Top 3 Products & Services
Dated: Aug. 13, 2004
Related CategoriesRAM Memory
RAM continues to grow up, even as the price goes down.
Computer memory has come a long way. The drive toward better performance started with PC66 SDRAM in the mid-1990s, and really heated up when the competing double data rate (DDR) and Rambus DRAM (RDRAM) standards delivered multiple RAM controller designs and bus speeds. The industry is now poised to make significant steps forward in RAM performance on the desktop, on video cards, and in embedded devices.
Companies have continued to offer far more RAM for far less cash each year. On average, the cost of a megabyte of RAM drops approximately 40 percent each year, according to Desi Rhoden, president and CEO of industry consortium Advanced Memory International and chairman of the JEDEC JC-42 memory committee. "The historical year-over-year, price-per-bit decline from 1978 -2002 is 36 percent," Rhoden said. "In 2001, there was a 75-percent decline, but that was unusual."
Unlike the CPU and hard-drive markets, which are dominated by a small number of manufacturers, the memory market is extremely competitive, with dozens of vendors selling modules produced by a wide variety of chipmakers. Although there are respected leaders, such as Samsung and Micron, no one company calls the shots in the mainstream dynamic RAM (DRAM) business.
That responsibility falls to JEDEC (Joint Electron Device Engineering Council), an industry standards body founded in 1960. While chipmaker R&D groups work constantly to build faster, denser, and cheaper chips, their concepts are normally ratified for everyday use through JEDEC, which keeps everyone, from the PC chipset makers to the companies that build the RAM sockets, on the same page.
JEDEC gave the world the DDR standard, which is holding steady at 400 MHz (3.2 GB/sec) and is now dubbed DDR1 because of its impending replacement. DDR2 is just around the corner, and it will offer a host of improvements. The new standard should push memory performance wider and faster than ever before.
Rambus' high-speed RDRAM system, which first appeared in 1999, beat the first generation of DDR to market. As the name implies, Rambus' technology lies in the memory interface and bus, which is significantly faster (but narrower) than DDR. Because Rambus, Inc. licenses its designs to chipmakers rather than publishing an open specification, RDRAM has been and will continue to be more expensive on a per-megabyte basis than DDR. According to Kingston Technology, supplier of both DDR and RDRAM, Rambus-based modules sell at about three times the price of standard DRAM.
Intel initially made the Rambus inline memory module (RIMM) format the standard memory component for the Pentium 4. When a customer revolt and pressure from competitors who shipped DDR-based P4 chipsets became too much to bear, Intel straddled the line to offer P4 platforms based on both RDRAM and the more common DDR SDRAM. Now Intel has come full-circle and is phasing out support for RDRAM entirely.
However, that doesn't leave Rambus abandoned in the P4 world. Chipmaker SiS has launched the R658 P4 chipset, which supports the high-end RDRAM offerings.
"Intel has made their last Rambus chipset, but SiS has adopted them, and there's a valid reason for using them," said Dean McCarron, principal analyst at Mercury Research. "It lets you get to higher levels of performance than you can get with the other technologies. But, there's a cost associated with it."
Whether SiS can drive RDRAM back to the desktop is questionable. While in the past third-party chipset makers effected industry change by leading the charge, it often occurred in more consumer-friendly areas of the price/performance curve. SiS is currently developing the SiS659 chip -set, which will offer quad-channel (four-controller) RDRAM. If done properly, this introduces a great performance boost at the RAM controller level without changing the speed of the memory bus or moving to more expensive memory modules.
Rambus has moved its chips up to 1,200 MHz (delivering 4.8 GB/sec in dual-channel mode), with more speed on the way. Rambus expects Samsung, its flagship manufacturing partner, to ship 1333 -MHz RDRAM in the first half of 2004.
Even as Rambus retools and repositions RDRAM, it has a new memory format in the works. Encouraged by the success of their relationship with Sony in the RDRAM-based Play-Station 2, and interested in the growing market for embedded computing devices, Rambus is heavily pushing Yellowstone, the code name for a new 3.2-GHz RAM format targeting the 6.4- GHz threshold. Potentially, it could reach 100 GB/sec of throughput. Yellowstone will have a prefetch of eight, meaning eight data "symbols" can be accessed at once. Toshiba has already agreed to manufacture the new memory format, and Yellowstone-based RAM should appear in the next iteration of the PlayStation.
The company hopes to capitalize on devices that need peak performance or a smaller chip count, or which are expensive enough that the high cost of the memory will be lost in the noise of a larger purchase-few IT executives haggle RAM prices on $10,000 network routers.
Rambus is also gunning for low-power (1.8V) operation with Yellowstone, and has shrunk the signal voltage swing to just 0.2V. Reducing the variance in power on a bus helps reduce electromagnetic interference on the lines, which improves performance. Yellowstone-based RAM has been test-produced using the .13 micron chipmaking process, but isn't expected to hit the market until at least 2004.
"Yellowstone-based DRAMs are initially targeted for consumer, graphics, and networking, since those are the applications which currently need over 10 GB/sec of memory bandwidth," said Michael Ching, product marketing manager for Rambus.
Despite the challenges Rambus has faced in the Intel-based PC market, some are quite confident of its survival.
"One of the virtues of Rambus is that [the company is] delivering pretty much top performance regardless of how much memory you're using," said McCarron.
DDR2 on the Horizon
The rest of the world is watching the DDR open standards evolve. The official DDR1 standard ends at the current DDR400/PC3200 step, although Kingston said it expects to sell DDR1 at 434 MHz to a niche market of ultra-high-performance gamers. Faster parts than the JEDEC spec exist because specialized buyers like graphics card makers will usually bankroll the development of newer chips, and those are expected to filter into the mainstream in part because of the amount of time it will take to get DDR2 into the market.
DDR2, for which JEDEC has defined standards all the way up to DDR2 533/PC5400, will be physically reminiscent of DDR1, but will come in a 240-pin module with the key (notch) in a different position than in DDR1. There will also be the familiar laptop form factor SO-DIMM, along with mini and micro-DIMM variants for specialized applications. Aside from the key positions and the pin density, the only other major change you may notice is the disappearance of pins from the module's surface. DDR2 chips will use ball-grid array (BGA) pinouts, rather than the gullwing packaging common on most SDRAM and DDRI modules.
"Initially the BGA packages did cost a little bit more, but as they move into volume there is no real significant difference," Rhoden said.
Although DDR2 modules will not be physically compatible with DDRI, electrically the chips are compatible, which creates a smaller learning curve for designing new DDR2 controllers into devices and PC chipsets.
DDR2 increases the memory prefetch to four, which means that modules can read four memory cells at the same time. (DDR1 had a prefetch of two.)
DDR2 also refines power consumption, dropping from DDR1's 2.5V to l.8V. Since voltage is squared in the power consumption equation, any reduction results in a substantial power cut. Of course, increasing speed tends to increase power, so reducing power consumption is necessary to offset the demand for faster RAM. Unlike DDR2 and its new Yellowstone line, Rambus is keeping mainstream RDRAM at the 2.5V level going forward.
Although it sounds attractive to cut power consumption to the bare minimum, technology needs to constantly improve so the memory devices can perform all of their functions within the smaller voltage range. A voltage drop has the potential side effect of slowing down electrical operations unless the distance that the signals must travel gets smaller.
"You have to wait for the geometries on semiconductors to get smaller so that they are fast enough to generate the speed you want at the voltage you want to ship it," Rhoden said. Shrinking transistors make it possible to cut voltage and still increase speed.
The next generation of DDR uses internally terminated memory cells, unlike in DDR1, which used resistors off of the memory die. This not only cuts the costs of the chipmaking process by removing the extra resistors, it also makes the signal stronger by reducing the distance data must travel.
The immediate future for DDR is the dual- channel design, which incorporates two memory controllers on a Northbridge chip that can access memory in 128-bit blocks, improving performance without significant end-user expense. Both Athlon and P4 motherboards already have chipsets that support the dual-channel design. No one has announced a quad- channel solution to answer the SiS659 quad-channel RDRAM design yet. The implementations allow for some, if not all, of the RAM to be used in traditional single - channel mode when economics demand such.
DDR2 on the desktop won't happen this year, and it may be limited to some very high-end boards in late 2004, depending on the supply of DDR2 modules to the marketplace and how quickly chipset manufacturers decide to integrate support.
"DDR2 is a story that's being written, and it really comes down to products being available and for applications to emerge," McCarron said. "Right now, GeForceFX uses DDR2, and that's pretty much it."
Graphics Memory Trends
Whenever specification authors negotiate the details, you can count on the graphics vendors to push the envelope, particularly when memory bandwidth is concerned.
"For system memory, you're worried about availability and latency," said Bryn Young, director of memory sales and operations for NVIDIA. "In the graphics frame buffer, it's all about bandwidth, so making sure you're not hobbled by bandwidth is job number one."
Because the DDR bandwidth path moves in small increments, the GPU vendors have had to encourage memory makers to produce reliable parts for the graphic cards' point-to-point environment. Unlike a system bus, where various resources compete for the attention of RAM, on the video card graphics RAM speaks only to the GPU.
Both ATI and NVIDIA have pushed the RAM on their boards well beyond the range of standard desktop DDR, with NVIDIA using dual controllers to reach the 1-GHz mark on the GeForceFX. "There's a standard DDR2 developed in JEDEC, but NVIDIA's product development could not wait the two years for JEDEC to finalize a spec. We'll go off and do what's best for our architecture," Young said.
Oliver Kiehl, a senior staff engineer at chipmaker Infineon who serves on JEDEC's JC42 committee defining DDR2, agreed in principle with the graphics world's contention that they need more options than the DDR makers provide. "The current DDR2 spec is not specifically optimized for graphics-it's a mainstream technology that can be used in anything," he said.
Young said that NVIDIA works with major RAM manufacturers to deliver chips that will meet the bandwidth requirements of next-generation products. Getting more bandwidth out of DDR2 designs is the goal.
"We've established gigahertz memory out there, so the next thing we deliver has to exceed a gigahertz," said Young. "I want to challenge the memory guys to do a lot better every season."
ATI has arguably gone even further than NVIDIA in pushing for memory standards change.
"We looked at DDR2 and saw ways to improve on it for graphics," said Michael Litt, ATI's senior technology manager. ATI, which was concerned about power and heat problems even at the specified 1.8V level (NVIDIA currently runs its DDR2 at 2.5V for higher performance), created a specification for "GDDR3," which is based on DDR2 design but cuts the power consumed by the internal resistors in the memory cells. The company said it hopes that JEDEC will accept it as a graphics branch of the DDR roadmap, and has pro-vided the documentation to memory manufacturers and vendors. According to Litt, four of the five biggest RAM manufacturers have agreed in principle to produce memory to the GDDR3 specification.
Back to the Future
RAM technologists are constantly improving their product, but developing faster memory is of little value without a reciprocal effort by the rest of the PC component manufacturers. In particular, without the CPU industry staying in lockstep on bandwidth, users don't get the full benefit.
"It's more the front-side bus of the processor that determines the speed of memory a system can take advantage of," Rhoden said.
Currently, information on DDR3 is hard to come by, although work is underway. Kiehl said that the design goals would be finalized in 2004, with the key features locked by 2005, but no one can make afirm prediction of when DDR3 will reach the market. Based on the current roadmaps, DDR3 would take standard desktop memory near or beyond the gigahertz range.
DRAM is in no danger of leaving the traditional silicon semiconductor manufacturing process behind. Scientists are researching esoteric new forms of data storage, such as chemical and organic storage in the form of DNA, which could pack tremendous amounts of data into a tiny area, far more dense than any known semiconductor process. Discovering ways to make easily manufactured and replicated constructs that work at modern computing speeds, in real circuits, remains a challenge.
Expect the gulf between system and graphics RAM to widen. "There's going to be divergence. I don't see how the system memory roadmap is going to stay close to the graphics memory roadmap," Young said. "On system memory, they're excited about going from 333 to 400 MHz; the leaps for graphics are going to be hundreds of megahertz at a time."
Litt said that graphics will continue to push performance, and that the desktop's role in the memory market will diminish overtime. This may shift the thrust of research away from the centrist PC and towards more polarized, specialized applications.
Whatever DDR becomes, and no matter if Rambus succeeds in establishing Yellowstone as the world's leading high-bandwidth memory, change is coming-and it will mean more and faster system memory, more instant storage, and a smaller price tag.
"The most important factor in terms of everyday memory is, and has been, price. And that gets cheaper by the minute," Rhoden said.
Now that you've gotten free know-how on this topic, try to grow your skills even faster with online video training. Then finally, put these skills to the test and make a name for yourself by offering these skills to others by becoming a freelancer. There are literally 2000+ new projects that are posted every single freakin' day, no lie!