psychologist

Ian Jackson tackles the sometimes difficult to understand subject of memory timings and tells you how to max out your bandwidth.

Far more press is given to the processor and video card when it comes to getting the best performance out of a PC, yet the difference between quality memory and generic types can yield gains just as significant as a CPU speed boost. This improvement is almost entirely due to differences in the rated latencies or 'timings' of the RAM. Before we look at memory timings in more detail, it is important to understand what kind of memory you have, its nomenclature and also how it works.


The memory used in most modern PCs is one of two kinds; either DDR or DDR2. The two types operate in a similar way, though at rather different speed ranges, virtue of the fact that DDR2's effective clock speed is considerably higher. Both types of memory are named in accordance with their theoretical maximum memory bandwidth in MB/s, though increasingly people are also referring to different types by their speed as well (for example, 'DDR400'). The most common speed of memory is PC3200, which has a peak theoretical bandwidth of 3.2GB/s and operates at a frequency of 200MHz. It is called DDR400 because DDR transfers its data on both the rising and falling edge of each clock cycle. Its 'effective memory speed' if it were SDR is therefore 400MHz.
PC3200 is currently the fastest officially recognised DDR memory speed as supported by JDEC, though many performance memory manufacturers produce memory a good deal faster. PC3500 was first to appear and become popular amongst overclockers, though you can now get PC4000, PC4400 and even PC4800 (DDR600) from a couple of manufacturers. Unfortunately such fast memory is very expensive to produce and yields are much lower than with more conventional chips. DDR2 is slowly taking over from DDR as it allows speeds as high as PC8000. It can achieve these speeds thanks to 4-bit data pre-fetch technology. Whereas with traditional DDR two sets of data are read from and written to the memory core, four sets of data are processed on DDR2 devices at a time. This means that for a given DDR 2 speed, the memory only has to operate at half the frequency of DDR. A DDR2-400 module therefore only runs at 100MHz, whereas a DDR-400 module has to run at 200MHz for the same performance.

Circuit Board
Sticks of memory are made up of printed circuit boards onto which memory chips are attached. As with a processor, the memory chip itself is an integrated circuit made up of millions of transistors and capacitors. In DRAM, a transistor and a capacitor are paired to create a memory cell, which can represent a single bit of data. A charged capacitor has a value of 1, and once discharged a value of 0. The transistor acts as a switch that lets the control circuitry of the memory chip read the capacitor or change its state. All types of RAM module that are used in PCs use volatile memory chips, which means they only retain information whilst they receive current. This is because capacitors 'leak' their charge, so without the memory controller or CPU recharging those capacitors holding a status of 1, they discharge and lose the information.

The speed at which this charge is dynamically refreshed is the speed at which the memory operates in MHz. Naturally there has to be a way to get information in and out of memory modules, so there is a whole support infrastructure of other circuits that allow the identification of each row and column (this is what RAS and CAS refer to), make sure they are being refreshed properly and read or restore the signal from a cell. When you change the memory timings, you change the speed at which these auxiliary circuits operate, and thus affect the overall performance of the module. In many situations changing from loose timings to tight ones can have as much as effect on overall system performance as upgrading your CPU to the next speed grade, or overclocking the actual speed of the memory by a significant amount.

All memory modules have a small chip on them called the SPD or 'serial presence detect'. It contains information about the module that helps the motherboard to set the correct settings in order to retain stability. Some of the most important settings stored in the SPD refer to the memory timings, and by default all motherboards will set the memory timings to those requested by the module. Fortunately, these SPD values can be overridden in the BIOS and set manually. This is useful for increasing performance, as often memory is capable of operating significantly faster than the SPD suggests. By using a less aggressive SPD setting, memory manufacturers can maximise yields, as a higher proportion of their modules will operate stably.

Balance
Memory speed and the tightness of the timings need to be balanced if the maximum performance is to be extracted, as although a module might be running at a huge frequency, the usable data might be delayed by the efficiency of the auxiliary circuits. DDR-2 typically has a much higher latency than DDR, which is why PC2-3200 is considerably slower than decent PC3200. The much higher clock speeds obviously compensate for this, but some applications that are very latency sensitive (for example some games) will see very little benefit from ever increasing bandwidth, but will show excellent gains when setting the timings to a more aggressive value. As you increase memory speed you will normally find that the timings have to be relaxed to retain stability. Relax the timings too far however, and you actually reduce performance versus running stock speed with the timings at full tightness. Getting the best out of your modules is therefore a bit of a balancing act and requires considerable testing and patience.

As mentioned before, information is stored in memory by separating it into rows and columns. The chips are then accessed via control signals including the Row Address Strobe (RAS), Column Address Strobe (CAS) and a number of other commands (DQ). The Command Rate is defined as the amount of time it takes for the RAS to be executed after the correct memory chip has been identified. Before the row can then become active (enabling the columns to be accessed) the controller has to wait for 2-4 cycles. This is known as the RAS-to-CAS Delay (tRCD). It then sends the read command, which is also followed by a delay in the form of the CAS latency. After the data has been accessed it is sent to the DQ. The controller then has to deactivate the row so that it can be read in the future. This is done within the RAS precharge time (tRP) and again typically takes 2-4 cycles. The last of the four most famous timings is the Active to Precharge Delay, which defines the fewest number of cycles a row has to be activated for before it can be turned off again. In DDR this value is typically between 5 and 8.

In overclocking forums you will typically see people quoting numbers when discussing which memory timings these are using. Strictly, these should be done in a specific order, where for example 2-3-3-6 would mean a CAS Latency of 2, a tRCD and tRP of 3 and a tRAS of 6. Unfortunately people are rather sloppy so you will see a number of different permutations being quoted, though it is normally relatively easy to understand what they mean.

CAS Latency
Of all the timings, the CAS latency is by far the most famous. This hankers back to the SDRAM days, where the CAS latency was the most important performance determinant. Although this is no longer the case, you will often see stores advertising their more illustrious memory as 'CAS 2'. Whilst this may not be the most helpful timing to quote, memory capable of running at a low CAS latency is normally quite able of running the other timings at least reasonably aggressively. The CAS Latency refers to the length of time in clock cycles it takes for a request sent from the memory controller to read a memory location and send it to the module's output pins. The column selection is the last step before the data is read from the physical location of the memory cell, so the CAS Latency measures the number of cycles between the time the request for data is sent to the actual memory location and the time the data is extracted from the module.
Cheap generic DDR memory will have a CAS latency of 3, mid ranged memory 2.5 and the best memory a value of 2. The benefit of moving from CAS 3 to CAS 2 is a significant increase in memory performance (as much as 2%), so if you are a benchmarking fiend, you will want this value as tight as possible. Some Intel DDR boards actually allow you to set the CAS latency to 1.5, but this normally causes serious instability and offers no increase in performance. In DDR-2 memory, the CAS latency is typically a lot less aggressive than in DDR, and has a value of 4 or higher. More expensive premium modules have a CAS latency of 3, which provides somewhat better memory performance as well.

tRCD Latency
The tRCD or RAS to CAS delay is becoming an increasingly important limiting factor in determining overall memory performance, especially when overclocking at higher memory speeds. It is the time it takes between the activation of the line (RAS) and the column (CAS). It can usually be set to 4, 3 or 2, with 2 being the fastest. Because RAS Precharge (tRP) is also set as 4, 3, or 2, the two settings are often grouped together or confused. The difference between a tRCD of 4 (the setting found on most inexpensive memory) to 3 yields a similar improvement to CAS latency in performance. Reducing this a step further to a value of 2 yields an even bigger improvement. A number of overclocking manufacturers, most notably Mushkin and OCZ are now sacrificing CAS latency in order to get the lower tRCD timings because it tends to yield better overall performance.

As with DDR, DDR2 benefits from lower tRCD timings. On inexpensive DDR2, especially with high speed DDR667, you can expect a value of 5 for tRCD. Performance will improve considerably with a value of 4 which even the cheapest of modules will normally allow with full stability, and again by reducing it further to 3. For a tRCD of 3 or less you will have to pay a heavy premium for performance DDR2 from a top memory manufacturer. Values of 2 are possible with the most expensive modules, but require a lot of extra voltage and can still produce instability on some motherboards. For most people 3 should be the lowest realistic value to be shooting for.

tRP Latency
The tRP or Ras Precharge is the time it takes from disabling the access to a line of data to the beginning of accessing another line of data. As mentioned before, it is often grouped together with tRCD as both have similar values of 4 3 or 2 on DDR and 5 down to 2 on DDR-2. However, unlike tRCD, the tRP value has a far less noticeable impact on performance, and going from 4 right down to the tightest 2 setting is unlikely to yield more than a 1% improvement in memory performance. If you are unstable at the most tight memory timings, you should try increasing the latency on the tRP value before tRCD or CAS, as it could improve stability without impacting performance.

tRAS Latency
tRAS or Active to Precharge Delay is also known as the 'Bank cycle time' and is the number of cycles necessary to develop the full charge differential between bit and reference lines to restore the data in the memory cells, or in other words, how long the memory has to wait until the next access to the memory can be initialised. This timing is normally set to a value of 8 on cheaper DDR memory, with it reducing to 6 or 5 on more expensive types. By reducing the tRAS from 8 to 6 you can normally see a very minor improvement in performance - a similar amount as achieved by tightening the tRP, though the gains certainly are not worth sacrificing the tightness on other timings for, especially if it comes at a price to stability.

On DDR2, the speed of the tRAS is somewhat higher than DDR, going all the way as high a 15 for some high speed modules. With such a slow tRAS, some slightly more significant performance gains can be had when tightening this value, though only by testing yourself in conjunction with your other timings settings can you get the best performance.

Command Rate
The Command Rate is - quite incorrectly - given less coverage than the other memory timings, but provides by far the largest increase in performance. Most motherboards allow values of either 2T or 1T to be set, though some also offer a 3T setting. The command rate is the time it takes from the activation of the memory chip to when the first command may be sent to the memory. The benefits of going from 2T to 1T can be an increase of 5% or more in performance, so sacrificing command rate for a higher clock speed or tighter timings in the other settings is generally not worth doing. Command rate is especially important on Athlon 64 systems, and can be set using a windows program called A64tweaker if your motherboard's BIOS doesn't have the setting available. Some motherboards have trouble running the memory at tight command rates beyond a certain speed, so if you are building a machine with this in mind, carefully read reviews of the boards you are considering to ensure you don't buy a dog.

Bank Interleaving and other settings
One setting available on many motherboards is Bank Interleaving, which can be set to disabled, 2-way or 4-way. Bank interleaving changes the way banks of memory are accessed and refreshed. A staggered effect is created to minimise the delay. Where possible you should set bank interleaving to 4-way for maximum performance, though not all boards allow this setting to be changed. Bank interleaving can sometimes increase performance by a similar amount to the command rate, so it is well worth hunting around your BIOS to see if it is present!

Some motherboards allow numerous other memory timings to also be changed in addition to the major ones looked at previously. Discussing all of these would require the rest of this issue, and as most of them shouldn't really be changed, wouldn't be particularly helpful either! Most of these are normally allocated by the SPD file on the memory, and setting them to incorrect values can negatively affect performance or stability. If you have hours to kill however, some experimenting with these other values could yield a few extra minor points in your preferred memory bandwidth tester.

Regardless of whether you are a power user, or just a casual tweaker, by changing the right memory settings you could see a sizable improvement in the performance of your computer. Always test your new settings using memtest (www.memtest86.com) before booting into the OS, or you could risk corrupting your data should your memory not be able to handle more aggressive settings.

Penjelasan tentang Timing Memory

Ian jackson me menjatuhkkan anggapan yang berkaitan dengan subyek tentang memory timing yang selama ini sulit dipahami dan memberi tahu anda bagaimana cara memaksimalkan bandwidth.


Ulasan press lebih banyak diperuntukkan bagi prosesor dan video card ketika dihubungkan dengan kinerja sebuah PC, namun perbedaan antara memory yang berkualitas dan memori generik juga bisa memberikan peningkatan yang signifikan pada kecepatan CPU. Peningkatan kecepatan ini hampir semuanya dikarenakan perbedaan pada latency atau “timing” RAM. Sebelum kita melihat pada timing memory secara lebih rinci, sangat penting bagi anda untuk memahami terlebih dahulu jenis memory apa yang anda punyai, nomenklatur (tatanama) nya dan bagaimana memory tersebut berfungsi.

Memor yang digunakan pada kebanyakan PC modern adalah terdiri dari dua jenis; bisa DDR atau DDR2. dua jenis memory ini beroperasi secara hampir sama, walaupun dengan kisaran kecepatan yang berbeda, karena kenyataan bahwa kecepatan clock DDR2 secara efektif adalah cukup tinggi. Kedua jenis memory ini dinamai sesuai dengan bandwidth memory maksimum secara teoritis dalam satuan MB/s mesikpun semakin banyak orang yang menyebutnya berdasarkan kecepatannya (sebagai contoh ‘DDR400’). Kecepatan memory yang paling umum dipakai adalah PC3200, yang memiliki bandwidth maksimal 3,2Gb/s dan beroperasi pada ferkuensi 200Mhz. Memori seperti ini disebut dengan DDR400 karena DDR ini mentransfer datanya pada setiap ujung atas dan bawah dari setiap siklus clock. Kecepatan efektif memory ini jika memory tersebut adalah SDR dengan demikian adalah 400.

PC3200 saat ini merupakan kecepatan memory DDR tercepat secara resmi yang diakui oleh JDEC, walaupun banyak pabrikan memory memproduksi memory yang lebih cepat lagi. PC3500 adalah yang pertama kali muncul dan paling populer di antara para overclocker, walaupun mungkin anda sekarang bisa mendapatkan PC400, PC4400 dan bahkan PC4800 (DDR600) dari beberapa pabrikan. Sayangnya memory yang sangat cepat tersebut sangat mahal harganya dan hasilnya lebih rendah jika dibandingkan dengan chip-chip konvensional. DDR2 secara perlahan-lahan mengambil alih peran DDR karena memungkinkan peningkatan kecepatan sebesar PC8000. Memory ini bisa mencapai kecepatan sampai demikian kaerna teknologi pre-fetch data 4-bit. Jika pada DDR lama dua set data dibaca dan ditulis dari dan ke inti memory, maka pada memory DDR2 ada empat set data yang diproses dalam satu waktu. Ini berarti bahwa pada kecepatan DDR2 tertentu, memory ini harus beroperasi pada setengah kali frekuensi DDR. Oleh karena itu modul DDR2-400 hanya berjalan pada frekuensi 100Mhz, sementara itu modul DDR-400 harus berjalan pada frekuensi 200Mhz untuk mendapatkan kecepatan yang sama.

CAS Latency
Dari semua pengaturan timing, CAS latency adalah yang paling menonjol. Keinginan ini kembali ke jaman SDRAM dulu, dimana CAS latency adalah penentuk kinerja yang paling penting. Walaupun sudah tidak lagi demikian, anda akan sering melihat toko-toko yang mengiklankan memorynya dengan “CAS 2”. Meskipun ini bukan timing yang paling tepat untuk dikutip, memory yang mampu berjalan pada CAS latency yang rendah biasanya lebih mampu berjalan pada timing lain dengan sedikit lebih agresif. CAS latency mengacu pada lama waktu yang dibutuhkan oleh clock cycle untuk membawa request yang dikirimkan dari memory controller untuk membaca lokasi memory dan mengirimkannya ke pin output modul. Pemilihan kolom adalah langkah terakhir sebelum data dibaca dari lokasi fisik sel memory, sehingga CAS latency mengukur jumlah cycle antara waktu permintaan data dikirimkan ke lokasi memory yang sebenarnya dan waktu data tersebut diambil dari modul.

Memory DDR generik yang murah mempunyai CAS latency 3, memory kelas menengah mempunyai CAS latency 2,5 dan memory yang paling bagus memiliki CAS latency 2. manfaat berganti dari CAS 3 ke CAS 2 adalah peningkatan yang cukup signifikan pada kinerja memory (kira-kira sebanyak 2%), sehingga jika anda adalah tukang benchmark, anda pasti ingin mendapatkan nilai ini seketetat mungkin. Beberapa motherboard DDR Intel memungkinkan anda untuk mengatru CAS latency sampe nilai 1,5, tetapi ini biasanya akan menyebabkan ketidakstabilan yang cukup serius dan tidak memberikan peningkatan kinerja. Pada memory DDR-2, CAS latency biasanya jauh kurang agresif daripada DDR, dan mempunyai nilai CAS sebesar 4 atau lebih. Beberapa modul memory kelas premium yang lebih mahal mempunyai CAS latency 3, yang memberikan kinerja memory yang lebih baik.

Labels: laptops and computers