- Dec 13, 2004
- 7,973
- 8
- 0
I've posted this a few times in response to various questions, so I thought I should make a dedicated thread about it, including the way I thought it up. As far as I can tell, its correct but post if I'm wrong with this.
To truly understand the formula of RAM speed for a A64, one must first understand how the A64 system arcitechure works. All data travels through the "Hyper Transport Tunnel", which is really fast, 800 or 1000MHz, and can carry 20GB/s data.
There are a lot of "clocks" in the A64 system, most of which operate independent of each other. There is the system clock, which is where the CPU gets its Speed from, where the AGP and PCI get theirpeed from in unlocked systems, and the Hyper Transport Tunnel gets its total speed from This is commonly called the HTT.
In most systems, however, the system clock only controlls the speed of the CPU and the Hyper Transport Tunnel, since the PCI/AGP and PCIe are all locked at their own optimal frequency.
Additional Comments by RichUK:
The PCI, AGP, System(FSB/HTT) BUS run on a 6:3:1 ratio from the system BUS clock crystal when running at 200Mhz which is standard for A64 systems. The system clock runs off of a clock crystal which controls the frequency of the electrical pulses across the system BUS, and every other BUS runs off of this main BUS for communication to the CPU and RAM.
JEDEC (Joint Electron Device Engineering Council) have set into play a nominal FSB/HTT frequency standard of 200Mhz (so the clock crystal for the system bus runs at 200Mhz). The PCI BUS was designed to run at 33Mhz and the AGP frequency was designed to work at 66Mhz.
The ratios (6:3:1) are as follows:
1:1 = System Clock/BUS = 200Mhz
1:3 = AGP Clock/BUS = 200Mhz ÷ 3 = 66.66? - 66Mhz
1:6 = PCI Clock/BUS = 200Mhz ÷ 6 = 33.33? - 33Mhz
System Clock = 200Mhz
AGP = 66Mhz
PCI = 33Mhz
When Overclocking you will find it is very beneficial to have a board with PCI/AGP locks as the only frequency you are really looking to increase is the CPU and Memory frequency. As i have previously stated the AGP/PCI Bus were designed to run at these frequencies, and clocked any higher you will find that you will get data corruption, on the PCI BUS the sound card is usually the first one to go (just doesn?t operate due to driver failure), and then the HDD's end up getting corrupt registries etc, note these run on the PCI bus as well.
You can Overclock a motherboard with unlocked PCI/AGP frequencies but it will not be able to scale as high as a board with locks as there are more variables for Overclock hold backs (Obviously due to the unlocked BUSes).
If you take for example a AMD Athlon 3200+ Winchester:
Multipliers and Dividers
System clock HTT = 200mhz (1:1) - 200Mhz (Ram running at PC3200 DDR400)
Processor = 200mhz x10 - 2000Mhz (or 2Ghz)
L2 Cache = 200mhz x10 - 2000Mhz (of course this is on die so it runs at the same clock as the processor)
AGP bus = 66mhz (1:3) - 66Mhz (always by default, also same clock as PCI-X for 64bit SCSI cards)
PCI Bus = 33Mhz (1:6) - 33Mhz (always by default)
AMD's have a Hyper transport Link or just abbreviated to HT Links, and this has a variable multiplier (LDT) by default of x5 for Socket 939 and x4 for Socket 754, and this is multiplied by the system clock which would be:
200Mhz x 5-LDT 1000Mhz HT 2000Mts (million transfers a second) for Socket 939
200Mhz x 4-LDT 800Mhz HT 1600(million transfers a second) for socket 754
AMD use marketing for the 2000Mts, which they claim because the HT is duplex meaning it can communicant back and forth on the HT Link at the same time thus doubling the data bandwidth (1000Mhz HT x 2), and as stated in the original Post 20GB\s is the theoretical 2000 million transfers a second that AMD claims its Socket 939 chips are capable of, of course none of the bandwidth gets fully used only about 50% at the very most.
The term FSB for A64's is a misnomer. RAM speed is linked to CPU speed. The computer derives CPU speed from the system clock and its internal multiplier, most commonly from 8 to 14x in current systems, though most BIOSes support up to 25x. RAM speed is derived by dividing CPU speed by the dividend of the CPU multiplier and the RAM divider.
Matmatically, this is shown by the function R= C/(M/[D]), where C is CPU speed, D is the RAM divider (normally either 1 or a fraction), M is CPU multiplier, and R is RAM speed.
For .5x multipliers, To get your RAM speed, you must round your CPU multiplier UP. AMD did this for stability reasons, as rounding down would have increased RAM speed, and therefore the chance for system instabilty.
Example: A clawhammer 3200+ runs stock 10x200. You overclock it to 10x240, but your ram can't run 240MHz, so you run it with a 166MHz, or .83 divider. 2400/(10/[.83]) gives you exactly 200MHz.
This formula, combined with the Hyper Transport Tunnel, also explains why A64's exhibit no performance degradtion when running ram dividers. To use my previous example, 240x10 with a 166MHz RAM divider looks exactly the same a 12x200 to the computer.
I hope this little guide helps anwswer some questions people have, because it doesn't explain just what, but why. Explaining why something happens helps people find their own answers, and come up with new ones.
To truly understand the formula of RAM speed for a A64, one must first understand how the A64 system arcitechure works. All data travels through the "Hyper Transport Tunnel", which is really fast, 800 or 1000MHz, and can carry 20GB/s data.
There are a lot of "clocks" in the A64 system, most of which operate independent of each other. There is the system clock, which is where the CPU gets its Speed from, where the AGP and PCI get theirpeed from in unlocked systems, and the Hyper Transport Tunnel gets its total speed from This is commonly called the HTT.
In most systems, however, the system clock only controlls the speed of the CPU and the Hyper Transport Tunnel, since the PCI/AGP and PCIe are all locked at their own optimal frequency.
Additional Comments by RichUK:
The PCI, AGP, System(FSB/HTT) BUS run on a 6:3:1 ratio from the system BUS clock crystal when running at 200Mhz which is standard for A64 systems. The system clock runs off of a clock crystal which controls the frequency of the electrical pulses across the system BUS, and every other BUS runs off of this main BUS for communication to the CPU and RAM.
JEDEC (Joint Electron Device Engineering Council) have set into play a nominal FSB/HTT frequency standard of 200Mhz (so the clock crystal for the system bus runs at 200Mhz). The PCI BUS was designed to run at 33Mhz and the AGP frequency was designed to work at 66Mhz.
The ratios (6:3:1) are as follows:
1:1 = System Clock/BUS = 200Mhz
1:3 = AGP Clock/BUS = 200Mhz ÷ 3 = 66.66? - 66Mhz
1:6 = PCI Clock/BUS = 200Mhz ÷ 6 = 33.33? - 33Mhz
System Clock = 200Mhz
AGP = 66Mhz
PCI = 33Mhz
When Overclocking you will find it is very beneficial to have a board with PCI/AGP locks as the only frequency you are really looking to increase is the CPU and Memory frequency. As i have previously stated the AGP/PCI Bus were designed to run at these frequencies, and clocked any higher you will find that you will get data corruption, on the PCI BUS the sound card is usually the first one to go (just doesn?t operate due to driver failure), and then the HDD's end up getting corrupt registries etc, note these run on the PCI bus as well.
You can Overclock a motherboard with unlocked PCI/AGP frequencies but it will not be able to scale as high as a board with locks as there are more variables for Overclock hold backs (Obviously due to the unlocked BUSes).
If you take for example a AMD Athlon 3200+ Winchester:
Multipliers and Dividers
System clock HTT = 200mhz (1:1) - 200Mhz (Ram running at PC3200 DDR400)
Processor = 200mhz x10 - 2000Mhz (or 2Ghz)
L2 Cache = 200mhz x10 - 2000Mhz (of course this is on die so it runs at the same clock as the processor)
AGP bus = 66mhz (1:3) - 66Mhz (always by default, also same clock as PCI-X for 64bit SCSI cards)
PCI Bus = 33Mhz (1:6) - 33Mhz (always by default)
AMD's have a Hyper transport Link or just abbreviated to HT Links, and this has a variable multiplier (LDT) by default of x5 for Socket 939 and x4 for Socket 754, and this is multiplied by the system clock which would be:
200Mhz x 5-LDT 1000Mhz HT 2000Mts (million transfers a second) for Socket 939
200Mhz x 4-LDT 800Mhz HT 1600(million transfers a second) for socket 754
AMD use marketing for the 2000Mts, which they claim because the HT is duplex meaning it can communicant back and forth on the HT Link at the same time thus doubling the data bandwidth (1000Mhz HT x 2), and as stated in the original Post 20GB\s is the theoretical 2000 million transfers a second that AMD claims its Socket 939 chips are capable of, of course none of the bandwidth gets fully used only about 50% at the very most.
The term FSB for A64's is a misnomer. RAM speed is linked to CPU speed. The computer derives CPU speed from the system clock and its internal multiplier, most commonly from 8 to 14x in current systems, though most BIOSes support up to 25x. RAM speed is derived by dividing CPU speed by the dividend of the CPU multiplier and the RAM divider.
Matmatically, this is shown by the function R= C/(M/[D]), where C is CPU speed, D is the RAM divider (normally either 1 or a fraction), M is CPU multiplier, and R is RAM speed.
For .5x multipliers, To get your RAM speed, you must round your CPU multiplier UP. AMD did this for stability reasons, as rounding down would have increased RAM speed, and therefore the chance for system instabilty.
Example: A clawhammer 3200+ runs stock 10x200. You overclock it to 10x240, but your ram can't run 240MHz, so you run it with a 166MHz, or .83 divider. 2400/(10/[.83]) gives you exactly 200MHz.
This formula, combined with the Hyper Transport Tunnel, also explains why A64's exhibit no performance degradtion when running ram dividers. To use my previous example, 240x10 with a 166MHz RAM divider looks exactly the same a 12x200 to the computer.
I hope this little guide helps anwswer some questions people have, because it doesn't explain just what, but why. Explaining why something happens helps people find their own answers, and come up with new ones.