CSC 456 Fall 2013/1c wa: Difference between revisions

From Expertiza_Wiki
Jump to navigation Jump to search
No edit summary
 
(55 intermediate revisions by 2 users not shown)
Line 1: Line 1:
'''Trends in cache size and organization'''
==Trends in cache size and organization==


'''Task'''
---------
1c.  Trends in cache size and organization
==Introduction==
Over the years, caches have grown larger--up to a point, and then L1 caches tended for awhile to decrease in size.  Why?  In the early 1980s, associativity increased; beginning about 1990, it decreased, and then by about 2000, it was increasing again.  Why?  When was the first machine with an L2 cache?  An L3 cache?  How fast were the various levels of caches, and how did this speed compare to main memory?  There is a wealth of information to bring to bear on this topic.


Cache size has grown over the years alongside the evolution of the microprocessor.  Intuitively one would expect cache sizes to keep growing larger and larger following some law similar to Moore’s Law.  In actuality however L1 cache sizes have all but maxed out for an individual processor. Observing the trend of cache growth it can be seen that some processor lines stopped growing from one iteration to the next and in some cases even decreased in size.  To go along with this, cache associativity has varied over the years.  While it is true that no cache organization is optimal for every situation certain organizations certainly perform better for most tasks on certain systems.  This wiki will try to analyze data on cache size and associativities to gain some insight into the trends and reasoning behind vendor choices of cache size and organization over the years.  Specifically  it looks from the late 80’s / early 90’s to the early 2000’s.


---------
<br />
To-Do: <br />
 
find first L3 and fill in <br />
==Cache Associativity==
order the list by year <br />
This table shows cache associativities found in some  mainstream processors from the late 80’s to the early 2000’s with one processor from 1968 just for reference.  As can be seen from the data, the late 80’s early 90’s tended towards a set associative cache with around four lines.  In the mid-90’s it tended towards lower associativity and direct mapping.  Then in the late 90’s and early 2000’s it tended back towards higher associativities with larger set sizes again.
analyze: average cc speed for a given level, main mem cc speed evolution <br />
---------


<br />
'''L1, L2, L3 Associativity'''
{|border=1


Theory: Cache Associativity decreased as cache size became larger because it became too expensive to have to search the cache each time once the cache was too large.  Also, bigger the cache size as a percentage of main memory, the less need for associativity.  But while caches and main memory have both grown, main memory size has grown faster in the 2000’s.  So when the percentage of cache to main memory goes down associativity needs to increase.
    | '''System'''
    | '''Year'''
    | '''L1 Associativity'''
    | '''L2 Associativity'''
    | '''L3 Associativity'''
    | Notes:
    |-
| IBM 360/85
    | 1968
    | Sector
    | N/A
    | N/A
    | First processor with a cache, clock speed 12.5MHz
    |-
| Intel 80486
    | 1989
    | 4-way associative
    | N/A
    | N/A
    |
    |-
| SuperSPARC
    | 1992
    | 4 & 5 way set
    | N/A
    | N/A
    | Used to render Toy Story, Core @ 40MHz
    |-
| Alpha 21064(DEC)
    | 1992
    | Direct
    | Direct
    | N/A
    |
    |-
| UltraSPARC
    | 1995
    | 2-Way & Direct
    | Direct
    | N/A
    | 64-bit w/ Core@200MHz
    |-
| Alpha 21164(DEC)
    | 1995
    | Direct
    | 3 way set
    | N/A
    |
    |-
| Pentium Pro
              | 1995
    | 2 & 4 way
    | ?
    | N/A
    | First on-die L2
    |-
| K6-III
    | 1999
    | 2 way
    | 4 way
    | n/a
    |
    |-
| Pentium 4
    | 10/2000
    | 4 Way
    | 8 Way
    | N/A
    |
    |-
| UltraSPARC III
    | 2001
    | 4 Way
    | N/A
    | N/A
    |
    |-
| Itanium 2
    | 2002
    | 4 -way
    | 8-way
    | 12 way
    |
|}






The Pentium/Pentium (1995)pro was the first processor to have the l2 cache on the processor chip. Before this, the l2 cache was an option to add on to the motherboard. [1]
<br  /><br /><br />


Systems to consider in table <br /> <br />
==Cache Size==
In accordance with Moore's law as the transistors on a chip increase we would expect cache sizes to increase with each generation of processors.  Main memory sizes have certainly kept increasing so we would expect to see a similar trend in caches.  Looking at the table below we can certainly see an increase in L1 cache sizes all the way up to the 2000's.  Analyzing the trend however we can see some irregularities in the 90's.  At certain stages we can see cache size growth stall and even decrease in some iterations for an individual vendor.  The Pentium to the Pentium Pro for instance both had 16 KB L1 caches.  The Pro however was the first processor to have an on-die L2.  From 1992 when the SuperSPARC came out with 36 KB of L1 to 1995 the UltraSPARC decreased to a 32 KB L1.  In this instance though the L2 size capacity increased.  So while sometimes an individual cache size may remain the same or even decrease this is usually accompanied by another change.  As can be deduced from the table however, the typical L1 cache size per core has leveled out at 64 KB around 1999.
<br /><br />
''' L1, L2, L3 Size by Year '''
{|border=1
| '''Processor'''
| '''System Type'''
| '''Year'''
| '''L1 size'''
| '''L2 size'''
| '''L3 size'''
|-
| IBM 360/85
| Mainframe
| 1968
| 16 to 32 KB
| —
| —
|-
| PDP-11/70
| Minicomputer
| 1975
| 1 KB
| —
| —
|-
| VAX 11/780
| Minicomputer
| 1978
| 16 KB
| —
| —
|-
| IBM 3033
| Mainframe
| 1978
| 64 KB
| —
| —
|-
| IBM 3090
| Mainframe
| 1985
| 128 to 256 KB
| —
| —
|-
| Intel 80486
| PC
| 1989
| 8 KB
| —
| —
|-
| SuperSPARC
| PC
| 1992
| 16 KB/20 KB
| 0 to 2 MB
| —
|-
| Pentium
| PC
| 1993
| 8 KB/8 KB
| 256 to 512 KB
| —
|-
| PowerPC 601
| PC
| 1993
| 32 KB
| —
| —
|-
| UltraSPARC
| PC
| 1995
| 16 KB/16 KB
| 512 KB to 4 MB
| —
|-
| Pentium Pro
| PC
| 1995
| 8 KB/8 KB
| 256 KB - 1 MB
| —
|-
| PowerPC
| 620 PC
| 1996
| 32 KB/32 KB
| —
| —
|-
| PowerPC G4
| PC/server
| 1999
| 32 KB/32 KB
| 256 KB to 1 MB
| 2 MB
|-
| IBM S/390 G4
| Mainframe
| 1997
| 32 KB
| 256 KB
| 2 MB
|-
| IBM S/390 G6
| Mainframe
| 1999
| 256 KB
| 8 MB
| —
|-
| Pentium 4
| PC/server
| 2000
| 8 KB/8 KB
| 256 KB
| —
|-
| IBM SP
| High-end server
| 2000
| 64 KB/32 KB
| 8 MB
| —
|-
| CRAY MTAb
| Supercomputer
| 2000
| 8 KB
| 2 MB
| —
|-
| UltraSPARCIII
| PC
| 2001
| 32 KB/64 KB
| 2 to 8 MB
| —
|-
| Itanium
| PC/server
| 2001
| 16 KB/16 KB
| 96 KB
| 4 MB
|-
| SGI Origin 2001
| High-end server
| 2001
| 32 KB/32 KB
| 4 MB
| —
|-
| Itanium 2
| PC/server
| 2002
| 32 KB
| 256 KB
| 6 MB
|-
| IBM POWER5
| High-end server
| 2003
| 64 KB
| 1.9 MB
| 36 MB
|-
| CRAY XD-1
| Supercomputer
| 2004
| 64 KB/64 KB
| 1MB
| —
|-
| Nehalem (i5,7, Xenon)
| PC, Server
| 2008
| 32 KB/32 KB /core
| 256 KB per
| 4 MB to 12 MB total
|-
| Sandy Bridge (i3-7, Pent.)
| PC, Server
| 2011
| 32 KB/32 KB /core
| 256 KB per
| 1 MB to 20 MB total
|}


Pentium <br />
amd <br />
Mips <br />
sun-microsystems: sparc <br />
ibm: power pc <br />
DEC: alpha <br />


Penalty <100 when before 2000
<br />
after 2000 started to increase to get to main memory <br />
<br />
< 20 1 level fine <br />
<=100 2 level <br />
>=200 3 level <br /> <br />
miss rate reported, spec benchmarks
>=200 3 level


miss rate reported, spec benchmarks
==Main Memory Issues==
Finally main memory latency needs to be analyzed to see how it can affect the cache.  The cache is a necessary piece of hardware in the first place due to the severe disparity between processor speeds and main memory which is usually implemented with SDRAM.  Below are a few examples of main memory speed and the introduction year for these standards.  The cache provides a buffer between the registers and main memory to reduce the effects of the processor waiting on information from main memory.  There are two main restrictions on this however.  Firstly, cache is expensive.  Secondly, when cache size is increased, so is the access time[10].  To maximize cache usefulness we need the L1 to be as fast as the processor or at least fast enough to load into the pipeline between an instruction being decoded and executed.  So as has been noted many years ago, the growth rate of processor speed is much greater than the growth in DRAM speeds[8].  The difference in speeds are speculated to grow large enough that a "Memory Wall" will be reached if a solution is not found[8].  This states that once the divergence is large enough a system's speed will be solely determined by its memory speed.  As can be seen from the table below CAS Latency (CL) times have slightly improved over the years, along with the data bus speed. (CAS Latency refers to the time to access a word in a given column in a row that is already open.  Main memory can be viewed as a 2D array where you access the row, then column to fetch a word.)  DDR3 bus speed is actually close to clock speed for today's processors.  Latency can still be affected by row lookups however because if a row is not already open then it must be opened and this is usually the most expensive step in terms of time.  As to the memory wall however, DRAM cannot be the sole culprit for processor speed growth decreasing.  As has been shown through the evolution of standard processor design, adding more levels of increasingly larger cache can help negate the effects of a growing memory latency.  Certain techniques can also be employed to combat the memory wall such as out-of-order (OOO) execution and speculative precomputation (SP) [11].  Physical cooling limits of current technology also limit processor speeds.  All the hardware issues stated however can be explained as showing lack of progress due to lack of expenditure.  Since the majority of funding for computers today derives from home-grade consumers, a technology cannot be invested in if it cannot be shown to have a strong chance of recovering its investment.  Currently the level of capital needed keeps getting higher and the improvement of each generation is getting smaller.  So in order to make the next generation fast enough it may make said processors too expensive to be mass marketed.  The trend of consumer computing towards mobile makes speedups less important than mobility too, further sidelining the memory wall.


<br />
SDRAM:  <1998 <br />
DDR:    2000 <br />
DDR2:  2003 <br />
DDR3:  2007 <br />




'''L1 & L2 cache'''
{|border=1


| System
{| class="wikitable"
| Year
|+Memory timing examples (CAS latency only)
| L1 Size (cache)
! Generation
| L1 Speed (cc)
! Type
| L1 Associativity
! Data rate
| L2 Size
! Bit time
| L2 Speed
! Command rate
| L2 Associativity
! Cycle time
| L2 On-Chip?
! CL
| L3 Size
! First word
| L3 Speed
! Fourth word
| L3 Associativity
! Eighth word
| L3 On-Chip?
|-
| Main Mem Penalty (cc)
| rowspan="2" | SDRAM
| Notes:
| PC100
|-
|align=right| 100 MT/s
| IBM 360/85
|  10 ns
| 1968
|align=right| 100&nbsp;MHz
| 16 - 32 KB
|  10 ns
| ?
| 2
| Sector
| 20 ns
| None
| 50 ns
| N/A
| 90 ns
| N/A
|-
| N/A
| PC133
| None
|align=right| 133 MT/s
| N/A
|  7.5 ns
| N/A
|align=right| 133&nbsp;MHz
| N/A
|  7.5 ns
| ?
| 3
| First processor with a cache, clock speed 12.5MHz
| 22.5 ns
|-
| 45 ns
| IBM 486
| 75 ns
| 1989
|-
| 8 kb
| rowspan="4" | DDR SDRAM
| ?
| DDR-333
| L1 Associativity
|align=right| 333 MT/s
| 256 kb
|  3 ns
|  
|align=right| 166&nbsp;MHz
|  
|  6 ns
| no
| 2.5
|  
| 15 ns
|  
| 24 ns
|  
| 36 ns
|
|-
|  
|rowspan=3| DDR-400
|  
|rowspan=3 align=right| 400 MT/s
|-
|rowspan=3|  2.5 ns
| Intel 80486
|rowspan=3 align=right| 200&nbsp;MHz
| 1989
|rowspan=3|  5 ns
| 8 KB
| 3
| ?
| 15 ns
| 4-way associative
| 22.5 ns
| None
| 32.5 ns
| N/A
|-
| N/A
| 2.5
| N/A
| 12.5 ns
| None
| 20 ns
| N/A
| 30 ns
| N/A
|-
| N/A
| 2
| ?
| 10 ns
| First processor with a cache, clock speed 12.5MHz
| 17.5 ns
|
| 27.5 ns
|-
| rowspan="11" | DDR2 SDRAM
|rowspan=2| DDR2-667
|rowspan=2 align=right| 667 MT/s
|rowspan=2  |1.5 ns
|rowspan=2 align=right| 333&nbsp;MHz
|rowspan=2|  3 ns
| 5
| 15 ns
| 19.5 ns
| 25.5 ns
|-
|4
| 12 ns
| 16.5 ns
| 22.5 ns
|-
|rowspan=4| DDR2-800
|rowspan=4 align=right| 800 MT/s
|rowspan=4|  1.25 ns
|rowspan=4 align=right| 400&nbsp;MHz
|rowspan=4|  2.5 ns
| 6
| 15 ns
| 18.75 ns
| 23.75 ns
|-
| 5
| 12.5 ns
| 16.25 ns
| 21.25 ns
|-
| 4.5
| 11.25 ns
| 15 ns
| 20 ns
|-
| 4
| 10 ns
| 13.75 ns
| 18.75 ns
|-
|rowspan=5| DDR2-1066
|rowspan=5 align=right| 1066 MT/s
|rowspan=5|  0.95 ns
|rowspan=5 align=right| 533&nbsp;MHz
|rowspan=5|  1.9 ns
| 7
| 13.13 ns
| 15.94 ns
| 19.69 ns
|-
| 6
| 11.25 ns
| 14.06 ns
| 17.81 ns
|-
| 5
| 9.38 ns
| 12.19 ns
| 15.94 ns
|-
| 4.5
| 8.44 ns
| 11.25 ns
| 15 ns
|-
| 4
| 7.5 ns
| 10.31 ns
| 14.06 ns
|-
| rowspan="12" | DDR3 SDRAM
| DDR3-1066
|align=right| 1066 MT/s
|  0.9375 ns
|align=right| 533&nbsp;MHz
|  1.875 ns
| 7
| 13.13 ns
| 15.95 ns
| 19.7 ns
|-
|rowspan=2| DDR3-1333
|rowspan=2 align=right| 1333 MT/s
|rowspan=2|  0.75 ns
|rowspan=2 align=right| 666&nbsp;MHz
|rowspan=2|  1.5 ns
| 9
| 13.5 ns
| 15.75 ns
| 18.75 ns
|-
| 6
|  9 ns
| 11.25 ns
| 14.25 ns
|-
| DDR3-1375
|align=right| 1375 MT/s
|  0.73 ns
|align=right| 687&nbsp;MHz
|  1.5 ns
| 5
|  7.27 ns
|  9.45 ns
| 12.36 ns
|-
|rowspan=4| DDR3-1600
|rowspan=4 align=right| 1600 MT/s
|rowspan=4|  0.625 ns
|rowspan=4 align=right| 800&nbsp;MHz
|rowspan=4|  1.25 ns
| 9
| 11.25 ns
| 13.125 ns
| 15.625 ns
|-
| 8
| 10 ns
| 11.875 ns
| 14.375 ns
|-
| 7
|  8.75 ns
| 10.625 ns
| 13.125 ns
|-
| 6
|  7.50 ns
| 9.375 ns
| 11.875 ns
|-
|rowspan=4| DDR3-2000
|rowspan=4 align=right| 2000 MT/s
|rowspan=4|  0.5 ns
|rowspan=4 align=right| 1000&nbsp;MHz
|rowspan=4|  1 ns
| 10
| 10 ns
| 11.5 ns
| 13.5 ns
|-
| 9
| 9 ns
| 10.5 ns
| 12.5 ns
|-
|-
| SuperSPARC
| 8
| ?/1992
|  8 ns
| (16+20) KB
| 9.5 ns
| ?
| 11.5 ns
| 4 & 5 way set
| 0 - 2 MB
| ?
| ?
| No
| N/A
|
|
|
| ?
| Used to render Toy Story, Core @ 40MHz,
|-
| Alpha 21064(DEC)
| 09/1992
| (8+8) KB
| 1
| Direct
| 128kb - 16mb
| 3-16 cc
| Direct
| No
| 8 cc
|
|-
| Alpha 21064(DEC)
| 09/1992
| (8+8) KB
| 1
| Direct
| 128kb - 16mb
| 3-16 cc
| Direct
| No
| 8 cc
|
|-
| UltraSPARC
| 06/1995
| (16+16) KB
| 1
| 2-Way & Direct
| 512KB - 4MB
| 1, pipe=3
| Direct
| No
| N/A
|
|
|
| 2-3
| 64-bit w/ Core@200MHz
|-
| Alpha 21164(DEC)
| 01/1995
| (8+8) KB
| 1
| Direct
| 96 Kb
| 2
| 3 way set
| Yes
| ?
|-
| K6-III
| 1999
| 32 kib
| 100 mhz
| 2 way
| 256 kib
| 100 mhz
| 4 way
| yes
| 1 mb
|
|
| no
|
|
|-
|-
| Pentium 4
| 7
| 10/2000
|  7 ns
| 8 KB (trace)
| 8.5 ns
| 2
| 10.5 ns
| 4 Way
| 256KB
| full speed
| 8 Way
| Yes
| None
|
|
|
| 4 (Mem controller)
|
|-
| UltraSPARC III
| ?/2001
| (32+64) KB
| 1
| 4 Way
| 2-8MB
| 2-3
| ?
| No
| 4 (Mem controller)
|
|-
|-
| Itanium 2
! Generation
| 2002
! Type
| 16 KB
! Data rate
| 1
! Bit time
| 4 -way
! Command rate
| 256KB
! Cycle time
| 5
! CL
| 8-way
! First word
| yes
! Fourth word
| 3 MB
! Eighth word
| 12
| 12 way (4 per MB)
| Yes
| ?
| Notes:
|}
|}


----
<br />
References
 
[http://download.intel.com/design/itanium2/manuals/25111003.pdf Itanium Specs(p.20)]<br />
==Conclusion==
[http://www.chips.5u.com/idxhst.html Ref1]<br />
 
[http://en.wikipedia.org/wiki/List_of_Intel_microprocessors Intel Processors]<br />
During the early days of the PC the size difference between main memory and cache size was nowhere near what it is today.  As low level cache sizes have tended to reach a maximum over the years, main memory keeps getting larger and larger.  We can loosely follow this by looking at operating system requirements for main memory over the years compared to an average PC processor cache for the time period.  Because operating systems are competitive the developers want to pack as much capability into them as possible, which they will usually do by writing an OS that requires at least the minimum average processor specifications at the time.  For instance Windows 1.0 required 256 KB of RAM to run[12].  Compare that to the 8 KB available in in the Intel 80486 in 1989, 4 years after Windows 1.0 was released.  This gives us a percentage of 3.1% (2^-5).  Windows 95 recommended 8 MB of RAM for an installation[13].  Compare this to the 16 KB available in the Pentium Pro, a ubiquitous processor back then.  It comes out to 0.2% (2^-9).  In 2001 XP came out with a RAM recommendation of 128 MB[14].  If we compare that to an Itanium 2 with 32 KB L1 which came out after XP in 2002 we get a percentage of 0.024% (2^-12).  Since then L1 caches have not changed much but we have RAMs on the order of gigabytes now.  This gap in sizes between main memory and low level cache can be seen as a reason for associativity increasing.  As the percentage of our cache size to main memory decreases, misses from direct mapping will increase dramatically.
[http://en.wikipedia.org/wiki/Intel_80486 First on-board L1]<br />
 
[http://faculty.washington.edu/lcrum/Archives/TCSS372AS07/Slides04_05.ppt Cache Trend Table‎]<br />
On the other hand as was pointed out earlier, associativity can be seen to have a slight trend decreasing in the mid-90's before increasing again.  As we noted in the previous paragraph L1 size compared to main memory was most likely the main cause for the increase in associativity.  However since this happened around the time cache size growth was seen to stall and even backpedal in some cases maybe there is a correlation.  It is possible that due to processor speeds increasing engineers could not develop a cache that was both larger and faster at the same pace.  So it had to either be larger or faster.  In order to keep up with these rapid speed increases at the time they may have had to sacrifice associativity since it can slow a cache down by searching.  Once these technological hurdles were overcome though size and associativity could increase again.
[http://www.eecs.berkeley.edu/Pubs/TechRpts/1999/CSD-99-1034.pdf Sector Caches]<br />
 
 
<br />
 
==References==
<references/>
<ol>
<li>[http://download.intel.com/design/itanium2/manuals/25111003.pdf Itanium Specs(p.20)] Intel Datasheet</li>
<li>[http://www.chips.5u.com/idxhst.html Cache Evolution] </li>
<li>[http://en.wikipedia.org/wiki/List_of_Intel_microprocessors Intel Processors] Wikipedia</li>
<li>[http://en.wikipedia.org/wiki/Intel_80486 First on-board L1] Wikipedia</li>
<li>[http://faculty.washington.edu/lcrum/Archives/TCSS372AS07/Slides04_05.ppt Cache Trend Table‎] </li>
<li>[http://www.eecs.berkeley.edu/Pubs/TechRpts/1999/CSD-99-1034.pdf Sector Caches] </li>
<li>[http://www.tomshardware.com/reviews/ram-speed-tests,1807-3.html DDR2/3 Speeds] </li>
<li>[http://www.eecs.ucf.edu/~lboloni/Teaching/EEL5708_2006/slides/wulf94.pdf Memory Wall] Wulf & McKee</li>
<li>[http://support.gateway.com/s/Servers/shared/pproprsr/pentpro.shtml Pentium Pro Specs] Gateway Datasheet</li>
<li>[http://www.cs.utexas.edu/users/cart/trips/publications/isca00.pdf Clock Rate vs. IPC] Argawal et al. </li>
<li>[http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=19C67C93D13D430FE9ECD17FD57D2142?doi=10.1.1.134.6195&rep=rep1&type=pdf Mem Latency-Tol. Methods] Wang et al. </li>
<li>[http://en.wikipedia.org/wiki/Windows_1.0 Win1.0] Wikipedia </li>
<li>[http://support.microsoft.com/kb/138349 Win95] Microsoft </li>
<li>[http://support.microsoft.com/kb/314865 WinXP] Microsoft </li>
</ol>
<br />

Latest revision as of 22:07, 6 October 2013

Trends in cache size and organization


Introduction

Cache size has grown over the years alongside the evolution of the microprocessor. Intuitively one would expect cache sizes to keep growing larger and larger following some law similar to Moore’s Law. In actuality however L1 cache sizes have all but maxed out for an individual processor. Observing the trend of cache growth it can be seen that some processor lines stopped growing from one iteration to the next and in some cases even decreased in size. To go along with this, cache associativity has varied over the years. While it is true that no cache organization is optimal for every situation certain organizations certainly perform better for most tasks on certain systems. This wiki will try to analyze data on cache size and associativities to gain some insight into the trends and reasoning behind vendor choices of cache size and organization over the years. Specifically it looks from the late 80’s / early 90’s to the early 2000’s.


Cache Associativity

This table shows cache associativities found in some mainstream processors from the late 80’s to the early 2000’s with one processor from 1968 just for reference. As can be seen from the data, the late 80’s early 90’s tended towards a set associative cache with around four lines. In the mid-90’s it tended towards lower associativity and direct mapping. Then in the late 90’s and early 2000’s it tended back towards higher associativities with larger set sizes again.


L1, L2, L3 Associativity

System Year L1 Associativity L2 Associativity L3 Associativity Notes:
IBM 360/85 1968 Sector N/A N/A First processor with a cache, clock speed 12.5MHz
Intel 80486 1989 4-way associative N/A N/A
SuperSPARC 1992 4 & 5 way set N/A N/A Used to render Toy Story, Core @ 40MHz
Alpha 21064(DEC) 1992 Direct Direct N/A
UltraSPARC 1995 2-Way & Direct Direct N/A 64-bit w/ Core@200MHz
Alpha 21164(DEC) 1995 Direct 3 way set N/A
Pentium Pro 1995 2 & 4 way ? N/A First on-die L2
K6-III 1999 2 way 4 way n/a
Pentium 4 10/2000 4 Way 8 Way N/A
UltraSPARC III 2001 4 Way N/A N/A
Itanium 2 2002 4 -way 8-way 12 way





Cache Size

In accordance with Moore's law as the transistors on a chip increase we would expect cache sizes to increase with each generation of processors. Main memory sizes have certainly kept increasing so we would expect to see a similar trend in caches. Looking at the table below we can certainly see an increase in L1 cache sizes all the way up to the 2000's. Analyzing the trend however we can see some irregularities in the 90's. At certain stages we can see cache size growth stall and even decrease in some iterations for an individual vendor. The Pentium to the Pentium Pro for instance both had 16 KB L1 caches. The Pro however was the first processor to have an on-die L2. From 1992 when the SuperSPARC came out with 36 KB of L1 to 1995 the UltraSPARC decreased to a 32 KB L1. In this instance though the L2 size capacity increased. So while sometimes an individual cache size may remain the same or even decrease this is usually accompanied by another change. As can be deduced from the table however, the typical L1 cache size per core has leveled out at 64 KB around 1999.

L1, L2, L3 Size by Year

Processor System Type Year L1 size L2 size L3 size
IBM 360/85 Mainframe 1968 16 to 32 KB
PDP-11/70 Minicomputer 1975 1 KB
VAX 11/780 Minicomputer 1978 16 KB
IBM 3033 Mainframe 1978 64 KB
IBM 3090 Mainframe 1985 128 to 256 KB
Intel 80486 PC 1989 8 KB
SuperSPARC PC 1992 16 KB/20 KB 0 to 2 MB
Pentium PC 1993 8 KB/8 KB 256 to 512 KB
PowerPC 601 PC 1993 32 KB
UltraSPARC PC 1995 16 KB/16 KB 512 KB to 4 MB
Pentium Pro PC 1995 8 KB/8 KB 256 KB - 1 MB
PowerPC 620 PC 1996 32 KB/32 KB
PowerPC G4 PC/server 1999 32 KB/32 KB 256 KB to 1 MB 2 MB
IBM S/390 G4 Mainframe 1997 32 KB 256 KB 2 MB
IBM S/390 G6 Mainframe 1999 256 KB 8 MB
Pentium 4 PC/server 2000 8 KB/8 KB 256 KB
IBM SP High-end server 2000 64 KB/32 KB 8 MB
CRAY MTAb Supercomputer 2000 8 KB 2 MB
UltraSPARCIII PC 2001 32 KB/64 KB 2 to 8 MB
Itanium PC/server 2001 16 KB/16 KB 96 KB 4 MB
SGI Origin 2001 High-end server 2001 32 KB/32 KB 4 MB
Itanium 2 PC/server 2002 32 KB 256 KB 6 MB
IBM POWER5 High-end server 2003 64 KB 1.9 MB 36 MB
CRAY XD-1 Supercomputer 2004 64 KB/64 KB 1MB
Nehalem (i5,7, Xenon) PC, Server 2008 32 KB/32 KB /core 256 KB per 4 MB to 12 MB total
Sandy Bridge (i3-7, Pent.) PC, Server 2011 32 KB/32 KB /core 256 KB per 1 MB to 20 MB total




Main Memory Issues

Finally main memory latency needs to be analyzed to see how it can affect the cache. The cache is a necessary piece of hardware in the first place due to the severe disparity between processor speeds and main memory which is usually implemented with SDRAM. Below are a few examples of main memory speed and the introduction year for these standards. The cache provides a buffer between the registers and main memory to reduce the effects of the processor waiting on information from main memory. There are two main restrictions on this however. Firstly, cache is expensive. Secondly, when cache size is increased, so is the access time[10]. To maximize cache usefulness we need the L1 to be as fast as the processor or at least fast enough to load into the pipeline between an instruction being decoded and executed. So as has been noted many years ago, the growth rate of processor speed is much greater than the growth in DRAM speeds[8]. The difference in speeds are speculated to grow large enough that a "Memory Wall" will be reached if a solution is not found[8]. This states that once the divergence is large enough a system's speed will be solely determined by its memory speed. As can be seen from the table below CAS Latency (CL) times have slightly improved over the years, along with the data bus speed. (CAS Latency refers to the time to access a word in a given column in a row that is already open. Main memory can be viewed as a 2D array where you access the row, then column to fetch a word.) DDR3 bus speed is actually close to clock speed for today's processors. Latency can still be affected by row lookups however because if a row is not already open then it must be opened and this is usually the most expensive step in terms of time. As to the memory wall however, DRAM cannot be the sole culprit for processor speed growth decreasing. As has been shown through the evolution of standard processor design, adding more levels of increasingly larger cache can help negate the effects of a growing memory latency. Certain techniques can also be employed to combat the memory wall such as out-of-order (OOO) execution and speculative precomputation (SP) [11]. Physical cooling limits of current technology also limit processor speeds. All the hardware issues stated however can be explained as showing lack of progress due to lack of expenditure. Since the majority of funding for computers today derives from home-grade consumers, a technology cannot be invested in if it cannot be shown to have a strong chance of recovering its investment. Currently the level of capital needed keeps getting higher and the improvement of each generation is getting smaller. So in order to make the next generation fast enough it may make said processors too expensive to be mass marketed. The trend of consumer computing towards mobile makes speedups less important than mobility too, further sidelining the memory wall.


SDRAM: <1998
DDR: 2000
DDR2: 2003
DDR3: 2007


Memory timing examples (CAS latency only)
Generation Type Data rate Bit time Command rate Cycle time CL First word Fourth word Eighth word
SDRAM PC100 100 MT/s  10 ns 100 MHz  10 ns 2 20 ns 50 ns 90 ns
PC133 133 MT/s  7.5 ns 133 MHz  7.5 ns 3 22.5 ns 45 ns 75 ns
DDR SDRAM DDR-333 333 MT/s  3 ns 166 MHz  6 ns 2.5 15 ns 24 ns 36 ns
DDR-400 400 MT/s  2.5 ns 200 MHz  5 ns 3 15 ns 22.5 ns 32.5 ns
2.5 12.5 ns 20 ns 30 ns
2 10 ns 17.5 ns 27.5 ns
DDR2 SDRAM DDR2-667 667 MT/s 1.5 ns 333 MHz  3 ns 5 15 ns 19.5 ns 25.5 ns
4 12 ns 16.5 ns 22.5 ns
DDR2-800 800 MT/s  1.25 ns 400 MHz  2.5 ns 6 15 ns 18.75 ns 23.75 ns
5 12.5 ns 16.25 ns 21.25 ns
4.5 11.25 ns 15 ns 20 ns
4 10 ns 13.75 ns 18.75 ns
DDR2-1066 1066 MT/s  0.95 ns 533 MHz  1.9 ns 7 13.13 ns 15.94 ns 19.69 ns
6 11.25 ns 14.06 ns 17.81 ns
5 9.38 ns 12.19 ns 15.94 ns
4.5 8.44 ns 11.25 ns 15 ns
4 7.5 ns 10.31 ns 14.06 ns
DDR3 SDRAM DDR3-1066 1066 MT/s  0.9375 ns 533 MHz  1.875 ns 7 13.13 ns 15.95 ns 19.7 ns
DDR3-1333 1333 MT/s  0.75 ns 666 MHz  1.5 ns 9 13.5 ns 15.75 ns 18.75 ns
6  9 ns 11.25 ns 14.25 ns
DDR3-1375 1375 MT/s  0.73 ns 687 MHz  1.5 ns 5  7.27 ns  9.45 ns 12.36 ns
DDR3-1600 1600 MT/s  0.625 ns 800 MHz  1.25 ns 9 11.25 ns 13.125 ns 15.625 ns
8 10 ns 11.875 ns 14.375 ns
7  8.75 ns 10.625 ns 13.125 ns
6  7.50 ns 9.375 ns 11.875 ns
DDR3-2000 2000 MT/s  0.5 ns 1000 MHz  1 ns 10 10 ns 11.5 ns 13.5 ns
9 9 ns 10.5 ns 12.5 ns
8  8 ns 9.5 ns 11.5 ns
7  7 ns 8.5 ns 10.5 ns
Generation Type Data rate Bit time Command rate Cycle time CL First word Fourth word Eighth word


Conclusion

During the early days of the PC the size difference between main memory and cache size was nowhere near what it is today. As low level cache sizes have tended to reach a maximum over the years, main memory keeps getting larger and larger. We can loosely follow this by looking at operating system requirements for main memory over the years compared to an average PC processor cache for the time period. Because operating systems are competitive the developers want to pack as much capability into them as possible, which they will usually do by writing an OS that requires at least the minimum average processor specifications at the time. For instance Windows 1.0 required 256 KB of RAM to run[12]. Compare that to the 8 KB available in in the Intel 80486 in 1989, 4 years after Windows 1.0 was released. This gives us a percentage of 3.1% (2^-5). Windows 95 recommended 8 MB of RAM for an installation[13]. Compare this to the 16 KB available in the Pentium Pro, a ubiquitous processor back then. It comes out to 0.2% (2^-9). In 2001 XP came out with a RAM recommendation of 128 MB[14]. If we compare that to an Itanium 2 with 32 KB L1 which came out after XP in 2002 we get a percentage of 0.024% (2^-12). Since then L1 caches have not changed much but we have RAMs on the order of gigabytes now. This gap in sizes between main memory and low level cache can be seen as a reason for associativity increasing. As the percentage of our cache size to main memory decreases, misses from direct mapping will increase dramatically.

On the other hand as was pointed out earlier, associativity can be seen to have a slight trend decreasing in the mid-90's before increasing again. As we noted in the previous paragraph L1 size compared to main memory was most likely the main cause for the increase in associativity. However since this happened around the time cache size growth was seen to stall and even backpedal in some cases maybe there is a correlation. It is possible that due to processor speeds increasing engineers could not develop a cache that was both larger and faster at the same pace. So it had to either be larger or faster. In order to keep up with these rapid speed increases at the time they may have had to sacrifice associativity since it can slow a cache down by searching. Once these technological hurdles were overcome though size and associativity could increase again.



References

<references/>

  1. Itanium Specs(p.20) Intel Datasheet
  2. Cache Evolution
  3. Intel Processors Wikipedia
  4. First on-board L1 Wikipedia
  5. Cache Trend Table‎
  6. Sector Caches
  7. DDR2/3 Speeds
  8. Memory Wall Wulf & McKee
  9. Pentium Pro Specs Gateway Datasheet
  10. Clock Rate vs. IPC Argawal et al.
  11. Mem Latency-Tol. Methods Wang et al.
  12. Win1.0 Wikipedia
  13. Win95 Microsoft
  14. WinXP Microsoft