The Yellowstone Project
Rambus is a small company located in Los Latus, California. This company that mainly focuses on intellectual property rights is not large because it does not have its own The factory, but its success with RDRAM memory that year has attracted people's attention since then. This company used to be one of Intel’s main support vendors. RDRAM has been one of Intel’s recommended memory configurations for a long time. However, since RDRAM has not made major technical architecture improvements for a period of time, the efficiency improvement is not obvious. Intel also gave up support for RDRAM memory, and the combination of 850E chipset + PC1066 RDRAM became the last full cooperation between Intel and RAMBUS.
Of course, the failure of RDRAM memory is mainly due to the gradual maturity of DDR memory. Compared with the original SDRAM memory, the physical structure of DDR memory has not changed much, but the performance can be greatly improved. With the continuous increase of DDR memory frequency and the emergence of dual-channel DDR memory solutions, a high-performance, low-cost memory solution has been completely occupied by DDR memory. Although Intel has withdrawn the use of RDRAM memory, some manufacturers are still promoting RDRAM memory. SIS is currently the only mainstream chipset manufacturer that has close cooperation with RAMBUS. Its newly released SiS659 chipset is a new product that uses RDRAM memory. .
In fact, looking at RDRAM memory objectively, this is still a very good memory solution, and its potential is very huge. The current RDRAM can get quite amazing transmission speed even if it only uses 16bit bit width, and keeping the same frequency as the processor bus also makes the system efficiency easier to play. However, the current situation surrounding RAMBUS is indeed very difficult. The high cost and few main partners have made RAMBUS isolated and helpless. If this continues, RAMBUS's prospects are not optimistic. Fortunately, on July 10, RAMBUS, together with Toshiba (Toshiba) and Elpida, launched a new generation of product XDR memory solutions, which are based on the well-known Yellowstone (YellowStone) plan.
The Yellowstone Project is a development plan established at the Rambus Developer Forum 2001 meeting in 2001. The core of the plan is to present RDRAM memory with a brand-new look. The Yellowstone Project took three years from the planning to the formal proposal. , The official launch also took two years.
Technology Prospects
The 8 times data transfer rate of XDR is mainly realized by the following technologies. First of all, the clock signal of the motherboard provides the basic signal of the clock phase-locked loop on the XDR, and the internal clock signal of the chip is increased to 1.6GHz, which is realized by the clock phase-locked loop inside the RDRAM particles. The data is transmitted at both ends of the clock signal at the same time, so it can reach a frequency of 3.2GHz, which is eight times the system clock frequency. In fact, the double clock rate RDRAM was introduced in 1992. At that time, its operating frequency was 256MHz, and the data transmission rate was 500MHz. It transmits the signal through the rise and fall delay of the clock signal, so that the clock frequency is changed. It has doubled, which is the same as the working principle of DDR SDRAM. The quadruple RAMBUS signal mode can use more levels of signals to transmit 4bit data in one clock cycle, which can be understood as the use of 2 times the clock frequency in the RAMBUS particles, and then through the rise and fall delay of the clock signal at the same time The signal is transmitted to achieve 4 times the data transmission rate. Now 8 times the data transmission rate can use 1.6GHz frequency under the system bus frequency of 400MHz, which is equivalent to transmitting 8bit data per clock cycle. In contrast, DDR SDRAM memory can transfer 2 Bits of data in one clock cycle, so it is named "Double Data Rate". According to this naming method, XDR may be called ODR, which is "Octal Data Rate (eight times data transmission)", which is also the key to such high frequency of XDR products.
Current computers seem to have only dual-channel and four-channel memory modes, but the XDR module technology officially announced that it can reach the eight-channel mode (51.2GB/s bandwidth), which can reach higher standards in the future. Such specifications indicate that XDR is not (or not only) facing the PC market, and Rambus also hopes that it can make a difference in the console, network equipment and other markets.
For XDR memory, there are several features that must be pointed out. The first is Differential RSL (Differential RSL transmission): DSRL technology allows XDR to use standard PCB boards and standard impedances, so the termination resistance can be It is set inside the chip, so there is no need for an additional middle junction, which is very important for the development of RDRAM, which can greatly simplify the use and configuration of RDRAM modules, and can reduce the cost of the system. However, DDR motherboards need cost and PCB board space to install termination resistors. And the terminator inside the chip does not need to spend any cost, it can also provide a cleaner signal.
Another benefit of DRSL is ultra-low voltage signal transmission. Its signal threshold voltage is only 0.2V, while the signal threshold voltage of DDR has dropped from TTL 3.3V to SSTL 2.5V, while XDR memory The DRSL is only 0.2V. LVTTL and SSTL are not pure square waves, and the signal amplitudes are different, while RSL and DRSL have very few signal voltages.
However, although the internal working principle has some similarities with DDR memory, the actual use of XDR memory is not as simple as current DDR memory, because the actual external interface says XDR memory is the same as DDR It is completely different. Relatively speaking, XDR memory is similar to RDRAM memory. Both use 16-bit memory controllers, but XDR has made some improvements to achieve higher operating frequencies.
There is an XMC on the left and bottom of the XDR controller. However, the operating frequency of each XDR memory module is 3.2GHz, so XDR memory can bring us 12.8GB/s bandwidth. Each memory channel of XDR needs to have its own termination (Termination) at the end, this design is also the same as RDRAM. XDR memory can include such a terminator in the actual memory module, without the need for a separate terminating module (C-RIMM design), but due to the characteristics of serial, the bus still needs a terminator.
We see here that the biggest difference between XDR memory and RDRAM memory is that XDR has independent data and address/instruction buses. The structure of Rambus requires data to pass through all memory modules, which also causes the high latency of RDRAM (compared to DDR). XDR solves this problem through two independent buses, of which the address/instruction bus still needs to pass through all The memory module, but the data can be directly entered into the corresponding module from the memory controller. Obviously, the simple internal structure of XDR is more reasonable and advanced than DDR3 that was still in the pipeline at that time.
Like RDRAM, XDR also requires an independent frequency generator chip. The frequency generator of RDRAM is the key to whether it can be overclocked on the motherboard, and it also determines the frequency at which RDRAM can work. If the quality of the RDRAM frequency generator is not good enough, it will limit the performance of the memory module. But all of this is not a problem for XDR. Cypress and ICS have already contracted to manufacture frequency generators for XDR.
Although Rambus will provide technical and personnel support when third-party manufacturers design their own XDR memory controllers, it is not easy for those who enter the DDR field to redesign memory controllers. For example, ATI and nVidia, both companies want to design a higher bandwidth graphics memory system. If hundreds of engineers and countless money have been invested in DDR memory controllers, it only takes a little bit to switch to DDR II memory. However, it is not easy to adopt the new XDR. XDR memory is more interesting. The architecture this time is not much different from the actual DDR and DDR II currently used, but XDR still has its own intellectual property rights. Samples of XDR will appear within this year and will be officially promoted in the middle and late next year. Samsung is still the core partner of RAMBUS as before. In addition, Toshiba and Elpida will also appear.
The biggest difference between DDR and XDR is the interface between the memory controller and the actual memory chip. This is not surprising. Rambus has positioned itself as an "interface" company. They claim that mid-range XDR memory is 8 times faster than the current DDR400 memory, and the latest XDR-II memory has reached the speed 16-20 times of DDR667.
Product positioning
At present, XDR can only show its skills on Sony’s PS3 game console. Although the requirements for memory in the network, desktop and workstation fields are getting higher and higher, However, all of the bandwidth we provide is not used in these areas. Both AMD and Intel's latest chipsets can provide 6.4GB/s bandwidth, but we have not found any difference in performance with the 3.2GB/s bandwidth system. In this case, the biggest bright spot of XDR memory, "high bandwidth", has no room for use on mainstream PCs. Of course, this only refers to current computers.
Because of this, Rambus has set the first target of XDR memory in the graphics card market. This is a significant and far-reaching move. In the field of graphics cards, video memory bandwidth is an important obstacle that restricts the performance of graphics cards. Designers need to continuously increase the interface bit width of the video memory to obtain a higher video memory bandwidth, which obviously greatly increases the cost of the video card. If you replace the current 256-bit DDR and DDR II memory with low-cost, low-pin count, and high-bandwidth XDR memory, then its bandwidth can well meet the needs of GPUs.
However, Rambus is not going to put all its hopes on the graphics card. They will plan to enter the high-performance network equipment, mobile PC, supercomputer, etc., and of course the game console market. Rambus has a good performance in the game console market. Both Nintendo (N64) and Sony (PS 2) use Rambus technology. At the same time, Sony has the technology license of Yellowstone. Rambus has a lot to do in this market.
Rambus did not give up on the PC market, just did not position it as the first target, and it is not the best time for XDR to enter. XDR hopes that in the next two or three years, CPU and I/O bandwidth will place higher requirements on memory. DDR and DDR II will not be able to meet their needs, and XDR will likely become a leader in the PC field through adequate preparations. A good choice. New technologies such as PCI-Express, SATA, Ultra640 SCSI, and processors above 5GHz will have better requirements for memory bandwidth. This is the opportunity for XDR.
And Rambus has made plans for the future PC memory modules. Since the previous Rambus memory module is called RIMM, this time the module using XDR memory is called XDIMM. Since XDR is a brand new technology, the technical details of XDIMM have not been disclosed too much, but we can still guess some of them.
From the picture provided by Rambus, the appearance of XDIMM is similar to the current 16-bit RDRAM module. The picture of XDIMM shows that it has the same size and similar interface pin design as RDRAM. The internal design of XDIMM looks more like a 32-bit RDRAM product, except that the "T" (terminator) is located inside the module instead of independently. The address/command bus passes through the XDR module and all XDR chips, and ends at the terminator, but each XDR chip has a direct channel with the memory controller. It is this difference that brings higher efficiency and smaller size Delay.
The XDIMM memory configuration of the PC is also disclosed in the technical document.
Product introduction
Currently, XDR manufacturers mainly include Toshiba, Samsung and Elpida. Toshiba launched engineering samples of XDR DRAMs with a frequency of up to 3.20GHz and a capacity of 512Mb as early as the end of 2003. This was the fastest memory in the world at that time. The sample numbers of this version are TC59YM916AMG32A, TC59YM916AMG32B and TC59YM916AMG32C. After that, Samsung introduced 256Mbit 4GHz XDR DRAM memory in early 2005. In March 2005, Toshiba released the second-generation XDR DRAM sample TC59YM916BKG with a running speed of up to 4.8GHz, with a capacity of 512Mb and an operating cycle of 40ns, while Samsung launched a 6.4GHz product. In September of the same year, Elpida's 512Mbit 3.2GHz XDR DRAM products were officially mass-produced.
Memory technology
In the past, Intel's backed Rambus was defeated in the battle with DDR memory standards. After learning from it, this year we will again use a new generation of XDR memory to compete with the future DDR memory. , And named it XDR2. Will XDR2 with a clock frequency of up to 8GHz pose a threat to DDR3?
[XDR2 Micro-Threaded Architecture]
XDR 2 is the second-generation high-speed memory technology launched by Rambus. XDR2 mainly relies on reducing memory loop interference. Join the original FlexPhase and Micro-Threading memory architecture of the previous generation of XDR to improve performance. Compared with the previous XDR's 6.4GHz clock frequency, the performance of this XDR2 memory has climbed again, enabling it to provide a clock speed of 8GHz.
XDR2 memory has a Micro-Threaded architecture, which is a major driving force for its speed improvement. Since XDR2 focuses on the graphics application field at the beginning of its design, the access operations commonly used in this technical field are not much the same as the main memory on a computer. Because the video memory often accesses some small-capacity discrete data collections, it is necessary to optimize this type of application. XDR2 uses the Micro-Threaded architecture, which can be optimized for this operation. Rambus calls it a micro-threaded architecture.
Because there were only two data channel structures on RDRAM memory before, and each channel was only 8 bits wide. A logic bank of RDRAM is composed of two sub-banks, and each sub-bank is connected to a data channel, so there is a total of 16 bits in width. When the memory is working, the two sub-banks are addressed at the same time and their data is transmitted to the data channel A and data channel B. The RDRAM core must transmit at least 64 bytes of data in a row access interval, and at least 32 bytes of data must be transmitted in a column access interval. However, in the application of graphics cards, such a large granularity often results in a waste of bandwidth, because such a large amount of data is generally not used when accessing a graphics object, which has a great relationship with the characteristics of graphics applications. Faced with such technical shortcomings, the new generation of XDR2 can rely on the Micro-Threaded architecture to make better use of higher bit widths.
[What is the difference between XDR2 and XDR]
The overall architecture of XDR2 and XDR is not much different, the most important manifestation is the different system clocks The frequency and data transmission frequency continue to climb above the related bus speed design.
XDR2 increases the frequency of the system clock from 400MHz of XDR to 500MHz. In addition, on the RQ bus used to transmit addressing and control commands, the transmission frequency is increased from 800MHz to 2GHz, which is the XDR2 system clock. 4 times. On the other hand, the data transmission frequency of XDR2 is increased from 3.2GHz of XDR to 8GHz, which is 16 times the system clock frequency of XDR2, while XDR is 8 times. Therefore, Rambus refers to the XDR2 data transmission technology as the 16-bit HDR data rate. The standard design bit width of XDR2 memory chip is 16bit, and the bit width can also be dynamically adjusted like XDR. According to the transfer rate of each data pin is 8GHz, the data bandwidth of an XDR2 chip is already as high as 16GB/s. Up. In contrast, even if the current fastest GDDR3-800 chip has a bit width of 32bit, the data transfer rate is only 1.6Gbps, and the single-chip transmission bandwidth of 6.4GB/s is only 40% of the XDR2 level. Obviously, both The data transfer rate gap is quite large.
[Technical Features of XDR2]
In addition to the highly concerned Micro-Threaded memory architecture design, RAMBUS in order to ensure the stable operation of the XDR2 system, At the same time, it is necessary to make XDR more efficient, so XDR2 continues to add new auxiliary designs to improve performance after inheriting the advantages of XDR. XDR2 integrates a number of important technologies. In order to reduce edge effects that affect signal quality in the memory subsystem, the unique FlexPhase circuit design on XDR2 can adjust speed, voltage and temperature in real-time operation. XDR2 also has the following technical features:
◎Realize zero refresh occupancy function
When refreshing the previous DRAM, it is used for all logic banks. The same address row is refreshed. When a bank is in working state or a bank is to be accessed, it is easy to conflict with the refresh operation. However, the new XDR2 system adopts a special logic bank design, which can adopt interleaved control to refresh, realize the zero refresh occupancy function, thereby avoiding the occurrence of normal addressing situations that are affected by the refresh operation.
With dynamic bit width adjustment function
XDR2 is similar to XDR memory and has the function of dynamic bit width adjustment. It can dynamically adjust the bit width of the interface such as 2bit, 4bit, 8bit, etc. The benefit of choice is to increase the flexibility of XDR2 design. When the bit width is changed, the access granularity will change accordingly. The column access granularity of 2bit, 4bit, and 8bit corresponds to 2 bytes, 4 bytes and 8 bytes, respectively, and the bandwidth is also reduced to 2GB/s and 4GB. /s and 8GB/s levels.
◎With flexible phase control adaptive function
When the speed of the memory continues to rise, XDR2's flexible phase (FlexPhase) synchronization circuit will play Function, it can be compensated to deal with the impact of voltage and temperature changes in real time. In the XDR2 system, the function of flexible phase control is completed by XIO, with an accuracy of up to 2.5ps. The flexible phase technology allows the signal to have the ability of data/clock synchronization and self-calibration, so that the design and wiring of peripheral timing tracking becomes very simple, and it helps to improve synchronization and improve bus utilization.
[A strong competitor of GDDR3]
In the application field, the current goal of XDR2 is still to target the high-end graphics card market. Now the high-end graphics cards of the two major graphics chip giants, NVIDIA and ATi, are equipped with GDDR3 video memory, and the introduction of XDR2 has become the main competitor of GDDR3 in this field. As far as a single chip is concerned, the current highest-speed GDDR3 bandwidth is only 6.4GB/s, while XDR2 has already reached as high as 16GB/s. In this way, while maintaining the same access efficiency, the performance of XDR2 is at least 2.5 times that of GDDR3, which alone has a great advantage. Obviously, XDR2 has gotten rid of the long-term disadvantage of Rambus memory in random access, which is closely related to the higher the frequency of the DDR architecture, the more delay cycles are required.
In terms of power consumption, although the standard design of GDDR3 is 1.8V, the highest speed GDDR3-800 has climbed to 2.0V, while XDR2-500 can still be maintained at 1.8V, which is similar to the first-generation XDR The memory is flat, which helps to reduce the energy consumption of the graphics card.
Future development
Since Rambus is only a pure-tech intellectual property company, all XDR memory and other related components need to be completed by third-party manufacturers. Currently Rambus has found partners like Samsung, Elpida and Sony, but the absence of Intel has cast some shadow on the future of XDR. Will Intel add XDR memory to future product designs? No one can answer now. Obviously, Intel has worked with Rambus in the past, but it doesn't seem to be very satisfactory. Intel thinks RDRAM is a good product, but it is not used in the right place. If the PC's requirements for memory bandwidth increase rapidly, then we will not be surprised that Intel chooses XDR. Everything needs time to give the final answer.
If Rambus can find a suitable partner, they can fulfill their various commitments on XDR. When they launched 800MHz RDRAM products, they were widely skeptical, but the final product dispelled people's doubts and even launched a higher 1066MHz product. Current XDR products with 3.2GHz and 4.0GHz operating frequencies will face heat dissipation problems, but Rambus is confident in solving such problems and does not believe that any exaggerated heat dissipation system is needed. When XDR enters the production stage, Rambus's partners will be able to use 0.11 micron or 0.10 micron technology, which will make XDR products much better in terms of heat dissipation than the current 0.13 and 0.15 micron process products.
The disadvantage of XDR lies in its high cost, which makes it difficult for it to succeed. In the case of DDR being preconceived, XDR can only work on performance or price.
Rambus has shown through the launch of XDR that they will not let the DDR family "dominate the rivers and lakes". Competition is always a good thing for ordinary users. The more intense the competition between XDR and DDR, the more we will It will benefit from it. It will take time to prove whether DDR or XDR is in charge.