L2 Cache 2mb vs 4mb
- Thread starter wahoyaho
- Start date
1M Super Pi time may drop by 1 second. You'll never know the change under normal use. That said, if you do heavy duty number crunching like video editing, then the large cache may improve speed by a few %. Still too small of a gain for the additional price premium. I personally would avoid the E2xx0 (1MB) unless you can't afford the extra $20 for the E4xx0.
For many kinds of applications, there is a VAST benefit to having as much cache as possible.
There will be large benefits to having more cache any time many repeated calculations are needed over a
small calculation program, or many repeated calculations over a localized set of data.
If you're doing things involving converting/editing of diigtal photos, digital video,
engineering / scientific computing, et. al. then the larger cache is the right choice.
For things like web browsing, email, video games, you'll notice less of a benefit of
large cache, but it's still helpful. Also having a larger cache helps take the most
advantage of slower RAM speeds or having less system RAM, so you could economize
somewhat on the speed of your RAM if you have lots of cache.
I always say "you can overclock any processor, but you can never over-cache one!"
so I'd always buy the one with more cache even if it was a bit slower for the same cost
than one that's 15% faster with less cache.
There will be large benefits to having more cache any time many repeated calculations are needed over a
small calculation program, or many repeated calculations over a localized set of data.
If you're doing things involving converting/editing of diigtal photos, digital video,
engineering / scientific computing, et. al. then the larger cache is the right choice.
For things like web browsing, email, video games, you'll notice less of a benefit of
large cache, but it's still helpful. Also having a larger cache helps take the most
advantage of slower RAM speeds or having less system RAM, so you could economize
somewhat on the speed of your RAM if you have lots of cache.
I always say "you can overclock any processor, but you can never over-cache one!"
so I'd always buy the one with more cache even if it was a bit slower for the same cost
than one that's 15% faster with less cache.
READ RAM 9993 MB/s
READ L2 30717 MB/s
READ L1 63085 MB/s
WRITE RAM 8998 MB/s
WRITE L2 24679 MB/s
WRITE L1 62936 MB/s
COPY RAM 9137 MB/s
COPY L2 33436 MB/s
COPY L1 125857 MB/s
Latency 48.9 ns
Latency L2 2.9 ns
Latency L1 0.8 ns
As you can see,
L2 cache is about 3 times faster than your RAM can be, and
L1 cache is about 6 times faster than your RAM.
L2 cache has 1/17th the latency of RAM, and
L1 cache has 1/61st the latency of RAM.
So you can see there will be a huge benefit to having more of
the L1 and L2 cache since if you access data again in RAM
when it cannot fit in the cache, you will do so at something like
1/10th the speed and that slows down everything.
READ L2 30717 MB/s
READ L1 63085 MB/s
WRITE RAM 8998 MB/s
WRITE L2 24679 MB/s
WRITE L1 62936 MB/s
COPY RAM 9137 MB/s
COPY L2 33436 MB/s
COPY L1 125857 MB/s
Latency 48.9 ns
Latency L2 2.9 ns
Latency L1 0.8 ns
As you can see,
L2 cache is about 3 times faster than your RAM can be, and
L1 cache is about 6 times faster than your RAM.
L2 cache has 1/17th the latency of RAM, and
L1 cache has 1/61st the latency of RAM.
So you can see there will be a huge benefit to having more of
the L1 and L2 cache since if you access data again in RAM
when it cannot fit in the cache, you will do so at something like
1/10th the speed and that slows down everything.
Vast benefit with more cache? E4300 (2MB) with 15% core speed advantage over the X6800 (4MB). Core speed is still KING.
In summary, going from 2MB to 4MB cache is equivalent to an extra 150MHz of core speed. I wouldn't call this a vast benefit. Enough said.
Disclaimer...I didn't run these tests.
http://www.anandtech.com/cpuch...howdoc.aspx?i=2903&p=3
In summary, going from 2MB to 4MB cache is equivalent to an extra 150MHz of core speed. I wouldn't call this a vast benefit. Enough said.
Disclaimer...I didn't run these tests.
http://www.anandtech.com/cpuch...howdoc.aspx?i=2903&p=3
That's quite an unwarranted genralization.
Sure, you're right *if* all you do with your PC is run sysmark 2004
Office Productivity, Internet Content Creation, et. al. benchmarks.
Let's face it, if you're running an *interactive* program like an Office/Web one,
or a program that can finish a single task in less than a few seconds, who cares,
the PC is spending more time waiting for YOU than anything else. The speed
differences in that case will be mostly due to how fast you type, how fast your
hard disk is, how fast your internet connection is, etc.
Now imagine you're running a program that actually does repetitive calculations
over a limited and well organized set of data, calculations that have to loop
millions of times over the data and code before they're through.
Now do the math, I posted actual benchmark numbers above.
L2 cache = 3x faster than your RAM even for very fast RAM.
So if you had no cache:
1 million loops * speed 1 = 1 million seconds.
With L2 cache:
1 million loops * speed 0.3 = 300,000 seconds, 3X faster because of L2 cache.
Go ahead, write the program, it's about 4 lines of code, time it for us.
Now try to buy a processor that's 3X faster than a modern mid-range processor.
You can't. They don't make 6 GHz processors.
Now try to buy a 2 GHz processor with 2X the cache, and voila, suddenly programs
like the ones I said had a VAST benefit from more cache are now running
REAL WORLD 3X to 6X faster if they locally access data that fits within the cache
you have a majority of the time.
If you just want to surf the web and do email, stick with your 500 Mhz Celeron,
it's plenty fast enough for that stuff.
If you want to do serious calculations and data analysis, get a fast CPU, but
get the most cache your money can buy.
That's also an unwarranted generalization...
What are you trying to prove by saying "I posted actual benchmark numbers above"? His numbers weren't from benchmarks that are just as synthetic and theoretical as yours when applying them to real-world performance? At least he can back his point up by going to the next page and looking at encoding times, etc.
Want more proof that core speed is still KING? We have E2160 @ 3.4GHz with 1MB cache vs. X6800 @ 2.93GHz with 4MB cache. The E2160 has 1/4 the cache of the X6800. All it take was a 16% bump in core speed for the E2160 to sail past the X6800!
Quote from the article:
"The table above shows the advantage of the overclocked Pentium E2160 very clearly. It loses to Core 2 Extreme X6800 with 4MB L2 cache only in a few applications. It means that you can squeeze the performance of Intel?s top dual-core processor from a sub-$100 CPU, no matter how unbelievable it sounds."
http://www.xbitlabs.com/articl.../pentium-e2160_14.html
Quote from the article:
"The table above shows the advantage of the overclocked Pentium E2160 very clearly. It loses to Core 2 Extreme X6800 with 4MB L2 cache only in a few applications. It means that you can squeeze the performance of Intel?s top dual-core processor from a sub-$100 CPU, no matter how unbelievable it sounds."
http://www.xbitlabs.com/articl.../pentium-e2160_14.html
I've heard that 1MB chips suffer a huge performance hit relative to the 2MB chips - as much as 15% or more in games. 2MB versus 4MB, on the other hand, is only around 1-2% in most apps, sometimes as high as 5%.
It seems that Intel chips generally have a bottleneck point relative to L2 cache beyond which there is a rather large dropoff - generally their Celerons were below this threshold, while while their P4/Core Duo/C2D chips are all above this threshold.
As long as you steer clear of the 1MB Pentium chips (which are basically the new "celerons"), you should be fine. I have a 3.4GHz e6400 (2MB) and with WMV 1080p encoding I seem to be clearly hard drive limited, so the 2MB of cache definitely isn't substantially holding back performance. whatever chip you get, just try and find a good stepping that overclocks well.
It seems that Intel chips generally have a bottleneck point relative to L2 cache beyond which there is a rather large dropoff - generally their Celerons were below this threshold, while while their P4/Core Duo/C2D chips are all above this threshold.
As long as you steer clear of the 1MB Pentium chips (which are basically the new "celerons"), you should be fine. I have a 3.4GHz e6400 (2MB) and with WMV 1080p encoding I seem to be clearly hard drive limited, so the 2MB of cache definitely isn't substantially holding back performance. whatever chip you get, just try and find a good stepping that overclocks well.
Perhaps, but one can always overcome this deficiency with core speed. All the numbers from reputable websites have confirmed this, even with many games. People need to understand that core speed is still KING. Don't fall into the marketing hype of larger cache.
Games and DVD encoding are repetitive tasks? Please, reconsider.
Sure they're more computationally repetitious in their CPU execution graphs and data access patterns than an email client or web browser
would be, but they're HARDLY good examples of compute bound or memory bound or cache bound situations.
In games they're inteactive and the vast majority of their execution performance dependency lies in the capabilities of
your special GPU processor. That's WHY many people spend from $200 to $600+ on high powered GPU cards -- to make GAMES
perform faster. Try running most any 3D high action highly graphic highly textured game at 1600x1200 resolution on JUST a CPU with
a basic GPU and see how far you get towards acceptable performance levels no matter WHAT CPU you have -- even the most
EXTREME EDITION, QUAD CORE, SINGLE CORE, OVERCLOCKED, WATER COOLED, whatever, will pathetically fail to achieve
what even a $150 graphics card can do.
Actually that's the POSTER CHILD example of proving my very point that the computational ARCHITECTURE is often the MAIN
factor in real world performance, not mere CPU MHz. It may surprise you to know that GPUs have very LITTLE cache, instead
they have massive numbers of independent processors each running at very high speeds coupled with lots of very high speed
memory. The reason for that is GPU rendering is anything BUT a computationally repetitive activity, it's called a
STREAMING activity where you calculate a given triangle's position / shading then you move on to the next one probably
never even processing the previous triangle AGAIN until the shader finishes which is a very long time and very large number
of triangles in the future. I said very clearly that cache benefits LOCALLY REPETITIVE processing of code/data, and
3D graphics in the traditional sense is not that.
The overall GAME performance is therefore GPU limited at any reasonably high resolution for most cutting edge games,
and what the CPU is left to do isn't at all the critical path or necessarily very local / repetitious.
Now as for DVD encoding, sure, it's computationally intensive, but again, it's PRIMARILY either a
SPECIAL PURPOSE HARDWARE limited activity, or a streaming not very repetitive (in data) action.
To compress a given line of video you need to know that line, the previous couple of lines, the next couple of lines, and
maybe the same lines of the previous and next frames of video. Once you have those data you can do your motion
estimation and spatial quantization and chromatic quantization et. al. You process those lines once and outside of
that small window of time and spatial context you NEVER TOUCH THEM AGAIN. Accessing data a few dozen times and
never touching it again isn't my (or a CPU's) definition of a super repetitive process. It's more of a streaming process.
At 60 fields per second of video that's 216,000 fields of video per hour and you're accessing any given
field maybe four times total since once you've read it, encoded it, you move on to the next totally unrelated
few fields. Sure a properly written software encoder can still make GOOD use of the CPU cache to deliver
substantial performance benefits of maybe 10x the performance compared to the same code that
either didn't use the cache right at all or which had to run on a CPU with WOEFULLY too little cache.
However the fraction of even POSSIBLY cacheable accesses to total ones is still small due to the small data access
repetition even if the coding algorithm probably is a lot more repetitive for the DCTs and quantizations etc.
Anyway that's all somewhat moot since anyone who's REALLY concerned about high video transcoding performance
probably has special purpose DSP encoder hardware inside their GPU or video chipset or digital video peripheral
that does about 90% if not 99% of the computing work involved in DCT / MPEG-2 / MPEG-4 / H.264 type encoding
and decoding processes which are specially designed to do just that task like ten to fifty times faster than any
general purpose CPU based software encoder can touch. Ever see ATI's AVIVO transcoder stuff?
The performance advantage of hardware MPEG-2 encoders/decoders for HDTV? I rest my case.
Saying that CPU cache only gives a moderate advantage to algorithms that are not even mostly
relevant to executing on the CPU or in a repetitive manner just points out how generally USEFUL
CPU cache is, giving decently noteworthy gains to programs that aren't even super great utilizers of it.
Now let's look at another REAL WORLD example that's a LOT more repetitive in certain configurations --
from the Folding @ Home forum. Observe the following gentleman's comment about a REAL WORLD
2X performance benefit on computations that take HOURS (much more than just a DVD encode!)
due to cache.
His 4MB Core2 can do the same type of work in half the time as his other machine that has
less cache.
As I've said before, you can (and many people should if they want the best performance)
OVERCLOCK *ANY* CPU QUITE A BIT, BUT YOU CANNOT OVERCACHE ONE AT ALL!
So if you happen to be running some program that's not hardware accellerated and needs
repetitive computations for CPU-hours per day / week, you CAN AND WILL get something
up to a 3X to 10X benefit (depending on the application) from having enough cache for
the program so that the 'hit ratio' is very high as opposed to having such little cache that
the hit ratio is well under 20%.
You just CANNOT buy at an attractive price OR OVERCLOCK a CPU 3X faster than a modern
mid-end one. So if you're doing very repetitive accesses for code / data that's taking
something like 80% of your total run time, it's the cache effectiveness that's MOSTLY
responsible for your performance, NOT the CPU MHz!
Look up the terms "compute bound", "memory bound", "cache bound" -- some things
need little cache and fast CPU/main memory (streaming programs), some things need super effective cache,
some things need a mix of both.
Assuming you overclock WHATEVER processor you buy to the max, and overclock whatever RAM
you have to the max, the KEY difference in whether you should even BOTHER to try to run certain
programs is if they fit in the cache well.
Know why you don't hear about a lot of programs that run miserably poorly on PCs without
large caches in common benchmarks? Because nobody even would TRY to run such a thing
on a PC UNLESS it had enough cache to perform well. Otherwise you would buy a special
workstation class or supercomputer class processor or beowulf cluster or whatever because
you'd just be wasting time and money to even try to run something that would be crippled
on a PC due to the small cache.
It is only recently that common PCs have even HAD 1MBy + cache amounts and it is no
coincidence that we now see a DRAMATIC increase in program POWER and CAPABILITY
for PCs because such programs are not even practical unless you have at least in
the 1MBy+ cache range, at least 2 CPU cores, and at least 2 GBy of RAM, none of which
has been possible with the majority of the consumer class of CPUs before the past year or two.
Sure they're more computationally repetitious in their CPU execution graphs and data access patterns than an email client or web browser
would be, but they're HARDLY good examples of compute bound or memory bound or cache bound situations.
In games they're inteactive and the vast majority of their execution performance dependency lies in the capabilities of
your special GPU processor. That's WHY many people spend from $200 to $600+ on high powered GPU cards -- to make GAMES
perform faster. Try running most any 3D high action highly graphic highly textured game at 1600x1200 resolution on JUST a CPU with
a basic GPU and see how far you get towards acceptable performance levels no matter WHAT CPU you have -- even the most
EXTREME EDITION, QUAD CORE, SINGLE CORE, OVERCLOCKED, WATER COOLED, whatever, will pathetically fail to achieve
what even a $150 graphics card can do.
Actually that's the POSTER CHILD example of proving my very point that the computational ARCHITECTURE is often the MAIN
factor in real world performance, not mere CPU MHz. It may surprise you to know that GPUs have very LITTLE cache, instead
they have massive numbers of independent processors each running at very high speeds coupled with lots of very high speed
memory. The reason for that is GPU rendering is anything BUT a computationally repetitive activity, it's called a
STREAMING activity where you calculate a given triangle's position / shading then you move on to the next one probably
never even processing the previous triangle AGAIN until the shader finishes which is a very long time and very large number
of triangles in the future. I said very clearly that cache benefits LOCALLY REPETITIVE processing of code/data, and
3D graphics in the traditional sense is not that.
The overall GAME performance is therefore GPU limited at any reasonably high resolution for most cutting edge games,
and what the CPU is left to do isn't at all the critical path or necessarily very local / repetitious.
Now as for DVD encoding, sure, it's computationally intensive, but again, it's PRIMARILY either a
SPECIAL PURPOSE HARDWARE limited activity, or a streaming not very repetitive (in data) action.
To compress a given line of video you need to know that line, the previous couple of lines, the next couple of lines, and
maybe the same lines of the previous and next frames of video. Once you have those data you can do your motion
estimation and spatial quantization and chromatic quantization et. al. You process those lines once and outside of
that small window of time and spatial context you NEVER TOUCH THEM AGAIN. Accessing data a few dozen times and
never touching it again isn't my (or a CPU's) definition of a super repetitive process. It's more of a streaming process.
At 60 fields per second of video that's 216,000 fields of video per hour and you're accessing any given
field maybe four times total since once you've read it, encoded it, you move on to the next totally unrelated
few fields. Sure a properly written software encoder can still make GOOD use of the CPU cache to deliver
substantial performance benefits of maybe 10x the performance compared to the same code that
either didn't use the cache right at all or which had to run on a CPU with WOEFULLY too little cache.
However the fraction of even POSSIBLY cacheable accesses to total ones is still small due to the small data access
repetition even if the coding algorithm probably is a lot more repetitive for the DCTs and quantizations etc.
Anyway that's all somewhat moot since anyone who's REALLY concerned about high video transcoding performance
probably has special purpose DSP encoder hardware inside their GPU or video chipset or digital video peripheral
that does about 90% if not 99% of the computing work involved in DCT / MPEG-2 / MPEG-4 / H.264 type encoding
and decoding processes which are specially designed to do just that task like ten to fifty times faster than any
general purpose CPU based software encoder can touch. Ever see ATI's AVIVO transcoder stuff?
The performance advantage of hardware MPEG-2 encoders/decoders for HDTV? I rest my case.
Saying that CPU cache only gives a moderate advantage to algorithms that are not even mostly
relevant to executing on the CPU or in a repetitive manner just points out how generally USEFUL
CPU cache is, giving decently noteworthy gains to programs that aren't even super great utilizers of it.
Now let's look at another REAL WORLD example that's a LOT more repetitive in certain configurations --
from the Folding @ Home forum. Observe the following gentleman's comment about a REAL WORLD
2X performance benefit on computations that take HOURS (much more than just a DVD encode!)
due to cache.
His 4MB Core2 can do the same type of work in half the time as his other machine that has
less cache.
As I've said before, you can (and many people should if they want the best performance)
OVERCLOCK *ANY* CPU QUITE A BIT, BUT YOU CANNOT OVERCACHE ONE AT ALL!
So if you happen to be running some program that's not hardware accellerated and needs
repetitive computations for CPU-hours per day / week, you CAN AND WILL get something
up to a 3X to 10X benefit (depending on the application) from having enough cache for
the program so that the 'hit ratio' is very high as opposed to having such little cache that
the hit ratio is well under 20%.
You just CANNOT buy at an attractive price OR OVERCLOCK a CPU 3X faster than a modern
mid-end one. So if you're doing very repetitive accesses for code / data that's taking
something like 80% of your total run time, it's the cache effectiveness that's MOSTLY
responsible for your performance, NOT the CPU MHz!
Look up the terms "compute bound", "memory bound", "cache bound" -- some things
need little cache and fast CPU/main memory (streaming programs), some things need super effective cache,
some things need a mix of both.
Assuming you overclock WHATEVER processor you buy to the max, and overclock whatever RAM
you have to the max, the KEY difference in whether you should even BOTHER to try to run certain
programs is if they fit in the cache well.
Know why you don't hear about a lot of programs that run miserably poorly on PCs without
large caches in common benchmarks? Because nobody even would TRY to run such a thing
on a PC UNLESS it had enough cache to perform well. Otherwise you would buy a special
workstation class or supercomputer class processor or beowulf cluster or whatever because
you'd just be wasting time and money to even try to run something that would be crippled
on a PC due to the small cache.
It is only recently that common PCs have even HAD 1MBy + cache amounts and it is no
coincidence that we now see a DRAMATIC increase in program POWER and CAPABILITY
for PCs because such programs are not even practical unless you have at least in
the 1MBy+ cache range, at least 2 CPU cores, and at least 2 GBy of RAM, none of which
has been possible with the majority of the consumer class of CPUs before the past year or two.
Sorry but most reputable review websites don't support your rambling logic. Have fun with your 4MB CPU.
That's utter nonsense. You speak of "many games" as if they're the
penultimate computational test of or usefulness of a computer.
How many games can you run at 1600x1200 x 60 FPS on your
uber-overclocked 4GHz water cooled 1 MB cache Pentium Extreme Edition or
whatever CPU using a Gf 5700 GPU?
Not many? None?
What happened to "you can always overcome this deficiency with core speed!"?
Now take your same overclocked 4GHz CPU with 1 MB cache and PC-6400 RAM
and do a few tens of millions of double precision floating point 16384x16384
matrix inversions or matrix multiplications on it and get back to me about
how that performs compared to a 2GHz CPU with 4MBy of cache.
Or run a cache intensive Folding @ Home work unit on it, same deal.
Many possible real world examples.
But it sounds like you just need an 8800GTX and E6300 and some good games
to be happy which is just fine, but don't make overly general statements
about how "clock rate is king". I'll show you just how fast your uber-clocked
under-cached CPU can sit there twiddling its thumbs idle uselessly just WAITING
for more data to come in from the memory 90% of the time totally useless.
MEMORY = 8 GBy / second max. performance.
CPU = 8 GBy / second max. I/O to the memory.
So that lets you do about A COUPLE of calculations for each item in memory
at max. POSSIBLE speeds before you CANNOT go faster due to the limits of
your FSB and Memory.
But if you do your code right on a dual core CPU you can execute 8 instructions
for each memory access, but where do those instructions get their data?
Not from memory since it's already running at max. speed!
Better be from the cache or *poof* suddenly you miss the opportunity to do
7/8th of your CPU's possible calculations becuase you're....
WAITING....
and
WAITING....
and
WAITING....
...for memory.
penultimate computational test of or usefulness of a computer.
How many games can you run at 1600x1200 x 60 FPS on your
uber-overclocked 4GHz water cooled 1 MB cache Pentium Extreme Edition or
whatever CPU using a Gf 5700 GPU?
Not many? None?
What happened to "you can always overcome this deficiency with core speed!"?
Now take your same overclocked 4GHz CPU with 1 MB cache and PC-6400 RAM
and do a few tens of millions of double precision floating point 16384x16384
matrix inversions or matrix multiplications on it and get back to me about
how that performs compared to a 2GHz CPU with 4MBy of cache.
Or run a cache intensive Folding @ Home work unit on it, same deal.
Many possible real world examples.
But it sounds like you just need an 8800GTX and E6300 and some good games
to be happy which is just fine, but don't make overly general statements
about how "clock rate is king". I'll show you just how fast your uber-clocked
under-cached CPU can sit there twiddling its thumbs idle uselessly just WAITING
for more data to come in from the memory 90% of the time totally useless.
MEMORY = 8 GBy / second max. performance.
CPU = 8 GBy / second max. I/O to the memory.
So that lets you do about A COUPLE of calculations for each item in memory
at max. POSSIBLE speeds before you CANNOT go faster due to the limits of
your FSB and Memory.
But if you do your code right on a dual core CPU you can execute 8 instructions
for each memory access, but where do those instructions get their data?
Not from memory since it's already running at max. speed!
Better be from the cache or *poof* suddenly you miss the opportunity to do
7/8th of your CPU's possible calculations becuase you're....
WAITING....
and
WAITING....
and
WAITING....
...for memory.
idiotekniQues
Platinum Member
- Jan 4, 2007
- 2,572
- 0
- 76
i read a lot of this when i was deciding e6400 vs e6600 back in january. every reputable review backed the small increase of up to 5% in some apps.
Thanks.
Have fun with your games -- I guess you spent too much time playing them
to learn to program or graduate from a computer architecture class
to teach you the error of your simplistic viewpoints on CPU function.
Not at all. I benchmarked the ACTUAL achievable performance benefit of cache L1 vs L2 vs RAM,
that does NOT depend on the application you run so it's not so much of a synthetic
benchmark as a fact of what your CPU and MEMORY CAN and CANNOT do.
It is like saying "I have an integer ALU so I can add numbers with this CPU", it's a functional
fact not a contrived synthetic theoretical benchmark of some arbitrary program.
And my point was therefore that it's a functional irrefutable FACT that
if you have the cache, your code and data memory accesses WILL
be 3X to 6X faster than POSSIBLE with RAM each time, every time,
all the time, no matter WHAT program you run.
If someone is saying that a 3X to 6X performance boosting factor
is irrelevant when it's obviously demonstrable that SOME applications
do use such capabilities heavily they're just wrong.
I never said that it'd be 3X faster for reading email, I said that
there is NO POSSIBLE WAY to get that 3X to 6X performance
benefit UNLESS you have enough cache for your program;
end of story, you have it or you don't.
If you don't have the cache you lose more ACTUAL performance than
clock rate can POSSIBLY make up for for a significant class of prorgams.
If those kinds of programs aren't what you care about then that's
fine, but I never said it'd make EVERY program 6X faster, just
SOME programs.
Have there been any benchmarks that compare heavy multitasking scenarios?
I'm interested in setting up a server to run virtual machines on. I need multiple VM's simultaneously, running apps, relational databases, web sites/services, domain controllers, SMTP servers, DNS, etc, basically setting up a virtualized environment mimicking the infrastructure of an entire business division.
I imagine that for something like this having 4MB of cache would give a noticable boost over 2 megs, since multiple apps and their respective datasets would easily surpass 2 or even 4 megs.
I'm interested in setting up a server to run virtual machines on. I need multiple VM's simultaneously, running apps, relational databases, web sites/services, domain controllers, SMTP servers, DNS, etc, basically setting up a virtualized environment mimicking the infrastructure of an entire business division.
I imagine that for something like this having 4MB of cache would give a noticable boost over 2 megs, since multiple apps and their respective datasets would easily surpass 2 or even 4 megs.
I'm sure there are such benchmarks, and I could help suggest places
to find them for you if desired.
However it doesn't seem like you really have much 'choice' if you're considering
the mainstream CPUs like the quad core Kentsfields either XEON 32xx series
or Q6xxx series. Don't they all have 2 MB L2 cache per core, 4 cores?
I understand that the upcoming 'Penryn' 45nm CPUs will have 3 MBy cache
per core, but I'm not aware of other CPU choices you have right now that'd
match that amount of cache. Though I do admit to not knowing much
about the XEON 5000 series or the high end server AMD Opterons et. al.
so maybe there's some cache choice.
Given the obvious advantages of having 4 cores / CPU at an attractive price
point, I'd think that'd be a primary consideration and the decent amount of
cache that comes along with them a bonus, though one could always want more.
I'm sure fast RAM and lots of it would also be a key benefit for your application.
If you're virtualizing that much stuff then even 8GBy / CPU looks a little anemic
unless you can split the load between a cluster of 2-4 quad-core CPU boxes
each with 8 GBy of memory which would have its fault-tolerance benefits
if you need to shift load from one to another while you're maintaining one box.
Yeah I just double-checked all the Intel Quad Core XEON
E5300 / X3200 CPUs have 8MBy L2 cache just like the Q6xxx ones, so
not much choice there.
However the XEON 7000 series has parts with up to 16 MBy L3, though those
are 2 cores / CPU, so they've got a ton of cache per core and very fast clock
rates.
http://compare.intel.com/pcc/d...milyid=5&culture=en-US
http://www.intel.com/products/...sor/xeon7000/index.htm
"Demo the benefits of larger cache ?"
http://www.intel.com/business/...n7100.htm?iid=xeon7000
http://www.intel.com/products/...sor/xeon7000/index.htm
So re: cache you've got a couple of choices, but your main choice
is what number of servers / blades you'll have, how much
memory, form factor, cost, redundancy, processor family / count, et. al.
Check VMWARE's site for white papers etc. I'm sure they have some
good benchmarks. Also obviously the Intel Xeon Server & Workstation
pages I linked, the AMD Opteron pages, and so on from SUN et. al.
I guess you spend too much time rambling on and on to learn to graduate from an English class. Semi-colons aren't substitutes for commas, nor vice-versa.
You need to understand people use computers for different things. This is Anandtech. People on Anandtech like to overclock and play games. CPU's at the same clock speed show negligible performance differences considering the price difference in the applications Anandtechers use most, even with different amounts of L2 cache. I don't argue that my mule can carry heavy loads across long distances at a horse racing event.
VirtualLarry
No Lifer
- Aug 25, 2001
- 56,570
- 10,202
- 126
People on Anandtech also like to participate in Distributed Computing. Some DC clients prefer larger caches to core speed.
That said, the performance difference that I've observed, going from 2MB L2 to 4MB L2 is mostly slight. But there are larger (negative) differences, going down to 1MB L2, or worse, 512KB L2. So get at least a 2MB L2 CPU. 4MB L2 is just gravy, although there are a few apps that like that much L2.
Thats pretty much my experience with cache as well, the gains are inconsequential *UNLESS* its a cripplingly small amount (aka celerons).
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 24K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
- Started by Tigerick
- Replies: 19K
-
-
-
Facebook
Twitter