Ok new topic. Why isn't SRAM used more often


monstercameron

Well-Known Member
Joined
Oct 19, 2010
Messages
1,001
Website
www.thinkteletronics.com
I noticed that sram doesnt get used as often outside of small caches but hasn't manufacturing processes shrunken far enough to allow workable densities -like 256-512Mb chips- and wouldn't this type of memory be more frugal, have higher bandwidth and lower latencies? and wouldn't it be great if any of the major ARM socs supported this memory...or am I missing something?
 
Look at this die shot of Sandy Bridge E:


http://images.anandtech.com/reviews/cpu/intel/SNBE/Core_I7_LGA_2011_Die.jpg


The size is 435mm^2. In the middle is 20MB of L3 cache. It takes up over 100mm^2 on Intel's 32nm process. Some of that is needed on tags instead of actual cache, but not much of it. SRAM is huge compared to DRAM. Even IBM is now using eDRAM for last level cache. Intel is also using integrated eDRAM for its Haswell offerings with the best GPU, not SRAM, and it's around the 128MB range.


You don't get more bandwidth with SRAM than a wide DRAM interface, and when you're looking at stuff that has to go through buses and off chip anyway the latency advantage goes down. Most of the latency in an SDRAM access on a current SoC isn't actually spent on the DRAM chip itself.


Furthermore, DRAM is even denser because Samsung et al have cutting edge processes optimized specifically for it, so it's ahead in manufacturing density. And it enjoys much better economy of scale since the same chips are used in just about everything. How you could entertain the notion of SRAM being more frugal is totally beyond me.


Major SoCs probably do support it though. SRAM is simpler for a memory controller to handle because it doesn't need refresh and the interface is usually asynchronous.
 
Exophase hit it on the head, SRAMs take up a huge amount of space on a die as each storage cell of an SRAM is comprised of a half dozen MOSFETs, while eDRAM takes considerably less space as each storage cell is pretty much just a single FET and and capacitor.


Most of the Custom Logic / ASICs I work with have both SRAM and eDRAM cores on them, but amazingly a good share of them only have SRAM on them.. I take it because it's cheaper to manufacture the FETs than the deep trench process needed to make the capacitors on the eDRAM also SRAM is old trusted and well established technology.
 
Last edited by a moderator:
Back
Top