|
|
L2 vs L3 cache
|
|
|
|
Professional Poster
Join Date: Mar 2002
Location: Boston
Status:
Offline
|
|
I currently have a g4 450 mhz cube with 2mb of L2 cache. I've been toying with the purchasing an imac or a tower. Now in looking at these machines I noticed that: First the L2 cache has dropped from 2mb to 256k, quite a decrease and for the towers an L3 cache has been added.
Questions:
1. Why the decrease in L2
2. What is L3
3. Did this increase performance of the G4s. My gut reaction is that they are slower due to the smaller l2 cache.
4. I guess how does it stack up aganst my "pokey" 450mhz g4 with 2mb of L2.
Thanks
Mike
|
~Mike
|
|
|
|
|
|
|
|
Professional Poster
Join Date: Mar 2000
Location: Ottawa, Ontario, Canada
Status:
Offline
|
|
Originally posted by Maflynn:
<STRONG>I currently have a g4 450 mhz cube with 2mb of L2 cache. I've been toying with the purchasing an imac or a tower. Now in looking at these machines I noticed that: First the L2 cache has dropped from 2mb to 256k, quite a decrease and for the towers an L3 cache has been added.
Questions:
1. Why the decrease in L2
2. What is L3
3. Did this increase performance of the G4s. My gut reaction is that they are slower due to the smaller l2 cache.
4. I guess how does it stack up aganst my "pokey" 450mhz g4 with 2mb of L2.
Thanks
Mike</STRONG>
I'm no expert but I can give you a few insites.
First the new L2 caches are on the actual CPU chip and run at the speed of the CPU. The trade-off for putting it on the chip and running faster, is that the cache has to be smaller, currently. The L2 cache on your G4 450, is I believe, off the chip but on the CPU daughtercard and is a backside cache that probably runs at half or quarter speed of the CPU.
The L3 caches on the new G4s are probably about equivalent to your old G4 450s L2 cache except that the new L3 caches are faster. The L3 cache is a cache on the CPU daughtercard.
So, the L2 cache on the CPU is a definite speed improvement even if it is smaller. How much does the L3 cache help, well I'll leave that to someone who's more of an expert. I believe there was a thread in this very forum that pointed to a benchmark of one of the towers with and without a L3 cache. Can't remember which is faster and by how much.
|
Mac Pro Dual 3.0 Dual-Core
MacBook Pro
|
|
|
|
|
|
|
|
Junior Member
Join Date: Feb 2002
Status:
Offline
|
|
Hello.
I have a cube as well.
First, about cache memory: The cache on your cube runs at 1/2 the clock speed of your cpu. It's L2 memory because it's the only cache memory it has. "L1" would be the processor registers, by the way. The newer power macs have a small amount of cache on chip (since it's first cache, it's called L2 memory) that runs at full cpu speed. Some models also have an additional external cache, now called "level 3" since it follows the on-chip L2, that runs at 1/4 the cpu speed - in other words, about the same speed as your cache memory. Having a small amount of very fast cache memory is good. Being backed up by a larger amount of slower cache is also good. The new models with on-chip cache are maybe 10-15% faster than if they had no on-chip cache. Also, those with off-chip cache are maybe 10-15% faster than they would be if they had no external cache memory. These are not exact numbers by any means.
Now to the rest of your question:
Do not replace your cube with an Imac if all you want is a faster computer!!! Repeat that 3 times or until you believe it, whichever takes longer. Your cube is not slower than an Imac. Or at least, it is not slower by any amount that you will feel. It has the same 100 MHz bus and a smaller but faster cache. Tests have shown that the new Imac (the faster one, even) is barely faster than a cube. Why? Well, the newer processors are less efficient, clock for clock, than the older ones. Of course, if you want an Imac for some other reason, then go right ahead.
In order to get anything close to a "wow" out of a speed increase, you really need at least a 2x faster system.
I got a pretty good wow when I went from a nearly 6 year old 300Mhz G3 upgraded 7500 to my cube, for example. But I had to use the old one again briefly before I really appreciated the improvement.
Realistically, the next step up that will give you a good sense of improved speed is the 933. It has a faster bus (133 Mhz), a faster processor with both levels of cache memory. It should come with a 7200 rpm drive vs the 5400 rpm drive that came with your cube, and that will help too. And it will have a faster video card as well - much faster if you have the usual Rage 128 pro. All the little things add up. Over-all, it's easily 50% faster than your cube in real-life stuff. Maybe even more than that, but not twice as fast. So it will feel faster, especially for X, but won't give you a big "wow!" feeling. Of course, there are down-sides. Money spent, a less elegant machine, and a much noisier one. Your choice.
Personally, I'm keeping my cube.
-Ph.D.
|
|
|
|
|
|
|
|
|
Senior User
Join Date: Apr 2001
Location: Vancouver,BC,Canada
Status:
Offline
|
|
I wish I had a Cube, and as much as I love them, I think they are pokey in comparison to a new 800MHz tower. The new chips are not less efficient IMHO, they are a step up from previous processors. The way L2 and L3 cache handle instructions greatly improves the efficiency of handling data calculations. ( I think...)When the processor is looking for the next instruction set in the pipeline it looks first in the L1, then the L2, then the L3, and finally goes to much slower RAM. I am no expert on chip architecture by a long shot, and learned most of what I know poking around on the web, and right here in the forums. And while a chip with only on-chip L2 cache running at proc. speed will not be nearly fast as a chip running the same set-up AND sporting the L3cache, an 800MHz G4 7450(?) with only L2 cache is WAY more punchy than any Cube, is it twice as fast? No, but I would say it's close to 60 percent "faster" (real world)
IMHO (And I LOVE the Cube.) Get a new tower!!!
|
Cheers,
raferx
|
|
|
|
|
|
|
|
Mac Elite
Join Date: May 2001
Location: ~/
Status:
Offline
|
|
The processor's cache is very high speed memory where the processor stores both data and instructions it is about to use or the results from processed instructions. I don't know of a processor ever made with more than three cache levels so I'll just concetrate on three.
For a while now the speed of the processor has far outpaced the speed of the system's memory. This means that for every clock tick of the memory there are several for the processor. Without a cache a processor would have to wait until the memory cycle completed in order to write and read data and instructions from it. This is obviously inefficient because the processor scaling any faster than the speed of the memory is next to useless. The speed increase will be hampered by the latency between the processor's periods it can exchange data with main memory. Enter stage left the cache, the cache is a smaller faster bit of memory that can keep pace with the processor itself. Level 1 cache (L1) is the fastest and smallest cache memory unit. In most modern processors the L1 cache operates at the same speed as the processor. The size of the L1 is usually measured in a handful of "mere" kilobytes. The more memory space you add the more circuitry needs to be added which not only increases the power consumption but also increases the latency. Thus the L1 is kept pretty small.
As processors got even faster they yet again outpaced the main memory even with their L1 cache that could store a batch of instructions or data to be worked on. Engineers approached the problem similarly to the original cache problem. The second cache level was added which could store even more data than the L1, also since this data was only meant to be an interim storage area between the processor and the main memory, it could be slower than the processor itself and thus could be made bigger than the L1. Even more space to store data and instructions while waiting for the main memory to catch up. Then yet again processors speeded up yet main memory stayed fairly slow. L3 caches were implimented which could be even slower (and bigger) than L2 caches but faster than main memory.
These concepts have been around for a while and it is up to chip designers to play around with the concepts to increase performance without making the processor too expensive to produce. In the case of the G4, the size of the L2 cache was reduced but it was made faster (up to the core clock speed) and in the 7450 circuitry was added to allow addressing to an L3 cache off the core of the processor. The smaller faster L2 cache meant higher instruction and data throughput and it could use the L3 cache to store stuff that couldn't fit. The extra throughput in theory keeps the processor busy a larger percent of the time because it doesn't need to wait for the main memory to cycle back to get more data or other instructions. For example a small graphic processing algorithm can be kept in the L2 (or even the L2 if it were small enough) and just the data can be grabbed from the various cache levels.
The smaller faster L2 cache combined with a large L3 cache is a good thing, however the L2 cache alone isn't going to change much as the throuput is roughly the same as it was when it was larger and slower. With the L3 however the smaller L2 becomes formidible.
|
2GHz 15" MacBook Pro, 120GB 5400rpm HD, 2GB RAM
|
|
|
|
|
|
|
|
Professional Poster
Join Date: Mar 2002
Location: Boston
Status:
Offline
|
|
Thank you all for posting, I have been leaning towards the 933mhz, the dp gig would be great if I could afford it. The iMac is cool but as some of you have mentioned the speed difference isnt that great at least $1800 worth. As I'm trying to make an educated decision I needed to know what the differences are.
Thanks
Mike
|
~Mike
|
|
|
|
|
|
|
|
Fresh-Faced Recruit
Join Date: Jan 2002
Status:
Offline
|
|
Originally posted by Maflynn:
<STRONG>I currently have a g4 450 mhz cube with 2mb of L2 cache. I've been toying with the purchasing an imac or a tower. Now in looking at these machines I noticed that: First the L2 cache has dropped from 2mb to 256k, quite a decrease and for the towers an L3 cache has been added.
Questions:
1. Why the decrease in L2
2. What is L3
3. Did this increase performance of the G4s. My gut reaction is that they are slower due to the smaller l2 cache.
4. I guess how does it stack up aganst my "pokey" 450mhz g4 with 2mb of L2.
Thanks
Mike</STRONG>
Official Response from Apple Computer, Hardware/CPU/motherboard/ASIC/ and BENCHMARKING TEAMS THERE! ( No, *I* do NOT work at Apple....but I work with people who do, and this is my area of expertise ;-)
1st you have a G4/450MHz with a 64k Level 1 cache <32k instructions+ 32k Data> and a L 2 1MB <Not 2!> Backside cache ( and the chip is the MPC 7400 -see Motorola web sight for more info!)
Your question is why is the New iMac using different cache approaches? Well, apparently, *SOME* people were not happy staying at 500MHz forever....Motorola improved the G4 in MPC 7410 < lower power & Heat> but still no speed increase.
The first step was making a longer pipeline, from 4 to 7 stages. Alone, this would cause performance issues, a 43% speed hit - if the prefetch (Guess which IS 95% accurate, still is 5% wrong) would need 7 cycles to clear the data, as opposed to the smaller, more efficient 4 stage pipeline in your Cube, and all early 400 - 500MHz G4's.
So, Motorola had to make this much more efficient per clock cycle. The old G4 did 3 instructions per clock cycle, The New ones ( MPC 7450/7451 which topped out at 867MHz Mac's) do 4 instructions per cycle ( minimum, excluding altivec ) Several other tweaks were added, dividing Altivec into 4 Units, adding 3 simple integer units to the Complex Integer unit, adding Floating Point, Vector, Integer abilities to the later revision of the G4. In Altivec, the newer chips can do 16 instructions at a time - NOT counting Integer & FP Units...but the longer pipeline *could* in real world use....cache misses, cause serious speed hits, often making a 733MH G4 (when it was our fastest Mac) slower than a 500MHz MPC 7400!
To compensate, the L1 64k was kept, then an ON CHIP L2 Cache ( If there is a formula it is On Chip cache @ Full CPU = double backside cache @ 1/2 CPU Speed < Faster CPU's make it difficult to find RAM for a cache which is even 1/2 as fast!> Motorola, also added a BIG 2MB L3 cache which on a 1 GHz is 250MHz DDR = which means it runs at 500MHz ( hence Double Data Rate Ram moniker ) This allowed 320k on chip Total cache at 100% clock speed, plus 2 MB of frequently reusable data to arrive MUCH faster to the CPU.
Finally, the Newer MPC 7450 dropped 60x bus support ( 66-100MHz bus speeds ) hence forcing system bus speeds to 133MHz minimum, and faster PC-133 RAM.
( Perhaps you are beginning to see how much work CPU design & implementation is....now?)
Finally, the same CPU which could run no higher than 867MHz, was treated with SOI ( Silicon-on-Insulator ) which adds about 30% speed, without having to redesign at a smaller size, while lowering power consumption! So, we hit 1GHz - and you may have seen the guy who clocked his to a dual 1200MHz machine, and it still runs FINE!
OK, so the PRO Desktop, has the same chip as a 7450 + SOI + doubled addressing from 8 to 16 bit, and includes the 2MB Backside cache. Titanium PB G4 550 & 667, The NEW iMac (so far...hint hint) are the only units using the MPC 7445. Which is exactly the same design, lower power consumption, and has no L3 Cache. ( So, a good indication of JUST the CPU < Laptops use slow 4200rpm HD vs. 5400-7200rpm drives in Desktop Mac's, Portables use slower-low power graphics chips, etc. etc. So check a 667 CPU benchmark TiBook and add 7% - 20% ( 700MHz vs 800MHz iMac versions ) you'll see your cube wayyy back there...hehehe.
Still, the lack of a solid 1MB L2, or a 2MB L3 does limit performance for a G4 of that clock speed ( depending on what you do of course ). Yet, the MAIN issue in a G4 is Altivec/ Velocity Engine. Which *Was* only for PhotoShop, Video, Audio PRO's not long ago....
OS X 10.2 is preparing a TON of G4 optimiations ( it's amazing seeing a FP intensive task - coded to Velocity Engine- and seeing it move 15x-30x faster on the same machine!!! ) Also, 10.2 offloads the GUI to Graphics cards...which I believe your Rage 128 is supported....compared to the 54 Billion instructions in a G*Force*3/G*ForceMX4 or the 1.23 Trillion in the Geforce 4 Ti - well, it makes a BIG difference!
I think you understand the "Why's" of your questions now, better?
The amazing thing on the 1GHZ G4 is, well remember the 500MHz G4 could sustain 4 Gigaflops? ( Billion Floating Point operations per/sec) The DP G4/500MHz sustained 7.3 Gigaflops. Now despite a longer pipeline, the DP 1GHz does 15 Gigaflops ( Not almost twice as fast, but MORE than twice as fast - P3 had a 14 stage pipeline, the P4 has 20 stages- The 1GHz P3 is faster than 1.5GHz P4!!! Motorola has an awesome design team!)
Is the New iMac about twice as fast as your 450MHz Cube? YES! CPU, bus speed, RAM speed, HD speed, Graphics card - more than twice as fast in most areas!
Any Cube Lover, MUST appreciate seeing the same "Miniaturization" techniques *Learned on the CUBE ( even convenction cooled like the Cube! But, just in case a large, VERY slow, VERY silent fan kicks on when/if it get's to hot - OFF during sleep of course.
As a Cube owner, it must be a GOD SEND to have this concept WITH a 15" LCD start at $1299 - and based on orders, it will be around a while!
I just wanted to corect the technical errors, and give an actual answer ( Hope it wasn't too technical!)
Hope it helped,
Sincerely,
Armas
|
BUY AAPL STOCK! It will be going WAYYY UP!
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Forum Rules
|
|
|
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts
|
HTML code is Off
|
|
|
|
|
|
|
|
|
|
|
|