1 reactions
Is there a DRAM successor coming?
DRAM based memory implementing the largest by capacity level in volatile memory hierarchy has had the same latency basically forever as far as the history of computers based on large integrated circuits is concerned, certainly for about the last 30 years of it. Specifically IIRC it's about 50 ns full end-to-end latency for an access, which is what DRAM-specific CAS of 10 ns and a slew of other partly related delays ends up adding up to.
There is no progress in this particular area whatsoever for many decades, in stark contrast to almost everything in computer building.
The only component I can think of that stagnated even remotely similarly was non-volatile memory access latency during the long lasted HDD era. That ended, however, with the broad adoption of SSDs and NVRAM.
Is there any reason to suspect any kind of replacement is coming to improve the access latency of DRAM within any observable future, within decades, not even years, or it's more likely that we will be stuck with 50 ns last level volatile memory access latency for our lifetimes?
Not that it's a tragedy, caches seem to work incredibly well for CPUs, and even if they don't work that well for GPUs, for the latter at this point in the design space and manufacturing technology DRAM bandwidth seems to be the bigger problem, not the latency. Still, being stuck on latency surely doesn't help any major processor in a modern computer or SoC.
#hardware #technology
DRAM based memory implementing the largest by capacity level in volatile memory hierarchy has had the same latency basically forever as far as the history of computers based on large integrated circuits is concerned, certainly for about the last 30 years of it. Specifically IIRC it's about 50 ns full end-to-end latency for an access, which is what DRAM-specific CAS of 10 ns and a slew of other partly related delays ends up adding up to.
There is no progress in this particular area whatsoever for many decades, in stark contrast to almost everything in computer building.
The only component I can think of that stagnated even remotely similarly was non-volatile memory access latency during the long lasted HDD era. That ended, however, with the broad adoption of SSDs and NVRAM.
Is there any reason to suspect any kind of replacement is coming to improve the access latency of DRAM within any observable future, within decades, not even years, or it's more likely that we will be stuck with 50 ns last level volatile memory access latency for our lifetimes?
Not that it's a tragedy, caches seem to work incredibly well for CPUs, and even if they don't work that well for GPUs, for the latter at this point in the design space and manufacturing technology DRAM bandwidth seems to be the bigger problem, not the latency. Still, being stuck on latency surely doesn't help any major processor in a modern computer or SoC.
#hardware #technology
| 1,000 mlx | |
| 21 | |
| 1 |