HPC
What's the Difference Between GDDR and DDR Memory?
September 21, 2023
3 min read
What Does GDDR and DDR Stand For?
DDR stands for double data rate and GDDR stands for graphics double data rate. But to be clear, DDR and GDDR architectures do not share the same standards like speed, bandwidth, etc. We can classify them as completely different technologies as DDR is developed for CPUs (and CISC) and GDDR is developed for graphics (and RISC). The full name GDDR6 SDRAM stands for graphics double data rate type 6, synchronous dynamic random-access memory where SDRAM is the foundation of the memory used in desktops (DDR4 and DDR5), laptops (LPDDR4 and LPDDR5) and graphics processing units (GDDR6 and also GDDR6x in NVIDIA GPUs).
GDDR - Bus and Bandwidth
GDDR differs from DDR in mainly the memory bus size and bandwidth. GDDR is memory optimized for bandwidth utilized by modern graphics cards as GDDR6 and GDDR6X found only in high-end consumer NVIDIA GPUs like the 4090. GPUs process larger amounts of repetitive and linear workloads like math calculations and deliver data in parallel; a larger bit bus and higher bandwidth are required to prevent a bottleneck. Otherwise, our visual assets would not load, or data-dependent AI training would have to wait for model parameters to update.
DDR - Speed vs Latency
DDR memory is not slower per se, but the trades off quite simple: DDR memory prioritizes lower latency as opposed to bandwidth. While still capable of high speeds, DDR’s lower latency is important for executing CPU tasks that require fast access to memory such as running applications, managing the OS, executing instructions, and instructing the GPU what things to calculate.
DDR5, the newest standard for CPU memory, has data transfer rates reaching from 4800 to 8400 MT/s (mega transfers) usually denoted after the RAM specification i.e.: DDR5 6400MT/s CL32. We want to focus on the CL which tells us the timing of the memory.
CAS Latency (column address strobe latency) refers to the time delay measured in clock cycles between a read command issued to memory and when a data request is available for access. It is one of the more important timing parameters that measures the speed and performance of a module.
For the example above, CL32 means that it will take 32 clock cycles for the module to access data once it has received a command. The lower the CL value such as CL16 and CL15 are generally considered faster and can offer better performance. It is worth noting that while DDR4 RAM DIMMs have CAS Latency options as low as CL16, whereas DDR5’s lowest CAS Latency DIMM is CL30. Not to fret! Because DDR5 has higher frequency/transfer rates, the effective latency is marginally greater.
| DDR Generation | CAS Latency | Frequency | Absolute Latency |
| DDR4 | CL16 | 3200 MT/s | 10ns |
| DDR4 | CL18 | 3600 MT/s | 10ns |
| DDR5 | CL38 | 5200 MT/s | 14.6ns |
| DDR5 | CL40 | 6000 MT/s | 13.3ns |
| DDR5 | CL32 | 6400 MT/s | 10ns |
The Next Generation of Memory
Below is a graph on the past projects on trends where memory technology would be by 2020. Looks like DDR5 and GDDR6 exceed the performance expectations extrapolated!
As of the time of writing, DDR5 is still relatively new, and we have no reason to believe that DDR6 is close to release. However, GDDR6 has been around for approximately 5 years now and Samsung has announced its first GDDR7 DRAM offering in July 2023 with Micron soon to follow in 2024. Samsung’s GDDR7 claims to provide 1.5TB of bandwidth and 32GBps per pin. Exxact is excited to see what this brings to the table for next-generation high-performance GPUs.
Have any questions? Interested in customizing a multi-GPU workstation featuring Threadripper PRO?
Contact Exxact Today to learn more!
HPC
What's the Difference Between GDDR and DDR Memory?
September 21, 20233 min read
What Does GDDR and DDR Stand For?
DDR stands for double data rate and GDDR stands for graphics double data rate. But to be clear, DDR and GDDR architectures do not share the same standards like speed, bandwidth, etc. We can classify them as completely different technologies as DDR is developed for CPUs (and CISC) and GDDR is developed for graphics (and RISC). The full name GDDR6 SDRAM stands for graphics double data rate type 6, synchronous dynamic random-access memory where SDRAM is the foundation of the memory used in desktops (DDR4 and DDR5), laptops (LPDDR4 and LPDDR5) and graphics processing units (GDDR6 and also GDDR6x in NVIDIA GPUs).
GDDR - Bus and Bandwidth
GDDR differs from DDR in mainly the memory bus size and bandwidth. GDDR is memory optimized for bandwidth utilized by modern graphics cards as GDDR6 and GDDR6X found only in high-end consumer NVIDIA GPUs like the 4090. GPUs process larger amounts of repetitive and linear workloads like math calculations and deliver data in parallel; a larger bit bus and higher bandwidth are required to prevent a bottleneck. Otherwise, our visual assets would not load, or data-dependent AI training would have to wait for model parameters to update.
DDR - Speed vs Latency
DDR memory is not slower per se, but the trades off quite simple: DDR memory prioritizes lower latency as opposed to bandwidth. While still capable of high speeds, DDR’s lower latency is important for executing CPU tasks that require fast access to memory such as running applications, managing the OS, executing instructions, and instructing the GPU what things to calculate.
DDR5, the newest standard for CPU memory, has data transfer rates reaching from 4800 to 8400 MT/s (mega transfers) usually denoted after the RAM specification i.e.: DDR5 6400MT/s CL32. We want to focus on the CL which tells us the timing of the memory.
CAS Latency (column address strobe latency) refers to the time delay measured in clock cycles between a read command issued to memory and when a data request is available for access. It is one of the more important timing parameters that measures the speed and performance of a module.
For the example above, CL32 means that it will take 32 clock cycles for the module to access data once it has received a command. The lower the CL value such as CL16 and CL15 are generally considered faster and can offer better performance. It is worth noting that while DDR4 RAM DIMMs have CAS Latency options as low as CL16, whereas DDR5’s lowest CAS Latency DIMM is CL30. Not to fret! Because DDR5 has higher frequency/transfer rates, the effective latency is marginally greater.
| DDR Generation | CAS Latency | Frequency | Absolute Latency |
| DDR4 | CL16 | 3200 MT/s | 10ns |
| DDR4 | CL18 | 3600 MT/s | 10ns |
| DDR5 | CL38 | 5200 MT/s | 14.6ns |
| DDR5 | CL40 | 6000 MT/s | 13.3ns |
| DDR5 | CL32 | 6400 MT/s | 10ns |
The Next Generation of Memory
Below is a graph on the past projects on trends where memory technology would be by 2020. Looks like DDR5 and GDDR6 exceed the performance expectations extrapolated!
As of the time of writing, DDR5 is still relatively new, and we have no reason to believe that DDR6 is close to release. However, GDDR6 has been around for approximately 5 years now and Samsung has announced its first GDDR7 DRAM offering in July 2023 with Micron soon to follow in 2024. Samsung’s GDDR7 claims to provide 1.5TB of bandwidth and 32GBps per pin. Exxact is excited to see what this brings to the table for next-generation high-performance GPUs.
Have any questions? Interested in customizing a multi-GPU workstation featuring Threadripper PRO?
Contact Exxact Today to learn more!
