CPI
Figure 15 CPI vs. Latency Components
6. Characteristics Across
Different Workloads
SPEC95 is a popular CPU intensive benchmark suite. It
is widely used to characterize CPU performance.
However, the behavior of other workloads can be quite
different. This section presents the characteristics of
8

desktop applications running on a Pentium®
Pro
processor. In particular, we present results for the
SYSmark/NT benchmark. These benchmarks are not
floating point intensive; Excel contains about 9% FP
instructions, MaxEDA has 4%, and the rest less than
0.5%. While the SPEC95 benchmarks were optimized for
the Pentium Pro processor using the latest compilers, the
SYSmark/NT benchmarks are based on old binaries that
have been shipping for many years and were probably not
generated with all optimizations turned on.
In this section, we compare the SYSmark/NT benchmark
statistics with the minimum, median, or maximum for
the SPECint95 and SPECfp95 suites. The data presented
here shows that the SPEC integer benchmarks should not
be used to predict the performance of real business
applications
Figure 16 shows the CPI across different workloads. The
business applications (using old binaries with non-
optimal code) incur higher CPIs than the median for the
SPECint95 benchmarks. Two of the five SYSmark/NT
benchmarks incur higher CPI than the median observed
for the SPECfp95 suite. The CPI is higher due to higher
L2 miss rates, Istream stalls, and resource stalls.
- 0.50 1.00 1.50 2.00 2.50
SPECint95_med
SPECint95_max
SPECfp95_med
Word
Excel
Pow erpoint
Texim
MaxEDA
Cycles Per Instruction
Figure 16 CPI for SYSmark/NT
Figure 17 shows the L2 cache misses. The three
Microsoft Office benchmarks incur much higher L2
misses than the SPECint95 median, but well below the
SPECfp95 median. The code and data sizes of these
business applications are much larger than the SPEC
integer benchmarks. Once again, there is fairly strong
correlation between L2 misses and CPI. Overall, Word
and Excel exhibit the highest L2 misses and stall cycles
among the SYSmark/NT benchmarks.
0 2 4 6 8 10 12 14 16
SPECint95_med
SPECint95_max
SPECfp95_med
Word
Excel
Pow erpoint
Texim
MaxEDA
Misses Per Thousand Instructions
Figure 17 SYSmark/NT L2 Cache Misses
Figure 18 shows the resource stalls. The SYSmark/NT
benchmarks incur higher resource stalls than the
SPECint95 median, but are well below the SPECfp95
median. The higher resource stalls can be attributed to
higher L2 misses during which the internal resources can
be consumed by instructions waiting to be retired.
0 0.2 0.4 0.6 0.8 1 1.2 1.4
SPECint95_med
SPECint95_max
SPECfp95_med
Word
Excel
Pow erpoint
Texim
MaxEDA
Resource Stall Cycles per Instruction
Figure 18 SYSmark/NT Resource Stalls
Figure 19 shows the instruction stalls. They are higher
than the SPECint95 median. Once again, this is due to
higher occurrence of string instructions in Word and
Excel that invoke the microsequencer and require the
decoders to stall. These workloads also have high context
switch activity resulting in ITLB flushes and Icache
misses.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
SPECint95_med
SPECint95_max
SPECfp95_med
Word
Excel
Pow erpoint
Texim
MaxEDA
I-StreamStall Cycles per instruction
Figure 19 SYSmark/NT Instruction Stalls
Figure 20 shows the uops per instruction. All the
SYSmark/NT benchmarks execute more uops than most
9

of the SPEC95 benchmarks. This is probably due to the
higher use of character string instructions. There is a
strong correlation between uops/instruction and CPI, a
trend not observed in the SPEC95 suite.
0 0.5 1 1.5 2 2.5
SPECint95_med
SPECint95_max
SPECfp95_med
Word
Excel
Powerpoint
Texim
MaxEDA
Micro-ops per instruction
Figure 20 SYSmark/NT Micro-ops Per Instruction
Figure 21 shows the speculation factor. It is in the bottom
half of the distribution for SPECint95. The speculation
factor is lower because there are fewer mispredicted
branches.
0.95 1 1.05 1.1 1.15 1.2 1.25 1.3
SPECint95_min
SPECint95_med
SPECfp95_med
Word
Excel
Pow erpoint
Texim
MaxEDA
Speculation Factor
Figure 21 SYSmark/NT Speculation Factor
7. Concluding Remarks
The Pentium®
Pro processor was designed to achieve
significantly higher performance than the Pentium
processor in the same process technology. It achieves this
performance through a superpipelined design that yields
a 25% faster clock, and with an out of order dynamic
execution engine that reduces the CPI. The data
presented here shows that the Pentium Pro processor
achieves a 15 to 45% reduction in CPI compared to the
previous generation design (Pentium processor) in the
same process technology, while running at a 25% faster
clock frequency. The processor’s out-of-order, speculative
execution engine does manage to overlap useful work
with pending memory accesses to reduce the impact of
cache misses. The impact of resource stalls is also
reduced by out of order execution. The branch prediction
scheme reduces branch mispredictions so as not to make
them a significant performance limiter. It performs well
even on old binaries that were not optimized for its
microarchitecture. Performance counter based
measurements show that the overall CPI achieved by the
Pentium Pro processor is about 20 to 50% lower than the
individual latency components due to overlapped
execution.
A detailed comparison of the Pentium Pro processor and
Digital’s Alpha 21164 RISC processor is reported in
another study [9].
8. Acknowledgments
The authors would like to thank the Pentium®
Pro
processor designers for producing such an interesting
microprocessor and for incorporating the performance
counting mechanisms that enabled this study. Special
thanks to Bob Colwell for his extensive review and
valuable comments.
9. References
[1] Robert P. Colwell and Randy L. Steck, “A 0.6um
BiCMOS Processor with Dynamic Execution”,
ISSCC Proceedings, February 1995, pp. 176-177.
[2] Linley Gwennap, “Intel’s P6 Uses Decoupled
Superscalar Design”, Microprocessor Report, Vol. 9,
No. 2, 16 February 1995, pp. 9-15.
[3] Intel Corporation, “Pentium Pro Family Developer’s
Manual, Volume 3: Operating System Writer’s
Manual”, Intel Corporation, Order Number 242692,
1996.
[4] Donald Alpert and Dror Avnon, “Architecture of the
Pentium Microprocessor,” IEEE Micro, June 1993,
pp. 11-21.
[5] David Papworth, “Tuning The Pentium Pro
Microarchitecture,” IEEE Micro, April 1996, pp. 8-
15.
[6] Jeff Reilly, “A Brief Introduction to the SPEC
CPU95 Benchmark,” IEEE-CS TCCA Newsletter,
June 1996. Also, see
http://www.specbench.org/osg/cpu95/.
[7] http://www.bapco.com/nt1.htm
[8] Tse-Yu Yeh and Yale Patt, “Two-Level Adaptive
Training Branch Prediction,” Proc. IEEE Micro-24,
Nov 1991, pp. 51-61.
[9] Dileep Bhandarkar, “RISC versus CISC: A Tale of
Two Chips,” submitted for publication.
10


Intel®
and Pentium®
are registered trademarks of Intel Corporation. Other brands and names are the property of their
respective owners.

Performance Characterization of the Pentium Pro Processor

More Related Content

What's hot (20)

Similar to Performance Characterization of the Pentium Pro Processor (20)

More from Dileep Bhandarkar (20)

Recently uploaded (20)

Performance Characterization of the Pentium Pro Processor