what's the right amount of RAM for a road warrior?
Nov 9, 2007, 3:29 pm
#46
Join Date: Aug 2001
Location: A Southern locale that ain't the South.
Programs: Bah, HUMBUG!
Posts: 8,014
Please cite a source for this assertion. The Crucial article is very clear that accesses in dual-channel configuration are serial, not parallel. And the Tom's Hardware analysis cited in Wiki measures bandwidth of single- or dual-channel as almost identical.
(Emphasis added.)
(Emphasis added.)
I'm not sure what precisely you're arguing. I've advocated stuffing the laptop full of RAM as you can, be that 1, 2, 3, or 4 GB. This generally involves installing two identical modules which will both offer more RAM and permit dual-channel operation.
Your comment was that going to 2GB soDIMMs would affect battery life vs. 1GB sticks. Here is a Micron power budget calculator for DDR2. Now I can't pretend to understand all of this but I can plug in 1gb vs. 2 and see there's not much difference. Please correct me here where I'm wrong. Two modules will consume more power than one. Two 2gb modules will not consume significantly more power than two 1gb module.
The actual quote is that Dual Channel is insignificant for "most users". The memory bandwidth increase is still there. Not only that; Tom's Hardware hasn't been a reliable resource for several years now, I wouldn't be citing them. Also, you CAN run in dual channel with mixed modules. I've done it. All else equal, if you can have two channel operation, why wouldn't you want the bandwidth increase? Why would you run one 2gb module and one 1gb module? I never advocated merely running one module. Memory draws the LEAST power among HDD, LCD, CPU, chipset, etc.
Nov 9, 2007, 4:03 pm
#47
Suspended
Join Date: Jul 2007
Programs: AAdvantage, SkyMiles, USAir, Singapore, BA
Posts: 602
I respectfully disagree - in a striped disk array, data from all drives in the array arrive at the system's DMA bus simultaneously, whereas in a dual-channel memory, data from the two banks arrive asynchronously.
Nov 9, 2007, 4:25 pm
#48
Suspended
Join Date: Jul 2007
Programs: AAdvantage, SkyMiles, USAir, Singapore, BA
Posts: 602
I'm not sure what precisely you're arguing. I've advocated stuffing the laptop full of RAM as you can, be that 1, 2, 3, or 4 GB. This generally involves installing two identical modules which will both offer more RAM and permit dual-channel operation.
Your comment was that going to 2GB soDIMMs would affect battery life vs. 1GB sticks. Here is a Micron power budget calculator for DDR2. Now I can't pretend to understand all of this but I can plug in 1gb vs. 2 and see there's not much difference. Please correct me here where I'm wrong. Two modules will consume more power than one. Two 2gb modules will not consume significantly more power than two 1gb module.
The actual quote is that Dual Channel is insignificant for "most users". The memory bandwidth increase is still there. Not only that; Tom's Hardware hasn't been a reliable resource for several years now, I wouldn't be citing them. Also, you CAN run in dual channel with mixed modules. I've done it. All else equal, if you can have two channel operation, why wouldn't you want the bandwidth increase? Why would you run one 2gb module and one 1gb module? I never advocated merely running one module. Memory draws the LEAST power among HDD, LCD, CPU, chipset, etc.
Your comment was that going to 2GB soDIMMs would affect battery life vs. 1GB sticks. Here is a Micron power budget calculator for DDR2. Now I can't pretend to understand all of this but I can plug in 1gb vs. 2 and see there's not much difference. Please correct me here where I'm wrong. Two modules will consume more power than one. Two 2gb modules will not consume significantly more power than two 1gb module.
The actual quote is that Dual Channel is insignificant for "most users". The memory bandwidth increase is still there. Not only that; Tom's Hardware hasn't been a reliable resource for several years now, I wouldn't be citing them. Also, you CAN run in dual channel with mixed modules. I've done it. All else equal, if you can have two channel operation, why wouldn't you want the bandwidth increase? Why would you run one 2gb module and one 1gb module? I never advocated merely running one module. Memory draws the LEAST power among HDD, LCD, CPU, chipset, etc.
To reiterate: maxing out memory won't speed up every computer running every application. You really have to know what you're doing to design a memory strategy that makes sense from a price/performance/heat/battery point of view.
Nov 9, 2007, 11:05 pm
#49
Join Date: Nov 2000
Location: Upcountry Maui, HI
Posts: 13,308
Please cite a source for this assertion. The Crucial article is very clear that accesses in dual-channel configuration are serial, not parallel. And the Tom's Hardware analysis cited in Wiki measures bandwidth of single- or dual-channel as almost identical.
(Emphasis added.)
(Emphasis added.)
http://www.kingston.com/newtech/MKF_...whitepaper.pdf
http://en.wikipedia.org/wiki/Dual-channel_architecture
http://www.hardwaresecrets.com/article/133/1
Yes, it really does access both memory channels in parallel, effectively doubling the bandwidth to memory, but you will never see a doubling of performance by switching to dual channel mode. In fact, as you noted, most synthetic benchmarks show only a small improvement when using dual channel mode over single channel mode. There is, in fact, a real world performance boost when using dual channel mode, but the amount of performance boost you actually see is highly application dependent. Since all the "magic" to support dual-channel mode is already built into the chipset's memory controller, you don't lose anything by taking advantage of it.
BTW, Tom's Hardware just isn't what it used to be. Not sure what article you are citing there, or which crucial article you are citing.
Edit: I guess you are citing the Tom's hardware article linked from the wikipedia article on dual channel memory. They try to explain why there isn't as big of a performance boost today as when the dual channel memory architecture was introduced. Basically, they say it's the cache effect. Large 2MB and 4MB caches tend to hide long latency to/from memory, and that's true to some extent. Eventually, the cache will become full and cache lines will need to be displaced. I imagine that if you're running applications like large files in photo shop that are memory bound, dual channel memory will do a decent job at increasing your performance. If you're just accessing random bytes from all over the place, it probably won't do anything to increase your performance.
Still not sure what Crucial article you were trying to cite.
-David
Last edited by LIH Prem; Nov 9, 2007 at 11:29 pm
Nov 10, 2007, 1:58 am
#50
FlyerTalk Evangelist
Join Date: Jul 2000
Location: in the vicinity of SFO
Programs: AA 2MM (LT-PLT, PPro for this year)
Posts: 19,781
Sure. But it starts up from sleep even faster... and given that a modern system can stay in "sleep" mode for a day or so (possibly longer, but the longest I've gone was about 20 hours), I'm not sure what the benefit of hibernate is these days.
Nov 10, 2007, 7:23 am
#51
FlyerTalk Evangelist
Join Date: May 2002
Location: Pittsburgh
Programs: MR/SPG LT Titanium, AA LT PLT, UA SLV, Avis PreferredPlus
Posts: 31,007
I was about to post the same thing. I haven't used hibernate in years. I shut down only when I need to reboot to clean up some problem. Otherwise I just put it in sleep. It wakes up in about 5 seconds.
Nov 10, 2007, 7:40 am
#52
Join Date: May 2004
Location: Exclusively OMNI/PR, for Reasons
Posts: 4,188
Nov 10, 2007, 7:48 am
#53
Join Date: Nov 2000
Location: Upcountry Maui, HI
Posts: 13,308
-David
Nov 10, 2007, 8:08 am
#54
FlyerTalk Evangelist
Join Date: Dec 1999
Programs: Marriott Lifetime Titanium
Posts: 15,352
One thing not mentioned in this article:
Buy the minimum amount of RAM for you laptop from the manufacturer and also place an order with Crucial, OWC, or any of the other 3rd party memory makers.
I have an iMac that I bought 2 weeks ago. It came with 1GB of RAM. A 1 GB upgrade from Apple was $150 (taking the system to 2GB). I bought 4 GB (2 2GB DIMMs) for $167 from OWC and maxed it out and put the Apple chip it came with on eBay. I like using my Apple for video, pictures, along with having tons of things going at the same time.
You might not need the memory now, but why bother waiting and then upgrading later? Sure, memory may be cheaper then, but then you'd probably have to discard anything you put in there now to handle larger memory DIMMs.
To answer your question, though, I have 2GB in my Dell D820 laptop and it is more than adequate for my work purposes. I am still on XP, though.
Buy the minimum amount of RAM for you laptop from the manufacturer and also place an order with Crucial, OWC, or any of the other 3rd party memory makers.
I have an iMac that I bought 2 weeks ago. It came with 1GB of RAM. A 1 GB upgrade from Apple was $150 (taking the system to 2GB). I bought 4 GB (2 2GB DIMMs) for $167 from OWC and maxed it out and put the Apple chip it came with on eBay. I like using my Apple for video, pictures, along with having tons of things going at the same time.
You might not need the memory now, but why bother waiting and then upgrading later? Sure, memory may be cheaper then, but then you'd probably have to discard anything you put in there now to handle larger memory DIMMs.
To answer your question, though, I have 2GB in my Dell D820 laptop and it is more than adequate for my work purposes. I am still on XP, though.
Nov 10, 2007, 8:10 am
#55
A FlyerTalk Posting Legend
Original Poster
Join Date: Jul 2003
Location: NYC (formerly BOS/DCA)
Programs: UA 1K, IC RA
Posts: 60,745
Wow. I didn't know that the easy to this question would generate this much discussion. 
I'll definitely buy the minimum from the manufacturer and get my upgrade independently. I think I'll start with 2 GB and then go from there. Thanks all!
I'll definitely buy the minimum from the manufacturer and get my upgrade independently. I think I'll start with 2 GB and then go from there. Thanks all!
Nov 10, 2007, 8:10 am
#56
FlyerTalk Evangelist
Join Date: Dec 1999
Programs: Marriott Lifetime Titanium
Posts: 15,352
Sometimes I don't power on that quickly. To me, hibernate is "off." I rarely reboot/power down unless I'm installing updates that require it.
Nov 10, 2007, 9:29 am
#57
Suspended
Join Date: Jul 2007
Programs: AAdvantage, SkyMiles, USAir, Singapore, BA
Posts: 602
Besides my 30+ years of work in the industry?
http://www.kingston.com/newtech/MKF_...whitepaper.pdf
http://en.wikipedia.org/wiki/Dual-channel_architecture
http://www.hardwaresecrets.com/article/133/1
Yes, it really does access both memory channels in parallel, effectively doubling the bandwidth to memory, but you will never see a doubling of performance by switching to dual channel mode. In fact, as you noted, most synthetic benchmarks show only a small improvement when using dual channel mode over single channel mode. There is, in fact, a real world performance boost when using dual channel mode, but the amount of performance boost you actually see is highly application dependent. Since all the "magic" to support dual-channel mode is already built into the chipset's memory controller, you don't lose anything by taking advantage of it.
BTW, Tom's Hardware just isn't what it used to be. Not sure what article you are citing there, or which crucial article you are citing.
Edit: I guess you are citing the Tom's hardware article linked from the wikipedia article on dual channel memory. They try to explain why there isn't as big of a performance boost today as when the dual channel memory architecture was introduced. Basically, they say it's the cache effect. Large 2MB and 4MB caches tend to hide long latency to/from memory, and that's true to some extent. Eventually, the cache will become full and cache lines will need to be displaced. I imagine that if you're running applications like large files in photo shop that are memory bound, dual channel memory will do a decent job at increasing your performance. If you're just accessing random bytes from all over the place, it probably won't do anything to increase your performance.
Still not sure what Crucial article you were trying to cite.
-David
http://www.kingston.com/newtech/MKF_...whitepaper.pdf
http://en.wikipedia.org/wiki/Dual-channel_architecture
http://www.hardwaresecrets.com/article/133/1
Yes, it really does access both memory channels in parallel, effectively doubling the bandwidth to memory, but you will never see a doubling of performance by switching to dual channel mode. In fact, as you noted, most synthetic benchmarks show only a small improvement when using dual channel mode over single channel mode. There is, in fact, a real world performance boost when using dual channel mode, but the amount of performance boost you actually see is highly application dependent. Since all the "magic" to support dual-channel mode is already built into the chipset's memory controller, you don't lose anything by taking advantage of it.
BTW, Tom's Hardware just isn't what it used to be. Not sure what article you are citing there, or which crucial article you are citing.
Edit: I guess you are citing the Tom's hardware article linked from the wikipedia article on dual channel memory. They try to explain why there isn't as big of a performance boost today as when the dual channel memory architecture was introduced. Basically, they say it's the cache effect. Large 2MB and 4MB caches tend to hide long latency to/from memory, and that's true to some extent. Eventually, the cache will become full and cache lines will need to be displaced. I imagine that if you're running applications like large files in photo shop that are memory bound, dual channel memory will do a decent job at increasing your performance. If you're just accessing random bytes from all over the place, it probably won't do anything to increase your performance.
Still not sure what Crucial article you were trying to cite.
-David
I built a personal computer from parts in 1976, so I have a little experience in the field, too.
The Crucial article: http://www.crucial.com/store/partspe...E=CT25664AC667
Nov 10, 2007, 3:24 pm
#58
Join Date: Aug 2001
Location: A Southern locale that ain't the South.
Programs: Bah, HUMBUG!
Posts: 8,014
The memory bus is still only 64 bits wide, so there's no way to cram 128 bits into it no matter what you do (see the picture on Page 3 of the Intel article). From the CPU's point of view, data is going to and from the memory system 64 bits at a time. All that dual-channel memory accomplishes is to get the next 64 bits to the bus sooner than if the whole system had to wait for the setup (row and column strobes) of the next memory operation. This is called "interleaving" - which I learned about when I went to FE school at IBM in 1966.
I built a personal computer from parts in 1976, so I have a little experience in the field, too.
The Crucial article: http://www.crucial.com/store/partspe...E=CT25664AC667
I built a personal computer from parts in 1976, so I have a little experience in the field, too.
The Crucial article: http://www.crucial.com/store/partspe...E=CT25664AC667
Which Personal Computer would that have been back in 1976?
That would be 64 bits per module. Dual Channel DDR operates at 128 bit width on most chipsets. For some, two memory controllers run each module at 64 bits separately allowing parallel access. Either way, you get a bandwidth increase. Even with dual channel you will not saturate the bus given current memory speeds..
Here're some benchmarks. You were saying???
Code:
SiSoftware Sandra Benchmark Results Int Buff'd iSSE2 Memory Bandwidth : 3658 MB/s Float Buff'd iSSE2 Memory Bandwidth : 3674 MB/s Results Interpretation : Higher index values are better. Int Buff'd iSSE2 Memory Bandwidth Assignment : 3626 MB/s Scaling : 3616 MB/s Addition : 3671 MB/s Triad : 3722 MB/s Data Item Size : 16 byte(s) Buffering Used : Yes Offset Displacement Used : Yes Bandwidth Efficiency : 57% (estimated) Float Buff'd iSSE2 Memory Bandwidth Assignment : 3681 MB/s Scaling : 3679 MB/s Addition : 3698 MB/s Triad : 3639 MB/s Data Item Size : 16 byte(s) Buffering Used : Yes Offset Displacement Used : Yes Bandwidth Efficiency : 57% (estimated) Performance Test Status Run ID : ZEPTO-6024W on 22 August 2007 at 19:11:03 Platform Compliance : Win32 x86 Memory Used by Test : 512MB NUMA Support : No SMP (Multi-Processor) Benchmark : No Total Test Threads : 1 Multi-Core Test : Yes SMT (Multi-Threaded) Benchmark : No Processor Affinity : P0C0T0 System Timer : 14MHz Page Size : 4kB Use Large Memory Pages : No Features SSE Technology : Yes SSE2 Technology : Yes SSE3 Technology : Yes Supplemental SSE3 Technology : Yes SSE4 Technology : No EMMX - Extended MMX Technology : No SSE4A Technology : No HTT - Hyper-Threading Technology : No Chipset 1 Model : Inventec Corp Mobile PM965/GM965/GL960 Express Processor to DRAM Controller Revision : A4 Front Side Bus Speed : 4x 200MHz (800MHz data rate) Width : 64-bit Maximum Bus Bandwidth : 6400MB/s (estimated) Logical/Chipset 1 Memory Banks Bank 0 : 1GB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Bank 1 : 1GB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Shared Memory : 8MB Channels : 1 Memory Bus Speed : 4x 333MHz (1332MHz data rate) Width : 64-bit Performance Acceleration Technology : No Memory Controller in Processor : No Maximum Memory Bus Bandwidth : 10656MB/s (estimated) Performance Tips Notice 5405 : System bandwidth appears FSB limited. Attempt to increase FSB. Notice 5008 : To change benchmarks, click Options. Notice 5004 : Synthetic benchmark. May not tally with 'real-life' performance. Notice 5006 : Only compare the results with ones obtained using the same version! Warning 5010 : Cannot use Large Memory Pages due to lack of privileges. Warning 2544 : System/Video shared memory greatly reduces performance. Use external video card. Tip 2 : Double-click tip or press Enter while a tip is selected for more information about the tip. SiSoftware Sandra Benchmark Results Int Buff'd iSSE2 Memory Bandwidth : 3847 MB/s Float Buff'd iSSE2 Memory Bandwidth : 3859 MB/s Results Interpretation : Higher index values are better. Int Buff'd iSSE2 Memory Bandwidth Assignment : 3773 MB/s Scaling : 3752 MB/s Addition : 3926 MB/s Triad : 3937 MB/s Data Item Size : 16 byte(s) Buffering Used : Yes Offset Displacement Used : Yes Bandwidth Efficiency : 60% (estimated) Float Buff'd iSSE2 Memory Bandwidth Assignment : 3823 MB/s Scaling : 3832 MB/s Addition : 3922 MB/s Triad : 3862 MB/s Data Item Size : 16 byte(s) Buffering Used : Yes Offset Displacement Used : Yes Bandwidth Efficiency : 60% (estimated) Performance Test Status Run ID : ZEPTO-6024W on 22 August 2007 at 21:34:45 Platform Compliance : Win32 x86 Memory Used by Test : 512MB NUMA Support : No SMP (Multi-Processor) Benchmark : No Total Test Threads : 1 Multi-Core Test : Yes SMT (Multi-Threaded) Benchmark : No Processor Affinity : P0C0T0 System Timer : 14MHz Page Size : 4kB Use Large Memory Pages : No Features SSE Technology : Yes SSE2 Technology : Yes SSE3 Technology : Yes Supplemental SSE3 Technology : Yes SSE4 Technology : No EMMX - Extended MMX Technology : No SSE4A Technology : No HTT - Hyper-Threading Technology : No Chipset 1 Model : Inventec Corp Mobile PM965/GM965/GL960 Express Processor to DRAM Controller Revision : A4 Front Side Bus Speed : 4x 200MHz (800MHz data rate) Width : 64-bit Maximum Bus Bandwidth : 6400MB/s (estimated) Logical/Chipset 1 Memory Banks Bank 0 : 512MB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Bank 1 : 512MB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Bank 4 : 1GB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Bank 5 : 1GB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Shared Memory : 8MB Channels : 2 Memory Bus Speed : 4x 333MHz (1332MHz data rate) Width : 64-bit Performance Acceleration Technology : No Memory Controller in Processor : No Maximum Memory Bus Bandwidth : 21312MB/s (estimated) Performance Tips Notice 5405 : System bandwidth appears FSB limited. Attempt to increase FSB. Notice 5008 : To change benchmarks, click Options. Notice 5004 : Synthetic benchmark. May not tally with 'real-life' performance. Notice 5006 : Only compare the results with ones obtained using the same version! Warning 5010 : Cannot use Large Memory Pages due to lack of privileges. Warning 2544 : System/Video shared memory greatly reduces performance. Use external video card. Tip 2 : Double-click tip or press Enter while a tip is selected for more information about the tip. SiSoftware Sandra Benchmark Results Int Buff'd iSSE2 Memory Bandwidth : 3980 MB/s Float Buff'd iSSE2 Memory Bandwidth : 4015 MB/s Results Interpretation : Higher index values are better. Int Buff'd iSSE2 Memory Bandwidth Assignment : 3951 MB/s Scaling : 3914 MB/s Addition : 4030 MB/s Triad : 4027 MB/s Data Item Size : 16 byte(s) Buffering Used : Yes Offset Displacement Used : Yes Bandwidth Efficiency : 62% (estimated) Float Buff'd iSSE2 Memory Bandwidth Assignment : 3941 MB/s Scaling : 3959 MB/s Addition : 4102 MB/s Triad : 4058 MB/s Data Item Size : 16 byte(s) Buffering Used : Yes Offset Displacement Used : Yes Bandwidth Efficiency : 63% (estimated) Performance Test Status Run ID : ZEPTO-6024W on 22 August 2007 at 19:33:08 Platform Compliance : Win32 x86 Memory Used by Test : 512MB NUMA Support : No SMP (Multi-Processor) Benchmark : No Total Test Threads : 1 Multi-Core Test : Yes SMT (Multi-Threaded) Benchmark : No Processor Affinity : P0C0T0 System Timer : 14MHz Page Size : 4kB Use Large Memory Pages : No Features SSE Technology : Yes SSE2 Technology : Yes SSE3 Technology : Yes Supplemental SSE3 Technology : Yes SSE4 Technology : No EMMX - Extended MMX Technology : No SSE4A Technology : No HTT - Hyper-Threading Technology : No Chipset 1 Model : Inventec Corp Mobile PM965/GM965/GL960 Express Processor to DRAM Controller Revision : A4 Front Side Bus Speed : 4x 200MHz (800MHz data rate) Width : 64-bit Maximum Bus Bandwidth : 6400MB/s (estimated) Logical/Chipset 1 Memory Banks Bank 0 : 512MB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Bank 1 : 512MB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Bank 4 : 512MB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Bank 5 : 512MB DDR2-SDRAM 5.0-5-5-15 (tCL-tRCD-tRP-tRAS) CR1 Shared Memory : 8MB Channels : 2 Memory Bus Speed : 4x 333MHz (1332MHz data rate) Width : 64-bit Performance Acceleration Technology : No Memory Controller in Processor : No Maximum Memory Bus Bandwidth : 21312MB/s (estimated) Performance Tips Notice 5405 : System bandwidth appears FSB limited. Attempt to increase FSB. Notice 5008 : To change benchmarks, click Options. Notice 5004 : Synthetic benchmark. May not tally with 'real-life' performance. Notice 5006 : Only compare the results with ones obtained using the same version! Warning 5010 : Cannot use Large Memory Pages due to lack of privileges. Warning 2544 : System/Video shared memory greatly reduces performance. Use external video card. Tip 2 : Double-click tip or press Enter while a tip is selected for more information about the tip.
Nov 10, 2007, 4:59 pm
#59
Suspended
Join Date: Jul 2007
Programs: AAdvantage, SkyMiles, USAir, Singapore, BA
Posts: 602
Sphere I (M6800, 750kHz, 4k RAM 1k ROM 9" 16x32 char. CRT)
Tell me what results you want the benchmarks to show, and I'll write the diagnostic that shows them.
168-pin Dual Inline Memory Modules have 64-bit data paths. Dual-channel memory controllers do not access two DIMMs simultaneously. The data goes in and out of the CPU 64 bits at a time. By overlapping setup in one module with gating data onto the bus in the other, some increased throughput can be realized.
Well, gee, I would expect doubling the memory controllers might have a dramatic effect. But that has nothing to do with the fact that dual-channel memory doesn't have twice the bandwidth of single-channel. Adding controllers is...cheating.
Tell me what results you want the benchmarks to show, and I'll write the diagnostic that shows them.
168-pin Dual Inline Memory Modules have 64-bit data paths. Dual-channel memory controllers do not access two DIMMs simultaneously. The data goes in and out of the CPU 64 bits at a time. By overlapping setup in one module with gating data onto the bus in the other, some increased throughput can be realized.
Well, gee, I would expect doubling the memory controllers might have a dramatic effect. But that has nothing to do with the fact that dual-channel memory doesn't have twice the bandwidth of single-channel. Adding controllers is...cheating.
Last edited by CessnaJock; Nov 10, 2007 at 5:06 pm
Nov 10, 2007, 7:18 pm
#60
Join Date: Nov 2000
Location: Upcountry Maui, HI
Posts: 13,308
Correct.
Yes, they do. The MCH issues 2 requests in parallel, one to each memory channel (or one to each DIMM). That's what's shown on page 3 of the Kingston article.
That's true in single channel mode. In dual channel mode, they transmit 128 bits of data at a time in between the processor and memory.
In order to delve deeper, we'd have to look at the architecture of the front side bus, which is proprietary. I believe that you're correct that there's 64 data pins for the FSB, but they don't have to transmit two entire packets for 128 bits of data. They transmit the data in two phases of the clock, without an intermediate handshake. In 64-bit single channel mode, they probably have to transmit two entire address/data packets to transmit 128 bits of sequentially aligned data between the processor and the MCH, so there's more overhead in single-channel mode for the same amount of sequential data. But, because the basic data width is still 64-bits, you won't see a doubling of performance when dual-channel mode is enabled.
I think we've covered this to death at this point.
-David
Dual-channel memory controllers do not access two DIMMs simultaneously.
The data goes in and out of the CPU 64 bits at a time. By overlapping setup in one module with gating data onto the bus in the other, some increased throughput can be realized.
In order to delve deeper, we'd have to look at the architecture of the front side bus, which is proprietary. I believe that you're correct that there's 64 data pins for the FSB, but they don't have to transmit two entire packets for 128 bits of data. They transmit the data in two phases of the clock, without an intermediate handshake. In 64-bit single channel mode, they probably have to transmit two entire address/data packets to transmit 128 bits of sequentially aligned data between the processor and the MCH, so there's more overhead in single-channel mode for the same amount of sequential data. But, because the basic data width is still 64-bits, you won't see a doubling of performance when dual-channel mode is enabled.
I think we've covered this to death at this point.
-David