Future Memory Technologies
For the systems seminar, this Friday, Nov. 15 (3:30, 149 CN), we will
discuss future memory technologies, and the prognosis for the CPU-memory gap.
I've divided this into three sections: Background,
Near-term technologies,
and Future technologies. I will talk on Background,
and I will then open
up the discussion to issues of near- and future technologies. As preparation,
I'd like each of you to just pick a link beyond the background links (pick
any one), and skim through it to get the flavor. In the following week,
we can pick out some articles to look at in depth.
Memory latency is comprised of 1) address decoding delay, 2) word line delay, 3) bit line sensing delay, and 4) output driving delay.
In more detail (taken from
Ace's Guide to Memory Technology: Part 1):
Basically, the latency of the whole memory (From FSB to DRAM) system
is equal to the sum of:
-
The latency between the FSB and the chipset (+/- 1 clockcycle)
-
The latency between the chipset and the DRAM (+/- 1 clockcycle)
-
The RAS to CAS latency (2-3 clocks, charging the right row)
-
The CAS latency (2-3 clocks, getting the right column)
-
1 cycle to transfer the data.
-
The latency to get this data back from the DRAM output buffer to the CPU
(via the chipset) (+/- 2 clockcycles)
Here are some background links:
Finally, there are some severe limitations of latency and bandwidth on the
bus, itself, transmitting information
between the CPU and RAM to be aware of:
- limitations on bus width:
bus skew --- the wider the bus, the more likely that signals on
one line will not be timed correctly with signals on another line.
- limitations of bus latency and energy dissipation:
the chip must develop
enough charge on each pin before sending out the signal --- as
chips use lower and lower voltages, this process can take longer.
- limitations on bus speed:
when the electricity does not have enough time to travel the length
of the wire during one bus cycle, then the standard electrical
transmission must be modelled as a transmission line (different
voltage levels at different places on the wire during one instant).
Transmission lines are likely to be very difficult to engineer on
a motherboard.
(Also, look at vendors' descriptions: Micron, Samsung, Hynix, Infineon,
Elpida (NEC/Hitachi), Nanya, Toshiba, Mitsubishi.)
- Reducing DRAM Access Latency
-
IDF Fall 2002: Dual Channel DDR, DDR-II, and RDRAM
-
Ace's Guide to Memory Technology: Part 2 (DDR RAM and Rambus RAM (RDRAM))
-
Ace's Guide to Memory Technology: Part 3 (RDRAM, DDR400 and DDR-II)
- Ars Technica RAM Guide (Part 3, DDR DRAM and Rambus RAM)
-
RLDRAM (Reduced Latency DRAM) by Micron and Infineon
(and
alt)
-
FCRAM (Fast Cycle RAM, aka NetRAM, Fujitsu and Toshiba)
-
Ars Technica: MRAM (magnetic RAM)
-
Ars Technica: embedded DRAM
- ESDRAM:
Ramtron/Enhanced Memory Systems, Inc.
Enhanced Memory Systems develops and markets patented EDRAMÒ high performance specialty memories that combine fast DRAM and SRAM on one chip. Enhanced Memory Systems is headquartered in Colorado Springs, Colorado and is a wholly owned subsidiary of Ramtron International Corporation (Nasdaq: RMTR).
-
WCDRAM: A fully associative integrated Cached-DRAM with wide cache lines (1997),
Gershon Kedem, Ram Prasad Koganti
-
Vector IRAM: A VLSI Architecture for Media Processing
-
Embedded DRAM design and architecture for the IBM
0.11-µm ASIC offering
-
Embedded DRAM page with pointers to various projects
-
Galileo Group (Computer Architecture)
-
* Datascalar: A Memory-Centric Approach to Computing
Stefanos Kaxiras, Doug Burger, James R. Goodman
Journal of Systems Architecture, Special Issue on Microprocessor Architecture, 45 (1999) pp. 1001-1022.
-
* Limited Bandwidth to Affect Processor Design (or PDF)
Doug Burger, James R. Goodman, and Alain Kägi.
Invited paper to IEEE Micro, special issue on advanced memory architectures, 17 (6), November/December, 1997.
-
* DataScalar Architectures
Doug Burger, Stefanos Kaxiras, and James R. Goodman.
Proceedings of the 24th International Symposium on Computer Architecture (ISCA), June, 1997.
-
* Memory Bandwidth Limitations of Future Microprocessors
Doug Burger, James R. Goodman, and Alain Kägi.
Proceedings of the 23rd International Symposium on Computer Architecture, May, 1996.
- IRAM
Intelligent RAM: (We predict that over the next decade processors and
memory will be merged onto a single chip. Not only will this narrow or
altogether remove the processor-memory performance gap, it will have
the following additional benefits: provide an ideal building-block
for parallel processing, amortize the costs of fabrication lines,
and better utilize the phenomenal number of transistors that can be
placed on a single chip. Let's dub it an "IRAM", for Intelligent RAM,
since most of transistors on this merged chip will be devoted to memory.)
-
CRAM (Computational RAM, place many SIMD processors
on RAM chip to do limited vector processing on the RAM chip;
This avoids the severe memory bandwidth/latency penalty)
- Note other technologies from link to Embedded DRAM page, above