Topic: Optical Interconnect CPU-DRAM
Introduction
The purpose of this project is going to discuss the topic is "Optical interconnect cpu-dram". The optical interconnect provide the connection through the optical cables. The optical is high technology to transfer data its maximum bandwidth can up to 100Gbit. So can we using optical in cpu-dram? CPU-DRAM is important unit in CPU processing. It wants to influence the performance directly on CPU. In the following we go to discuss more about the optical interconnect cpu-dram.What problem is the interconnect CPU to DRAM at present?
Previous architecture of computer, CPU does not communicate to Memory (DRAM) directly. There has used Front-side bus and Memory bus to Northbridge (memory controller hub) to do the data exchange, which its circuits are PCB. (Figure 1)CPU is faster than Memory, which is bottles neck of architecture. The clock rate is a problem. Although it has a way to solve it, which the DRAM is used FB-DIMM to connect it; the latency problem is also in there. The speed is affected by the clock rate that is physical. FB-DIMM is interface that technology is in the DRAM. The interface allows providing high capacity and high bandwidth. And then, the DRAM clock rate approaches the CPU clock rate, But it has new architecture to be faster than that and it unlimited to solving the clock rate in CPU to DRAM.
(Figure1)
What is OC-DIMM?
The new architecture is improving the PCB according to exist architecture which CO-DIMM is derived from FB-DIMM. That is point to point similar with FB-DIMM, but CO-DIMM is connect CPU directly.The PCB is not a traditional circuit, which change to use optical fibers integrated which is in a standard FR4-stack are used to realize an optical layer as connecting. CPU and DRAM also want to improve or change the adapter or bridge, which is connecting the O-PCB.(Figure1)In fact, there have 32 wavelengths going in southbound direction and 32 for northbound; thereby connect the 8DIMMs or more on the DRAM channel. The O-PCB is cable, which is an optical wire. The bandwidth of wires from 10Gbit/s up to 100Gbit/s, so that, the throughput is so large. Otherwise, the clock rate is not a bottleneck in the architecture.And then, it supports one or more CPUs to connect the DRAM. Throughs the OC-DIMM interface, IO of CPU and DRAM also use optical IO to connect the O-PCB. It is wave technology in there, so it has access protocol as a read/write.The protocol has sending a packet is per DRAM clock rate ,and multiple wavelengths are grouped to talking between CPUs and DRAM from read/write bus.(Figure 2)
(Figure 2)
What is better in comparison with new technology and tradition technology?
OCDIMM is deriver from FBDIMM. We can know more OCDIMM if we understand FBDIMM how it work.FBDIMM is a memory technology which can be used to increase reliability and density of memory systems. Each DIMM is connected through a point-to-point pair of differential links and advanced memory buffer (AMB). The serial interface has two buses Northbound and Southbound. Northbound is used for reading data from DIMMs and Southbound is used for sending DRAM commands and data to DIMMS. The first generation FBDIMM link is being specified to operate from 3.2 to 4.8 Gb/s.
But FBDIMM have a weak point. It is not scalable beyond 8 DIMMs per channel. Because we add more DIMMs on a channel the latency increases significantly.
FBDIMM memory controller will add a fixed delay before reading data from AMB. This allows time for previous transactions to complete before starting the next one transaction.
In OCDIMM, the Northbound and Southbound part are replaced by optical fibers. It can carry the data by 10Gps or higher, and this is FBDIMM can’t capable.
How does performance of the optical interconnect CPU into DRAM?
Before we talk about the performance of the optical interconnect CPU to DRAM, we want to know what is Optically Connected Dual In-line Memory Module(OCDIMM), OCDIMM is uses multi wavelength optical interconnect technology. It is use to transfer data CPU and memory.
(Figure 3)
This figure(Figure 2) shows when change the write transaction percentage, bandwidth speed also will change. When we increase write transactions will reduce total bandwidth. We use the DDR2-677Mbps to practice on a single channel. In case of 0% and 100% write transaction, realistic a bandwidth of OCDIMM is 7.4GB/s and 4GB/s.
We can see that, When the Write transactions is 20%, the bandwidth is faster.
(Figure 4)
This figure(Figure 3) shows the relation between Number of DIMM on a channel and Bandwidth. In case of number of DIMM on a channel from 1 to 8, bandwidth speed will increase from 4.7GB/s to 3.9GB/s, but number of DIMM on channel from 8 to 32, the bandwidth speed is obviously drop down from 7.9GB/s to 5.4GB/s.
We can see that, when number of DIMM on a channel is 8, the bandwidth is faster.
Conclusion
In conclusion, we discuss a lot of questions it can be seen then that differences between optical and tradition technology. This can be seen in the performance, which can be more effectively in CPU processing. In the feature, the technology of optical interconnect is ripe that it can be using in difference ways.Reference
- http://sierra.ece.ucdavis.edu/Papers/2008/Yoo_HOTI2008_DRAM_WDM.pdf
- http://www.cs.ucdavis.edu/research/tech-reports/2010/CSE-2010-12.pdf
- http://en.wikipedia.org/wiki/Fully_Buffered_DIMM
- http://en.wikipedia.org/wiki/Optical_fiber
- http://www.cs.ucdavis.edu/research/tech-reports/2010/CSE-2010-12.pdf
- http://www.jedec.org/standards-documents/docs/jesd-8-18a
- http://www.minatec.org/nocs2010/Presentations/NOCS2010%20-%20NA8%20-%2003%20-%20Vincent%20Meija%20-%20Performance%20Evaluation%20of%20Multicore%20System%20with%20Optically%20Connected%20memory%20Modules.pdf
- http://lib.semi.ac.cn:8080/tsh/dzzy/wsqk/SPIE/vol4942/4942-269.pdf
- http://john.mer.utexas.edu/A235FB325D345AC9F0E/OFC2011/2011%20OFC/Conference%20CD/files/OThV4.pdf
- http://spectrum.ieee.org/semiconductors/optoelectronics/get-on-the-optical-bus
