New Intel Chipsets Herald DDR’s Successor
This article originally appeared on CPU Planet.
While Intel and AMD prepare their slates of new CPU releases and transitions from 0.13-micron to 90-nanometer process manufacturing, two other architectural changes await the PC in 2004.
One is the long-awaited arrival of the PCI Express bus as a replacement for today’s PCI and AGP alike; the other involves the world of system memory.
All signs point to the new year being the start of the DDR-2 revolution.
Actually, “evolution” would be a better word; just as double-data-rate (DDR) replaced standard SDRAM over the past two years, DDR-2 is the next step in the chain, anointed by industry support and economics as well as technological advances.
Instead of the usual case of “bigger, better, faster” at work, DDR-2 is more of a slight improvement on current DDR design that brings a combination of higher performance and lower cost, while trying to sidestep a few potholes along the way.
In case your own memory needs refreshing, the move from SDRAM to DDR consisted of doubling the data rate by transferring data on both the rising and falling ends of each clock cycle.
This resulted in a doubling of the base clock speed, with the 133MHz of PC133 SDRAM becoming the 266MHz of DDR266 (a.k.a. PC2100 for its 2.1GB/sec transfer bandwidth).
DDR-2 changes the basic specifications once again, while maintaining backwards compatibility with DDR commands — the new memory shares the same basic design and features, while increasing the actual data fetch rate.
Instead of current DDR’s fetch rate of 2 bits per clock cycle, DDR-2 memory can fetch 4 bits per clock.
This means that, with equivalent core speeds, DDR-2 offers theoretically double the data throughput of DDR or quadruple that of conventional SDRAM. (The key word, however, is “theoretically”; more on that in a moment.)
DDR-2 also lowers the voltage requirement to 1.8V, compared to standard DDR’s 2.5V (some of today’s ultra-high-end DDR modules require up to 2.75V to hit their top-rated speeds).
DDR-2 also incorporates some recent chipset features, such as programmable on-die termination (for better signals at high speeds), along with other memory-timing and -configuration controls.
Clock Speeds And Other Fairy Tales
If DDR worked some sleight-of-hand to achieve its rated speeds, DDR-2 is a master magician.
Today’s mainstream desktops’ DDR400 memory modules perform at an effective 400MHz, transferring data on both ends of the clock with an actual core speed of 200MHz.
That’s quite a feat, but DDR-2/400 doubles the fetch rate to achieve the same theoretical performance while actually running at a mere 100MHz.
We say “theoretical” because the requirement of 4 bits per clock cycle tends to introduce additional latencies; if the DDR-2 memory is transferring fewer than 4 bits of random data per clock, the additional data fetch rate goes unused.
Exactly how this affects performance will depend on the application, but in testing 3D graphics cards, we’ve found that DDR-2 had to be clocked slightly higher than DDR memory to make up for latencies.
As they ramp up for a 2004 release, memory manufacturers seem to be targeting the DDR-2/400 through DDR-2/800 ranges — running, respectively, at 200MHz with 3.2GB/sec of data bandwidth and 400MHz with 6.4GB/sec of bandwidth.
It’s easy to see why memory makers are jumping on the DDR-2 bandwagon, as transitioning today’s DDR400 (PC3200) technology into DDR-2/800 lets them advertise double the bandwidth, without a big increase in production costs.
That also means higher profits, since the faster memory will obviously command a premium price.
That’s a little ironic, since the main benefit of DDR-2 is a lower cost per module — the ability to lower core speeds while maintaining performance.
Today’s low-cost DDR266 would transform into DDR-2/533, while even slower DDR could be converted to DDR-2 for notebook use.
This doubling of the data fetch rate decreases emphasis on the core-speed race now raging (as gaming and overclocking enthusiasts pay extra for elite PC3700, PC4000, and faster modules), and instead creates an environment of increased efficiencies.
The Heat is On, Literally and Figuratively
One issue that continues to plague DDR-2 is heat, and not the good kind. In video-card comparisons, similarly-rated DDR and DDR-2 BGA modules exhibited strikingly different temperatures — while ATI and Nvidia cards sporting standard DDR may be warm to the touch, you want to keep your fingers from getting anywhere near their DDR-2 counterparts’ hefty heat sinks.
This may not yield a direct parallel in larger, more-heat-spreader-surface-area desktop modules, but it does indicate that boosting the data fetch rate produces additional heat.
This has created a virtual requirement for DDR-2 production to hit smaller process technologies in order to remain viable in the desktop arena.
Top-end DDR-2 modules are currently being produced using an 0.10-micron process, with the target for a mass-market launch being 90-nanometer or 0.09-micron.
The smaller process not only facilitates lower heat levels, but allows higher manufacturing yields per wafer, i.e., lower production costs.
Of course, manufacturing improvements that benefit potential DDR-2 production are good news for standard DDR, as well.
The similar core designs of DDR and DDR-2 have spurred a game of cat-and-mouse between the two technologies; memory makers have been working hard to achieve ever-shrinking die sizes, and — with DDR-2 not coming to market as quickly as anticipated a year ago — have shifted these into DDR production instead.
When DDR-2 was first announced, standard DDR was using 0.13-micron process technology, but due to heavy competition, it’s now common to find DDR built on an 0.10-micron line.
Since DDR already runs cooler than DDR-2, the ability to use these “designed with DDR-2 in mind” die sizes has led to increased DDR speeds — decreasing the relative market advantage of DDR-2 even before its mainstream debut.
Again, check out what those gaming enthusiasts are buying: Memory manufacturers expected to release DDR-2/533 as a natural successor to DDR400/PC3200, but now we’ve got 533MHz (PC4200) DDR for sale in the mass market, with overclockers already moving on to DDR550 and DDR566, and talk of 600MHz DDR speeds right around the corner. Technology waits for no man, and that seems doubly true of system memory.
Intel has been the first CPU (as opposed to graphics-processor) vendor to endorse DDR-2: its upcoming “Grantsdale” chipset for the “Prescott” Pentium 4 successor will feature dual-channel DDR-2/533 memory support, which should match up nicely with the proposed step from an 800MHz to 1066MHz front-side bus.
Since Grantsdale will support a new CPU package and higher bus speed, Intel is unlikely to add DDR-2 support to existing Pentium 4 platforms — and moreover, having obviously learned a hard lesson with RDRAM, will offer regular DDR compatibility in Grantsdale as well.
This may just be smart business, but it could also mean Intel is keeping its options open until the real-world DDR-2 verdict is in.
Portable PCs, however, are a segment in which DDR-2 promises to yield significant returns, and Intel is expected to move its Centrino platforms in this direction.
The lower 1.8V requirement and relatively slower core speeds mean DDR-2 modules could produce less heat and use less battery power than standard DDR. The current plan is to institute a dual-channel DDR-2 architecture for notebook memory.
AMD, whose 64-bit processors incorporate the memory controller that Intel leaves in the traditional place on the chipset, is more of a dark horse when it comes to DDR-2.
The company has said it will offer Athlon 64 support sometime in the future, but the question remains of how much work needs to be done at the processor level; DDR-2 is command-compliant with DDR, so depending on the Athlon 64/Opteron core architecture, it could be an easy (or difficult) switch to implement an on-chip DDR-2 controller.
There’s also the option of turning the CPU’s integrated memory controller off and using a secondary chipset for memory support, though the performance benefits of going this route for DDR-2 are questionable.
DDR-2 may be the designated heir to today’s desktop memory, but it brings a great deal of questions to the table.
This is not the usual upgrade, with performance benefits pushing aside other factors; instead, DDR-2 represents a series of the tradeoffs that will need to be fully evaluated.
The prospect of more performance at lower production costs certainly appeals to manufacturers, but real-world latencies, throughput, and heat production are still unknown. Maybe we should all follow Intel’s lead and hedge our bets accordingly.