Until I found out about this, I had a lot of problems and so does a lot of other people, so I thought you would want to know.
I couldn't even install Windows! It would hang on loading the drivers.
I couldn't even install Windows! It would hang on loading the drivers.
The Athlon64 itself uses a single memory channel and only one controller on the CPU. Because of this, the Athlon 64 (non-FX) can actually use plain vanilla, unbuffered, non-ECC modules. There are, however, still some peculiarities of the Athlon64 systems in how they are handling memory which effects memory compatibility.
One of the most critical issues in this resepect relates to the first step in the memory access sequence, that is, the opening of a memory page and subsequent steps until a read command can be issued. All of these steps fall into the time frame referred to as RAS-to-CAS delay or tRCD. In a typical system, the memory controller on the chipset is running at bus speed. On the Athlon64, it is running at CPU speed, which means that it is A LOT faster. That means that the internal sequence of commands also goes faster, and that's where all systems are getting hiccups if the memory chips used can only run at a high tRCD.
A very trivial explanation is that the controller expects that the memory is as fast as it itself but we don't have memory yet running at a 2 GHz and beyond clock rate. What it comes down to is that, depending on the BIOS tuning (by the manufacturer), the read command is issued a bit too early and violates the tRCD. The workaround is to add one additional cycle to the actual memory latency, that is, a 2:2:2-rated DIMM will run stable at 2:3:2; a 2:3:2 rated module will, most likely, need to be set to 2:4:2 in order to run stable (where the latencies are CAS:tRCD:tRP). Bottom line is that when shopping for memory for the Athlon64, it is highly advisable and in most cases mandatory to only consider memory with a tRCD of 2 or, to make it simple, memory that is rated at 2:2:2 a the designated speed. Alternatively, registered DIMMs can be used as well without problems, even though they are slightly more expensive and will give lower performance.
One of the most critical issues in this resepect relates to the first step in the memory access sequence, that is, the opening of a memory page and subsequent steps until a read command can be issued. All of these steps fall into the time frame referred to as RAS-to-CAS delay or tRCD. In a typical system, the memory controller on the chipset is running at bus speed. On the Athlon64, it is running at CPU speed, which means that it is A LOT faster. That means that the internal sequence of commands also goes faster, and that's where all systems are getting hiccups if the memory chips used can only run at a high tRCD.
A very trivial explanation is that the controller expects that the memory is as fast as it itself but we don't have memory yet running at a 2 GHz and beyond clock rate. What it comes down to is that, depending on the BIOS tuning (by the manufacturer), the read command is issued a bit too early and violates the tRCD. The workaround is to add one additional cycle to the actual memory latency, that is, a 2:2:2-rated DIMM will run stable at 2:3:2; a 2:3:2 rated module will, most likely, need to be set to 2:4:2 in order to run stable (where the latencies are CAS:tRCD:tRP). Bottom line is that when shopping for memory for the Athlon64, it is highly advisable and in most cases mandatory to only consider memory with a tRCD of 2 or, to make it simple, memory that is rated at 2:2:2 a the designated speed. Alternatively, registered DIMMs can be used as well without problems, even though they are slightly more expensive and will give lower performance.
Comment