Ik moest ook weer even opnieuw kijken wat nou precies de verschillen zijn.
Hier is een video die het uitlegt:
https://www.youtube.com/watch?v=pbgvzVgfoSc
De reactie van
matsv201 onder de video is ook wel informatief:
There is some non correct statements in this video. GDDR2 and DDR2 uses basically the same structure, well they use the same basic modulation metod, but the GDDR2 is a port memory (channel terminator) but the DDR2 is a buss memory (non channel terminating). This is also true about GDDR3 and DDR3. But GDDR4 and GDDR5 uses the same modulation method (one generation newer than GDDR3) both uses the same. The difference between 4 and 5 is not the modulation but accessing and caching. DDR4 uses the same modulation and a similar caching/accessing method of GDDR5, so GDDR5 and DDR4 is quite similar. Also, DDR4 is the first (non G) DDR memory using channel termination. Its a ported technology. So DDR4 and GDDR5 is quite simular. There is some difference in addressing and GDDR5 is made for soldered (BGA) mounting only. You can´t use (high speed) GDDR5 in a socket, that's the main reason why its not available for CPU, because on a CPU you want to buy separate memory, but in a console, it does not matter (why PS4 uses GDDR5 and Xbox 360 uses GDDR3).
Is a common miss understanding that GDDR5 have slower latency than DDR3 (as falsely stated in the video). GDDR5 have a higher CAS latency, but that's because it have a higher transfer rate. CAS latency is invert related to the transfer-rate. DDR3 and GDDR5 have i practical application the exact same time period latency.
The reason is that GDDR 1-5 and DDR 1-3 uses the exact same type of ram. Every single chip is a SRAM modul. The speed an CAS is depended on the number of SRAM modules that is multiplexed inside the chip. The SRAM memory is not getting any faster or slower regardless of if its mounted inside a GDDR or DDR. The latency is always the same (for the same quality and line-width). The speed of the memory is increased by having a higher number of multiplexed unit. This increases the CAS time, but NOT the REAL time.
In the olden days people often upgraded there memory's when the memory got cheaper. Buying two modules up front, and adding 2 more later on. To make this possible you have to use a bus solution. A bus solution consumes a lot of power and is slower than a port solution. In a modern computer its only the memory uses a bus solution, witch is ironic because that memory uses the highest bandwidth in total. A DDR3 uses 64 pins + addressing comparied with a PCI-E GPU that uses 8 or 16 pins in total. Memory consumes a lot of pins in the processor socket (and chip) in term dictating a minimum die size of the chip. (in the K7 AMD solved this by making the chip realty rectangular). A additional reason why bus memory was prefered was that a bus DIMM can have 2 or more stacks of memory chip on a single DIMM. Some higher capacity DIMM have 1 stack on both side, and some might have 2 stacks on both side. Making possible for a PC of having up to 16 stacks of modules on one single Chanel. With port memory its only possible to have one stack per channel. Now days its no problem in making a 2GB chip or in same cases 4GB chip, thats why the PS4 can have 8 GB of GDDR because it uses brand new higher capacity GDDR modules.
DDR4 solves a lot of this problem, both lowering the number of pins needed and raising speed (witch will enable CPU to continue using a 128 bit memory system in the future), also enabling cheaper processors of using a 64 bit system with good performance. Thow DDR4 still uses socket, the maximum frequency is limited, so the performance will not match GDDR5, but it will be quite close, aproximlty 20% slower. We probably have to wait for ReRAM or MRAM to get a on Chip RAM memory (for PC, in phones it already exist)
[Reactie gewijzigd door Berlinetta op 1 augustus 2024 22:13]