I wrote a paper in ISCA with two collaborators, in which we investigate the possibility of performing computation with STT-MRAM memory technologies where WRITE OPERATIONS are very expensive (5x more expensive than DRAM, both speed, and energy-wise) , and READ operations are a lot less expensive then DRAM (2x less expensive). The density of STT-MRAM can be 3x higher than DRAM, thereby potentially giving you a lot less ENERGY-HUNGRY RAM, that has 3x higher capacity.
This memory is being heavily researched and is expected to be mainstream within 3-5 years. Here is my question: If STT-MRAM was the dominant RAM in the market, this changes the way we view data manipulation. Currently, the fact that READ and WRITE operations are the same speed, and same energy, and same (everything) makes us ignore when we are reading and when we are writing. However, if WRITE operations are more expensive, things change ... Check this out:
Look at the code fragment:
d=b+c;
d=d*q;
A typical way for a compiler to arrange this code is to perform the addition (TWO READ operations from memory), and WRITE into d first, and read fromd, and read from q (THIRD READ) and perform the second multiplication operation, and write back to d.
This is THREE READ and TWO WRITE operations.
Instead, if the variables are small enough, I could do:
d=LOOKUP_ADD_MULT(b,c,q).
The look up table (LUT) is a very large (1MB, 16MB, or more) and allows me to use massive storage areas to eliminate WRITE operations. Notice, I only have a single write operation now. What happened is that, I am storing "pre-computed results" into a LUT and there is no reason for the intermediate write operation now ...
NOTE: everything is in MEMORY, since this code is a part of a large read of a 1Giga-Entry array. This simple example shows how READ+READ+WRITE+WRITE could be turned into LOOKUP+WRITE, saving energy.
Also NOTE: LUTs are the heart of FPGAs.
Do you see this type of memory eventually taking over from DRAM ? or, is this too non-traditional and will never catch on ? Can you conceptualize a programming environment where READ operations are 80% of all arithmetic operations, and WRITE's are 20% by re-working the code to use massive LUTs?
Hello,
I did never program FPGAs, so maybe I do not get the point with your example. For me it looks like every compiler should avoid the write and keep d in a register. Even if the compiler does not keep it in a register, it should be in L1 cache and must not be written to memory twice, does it? So for me your LUT sounds like a software-managed cache as it is (sometimes) used in GPGPUs. Could you explain if this is what you ment or what is the difference?
Best Regards,
Dirk
The trick is the INSTRUCTION SET ARCHITECTURE (ISA). Some ISA's allow three-operand instructions (e.g., MIPS), some don't (early x86). For example, a=a+b would be a two-operand instruction, whereas a=b+c would be a three-operand. So, the compiler can only work with the ISA it has. On top of that, something has to be eventually read from memory, and written to memory. LUTs are always READs from memory ... The question is : can you reduce the number of times you have to access the memory for WRITE operations ? at the expense of increasing the LUT-based READs ? This will require the compiler to be conscious of the READ/WRITE ratio ... whereas, it really didn't matter before. All that mattered is the number of MEMORY ACCESSes. But, now, we are making a huge distinction between MEMORY READs AND MEMORY WRITEs.
Hi there! My research areas are the telecommunication systems and the modern networks (both wired and wireless), and i can say that this kind of memory will have a huge impact on the active networking devices. For example - the routers. They have huge routing table, which are mainly use for route selection. Once the network converges the routing tables are very rarely modified (if the network topology changes) so WRITEs are rarely performed. On the opposite side the router reads the data from the routing table as soon as it receives a packet (and in the modern networks this is millions of times per second), performs some comparisons and then selects either to forward the packet using the corresponding port or to simply delete the packet. In this sense the more expensive operations (the WRITEs) will be performed only during the first several minutes of operation time of the devices and rarely when a change is introduced in the network, while the READs will be performed millions of times per second for the rest of the exploitation time of the devices, which can be up to years (if there are no power surges or malfunctions). On a second thought i can think of several other application areas of this kind of memory in the area of the telecommunication systems and networks - in WSNs, as a potential replacement of NVRAM for network devices, for security purposes, and many, many other!!
Plamen , this is good ! Now, remember one more thing, to make your points even more exciting, this type of ram is actually Non-volatile ! The question is : can you take advantage of GIGABYTES, or, do you only need MEGABYTES ?
In this case the option to replace the NVRAM is out of the discussion.
The size of the memory is always application related. Some devices, like the routers and the switches, will operate fine with just 64MB or 128MB (although they are also application dependent and might require 256MB, 512MB or more), while other might need Gigabytes to work. One more idea - replacement for BIOS or boot-up partitions (battery powered). You upload the data only once, but you read it every time the device boots up (and since it is battery powered you don't need to worry about anything - there are many possibilities if you use LiPo or LiFe batteries and provide a way to recharge them) . Even better application - as replacement for the storage drives in the SANs and the CDNs (if you can provide a reliable UPS you will need even TERABYTES here). A user will upload a file on the server (like audio, video, etc.) and it might be requested and used by millions of users. Imagine YouTube servers with this kind of memory as a storage - a user uploads a video file in his/hers channel and then millions of users request (READ) and play the video.
Ok, the next question becomes: how important is the ENERGY CONSUMPTION OF THE MEMORY in the application you are suggesting. BEcause, there are two advantages of this memory : 1) LOW ENERGY CONSUMPTION, 2) AVAILABILITY OF MORE MEMORY
- Badges
- Science topic
- Similar topics
- Engineering
- Computer Engineering