Memory Space and Memory Form-Factor Considerations
The type of memory form-factors that are used in the memory subsystem are most often determined by:
1)The physical space and more specifically the mechanical envelope (volume) that will be allocated for the memory.
2)How much latency and speed degradation that can be tolerated by the processor (due to physical distance from the processor to memory).
3)Chassis constraints such as proximity to nearby components, thermal cooling requirements (proximity to fans, use of heat spreaders or heatsinks, liquid cooling, spacing between memory elements etc.),
4)Memory placement (proximity to the processor, vertical or horizontal mount etc..)
5)Ease of access to the memory (for installation, servicing and upgrading).
When using embedded DRAM (eDRAM) memory in SoCs, the criteria that can be used to determine the "on-chip" vs. off-chip DRAM memory distribution is often based on the space available for memory on the SOC for pipelining, caching and low latencies access as well as the associated cost for embedding DRAM into the SoC. Current 90nm embedded DRAM process is CMOS logic-based and incorporates an add-on memory module. The embedded process eliminates the I/O power consumed when interfacing with external DRAM devices, provides a wider bus, lowers material cost and has high-bandwidth. It also consumes less active and standby power than SRAM.
However, embedded memory technology is not at density level which can replace high density DRAM on the motherboard, since embedded memory for processors and SoC is currently less that 100MBytes, while off-chip memory placed on the motherboard often begins in the 1GByte range.
Like processors, memory technology has been following Moores Law, where there is a 2x improvement every 2 years (ie: 2x the density every 2 years). The table below shows the current trend in memory densities:
with ECC using 36 DRAMs
2GB with 128Mx4bit
4GB with 256Mx4bit
16GB with 1Gx4bit
with ECC using 5 DRAMs
128MB with 16Mx16bit
256MB with 32Mx16bit
512MB with 64Mx16bit
1GB with 128Mx16bit
with ECC using 1 DRAM core
8MB with 1Mx72bit
16MB with 2Mx72bit
32MB with 4Mx72bit
64MB with 8Mx72bit
If memory size is an important system criteria, memory modules are preferred over other form-factors. They are available in several JEDEC standard DIMM such as RDIMMs, UDIMMs, SO-DIMMs, Mini-DIMMs, MicoDIMMs and FBDIMMs.
Listed below, are some of the major reasons why DRAM memory modules are used over other types of memory form-factors:
Lowers the Total Cost of Ownership (TCO)
a) Memory in modular form can be purchased to arrive just-in-time (JIT) when
needed for final assembly and test, just before systems ship. This would lower
inventory carrying cost and free up cash. It can also eliminates the exposure of
memory component inventory to price declines (a financial loss) during times of
DRAM price volatility. If the memory for the system is chip-on-board (COB), the
individual memory components may need to be stocked longer in inventory to
allow time to manufacture the motherboards.
b) Systems can be initially sold/purchased at a lower system cost using a
minimum memory configuration (leaving a few empty sockets left to spare for
future upgrades). The end-user, can then add more memory if required for new or
larger software application(s) or to improve the speed of the computer, simply by
adding larger memory modules or by populating the empty memory sockets.
c) Since memory can also be an expensive part of the computer system cost,
memory modules can be removed and reused in other systems later, when the
original system is no longer operational or has been retired from service.
Common uses for older memory are:
i) upgradeability and improved performance for other systems still in service
ii) spare parts to maintain older systems still in service
iii) Older memory can be sold for cash (on web brokering service or for trade-
in discounts on new purchases)
Provides Higher System Densities
Modular memory provides more memory then discrete monolithic DRAMs that
are mounted to directly to the mainboard. Systems that require the maximum
memory capacity can use memory module that is very compact using various
stacking techniques which increase density per volume. DIMM socket orientation
can also provide memory placement over other motherboard components to save
Memory may fail overtime creating SBE and multi-bit errors that prevent the
computer system from operating or which slows the system down because of ECC
handling software overhead. This memory can be replaced quickly from its socket
if it was in module form thereby reducing system down time for troubleshooting
and repair. It is also difficult to troubleshoot failing memory components if they
are mounted directly to the mainboard. Rework of memory on the mainboard can
also compromise its quality. If the mainboard requires built in memory test
programs to diagnose the failing chip, that memory testing expertise may not be
available to system designers or BIOS programmers as would be better handled
by memory test programs designed specifically to test memory modules.
Enhanced Testability and System Compatibility
a) If a system has memory directly mount to the mainboard, and it fails, the
processor may not be able to run diagnostics programs to find the root cause of
the error. Memory is a bus oriented operation with the memory controller (or
chipset), and acts as a group of components, thus any timing skews from
mismatched DRAMs not screened as a group (as in a module) could lead to
erratic, unpredictable or intermittent operation. Also, test probing DDR2 memory
which is only available in BGA packages may be not possible.
b) System motherboards often use unique design topologies and memory controllers
which may push the limits on the AC and DC parameters of the DRAM. Memory that
is in modular form can be pre-tested at the system level using system software to filter
out memory that is not compatible with these systems.
If the mainboard in not lead-free and the chipset is for DDR2 memory, then you
may not be able to assemble lead-free RoHS DDR2 DRAMs onto a non-lead-free
mainboard with a high degree of solder joint reliability.
The table below highlights some of the differences for discreet memory chips on the motherboard (a.k.a. Chip-on-board, COB) and socketed memory modules and embedded memory on an SoC.
CHIP ON BOARD (COB)
Least board space efficient
Less board space efficient
Most board space efficient
Populate board with multiple components
Simple socket interface, module assemble by memory supplier
Simple component, complex placement/rework process
Board layout complexity
More complex due to multiple components
Straightforward, simple layout for motherboard
Multiple motherboard layer layout
Most vibration proof
Least rugged (standard module)
Short lead-time, testing done by manufacturer
Shorter lead-time, module tested by suppliers. More predictable supply
Long lead-time, high cost
Most impedance control
Possible impedance mismatch due to memory connector socket
Least impedance control
Field failure support
Whole motherboard replacement required
Easy on-site module replacement
Whole motherboard replacement required
Manufacturing process of memory
Uses Finished/tested package component from semiconductor
Uses Finished/tested package component from semiconductor, assembled onto a module
RAMpedia is a DRAM memory encyclopedia and reference tool
to help computer hardware board designers with DRAM memory subsystems.
This information is provided as a courtesy by Virtium Technology. www.virtium.com