Elmer, RE: "Off die L2 slow and narrow"
I found this on the RMBS board.
To: +REH (4255 )
From: +REH Thursday, May 21 1998 8:28AM ET
Reply # of 12237
I found this article from Windows Sources very informative - it's from last month but I doubt all has seen it so here it is:
With all the advances in microprocessors, it's easy to forget that memory is as important to overall system performance as any other component. As a matter of fact, it may be the most important component of all. Nothing can shackle a fast CPU faster than a poor memory subsystem. Despite this fact, memory technologies have not advanced nearly as rapidly as microprocessors.
During the course of the PC's development, overall memory performance has merely doubled, while CPU performance has increased by a factor of 100. The PC era has also witnessed six generations of processor families--8086, 286, 386, 486, P5 (Pentium), and P6 (Pentium Pro and Pentium II). But the family tree of DRAM technology for the same period is less expansive--Fast Page Mode, EDO, and Synchronous DRAM.
Cache Advance
To compensate for the memory-performance deficit, developers have used cache memory on the CPU itself (Level 1) and in the system (Level 2). Thus, every Intel CPU since the 486 has had a small amount of L1 cache built right into the chip. For example, the 486 had 8K, the Pentium and Pentium Pro have 16K, and the Pentium/MMX and Pentium II all have 32K of L1 cache.
The amount of Level 2 cache in systems has increased as well, from 32K to 64K in the era of the 386 and 486 to 512K in today's typical systems. In fact, it's not unusual to see caches of 1MB or greater in high-end servers and workstations.
Having cache memory on the processor and in the system compensates for the performance limitations of the slower DRAM. Though the quantity of cache memory is small, it's of a type (SRAM, or Static RAM) that is inherently faster than even the fastest DRAM, partly because SRAM doesn't need a refresh cycle. (A refresh cycle is the time it takes to electrically charge memory to avoid data loss; during this cycle, data can't be accessed from memory.)
Although the use of cache memory has bought the computer industry some time, putting memory on CPUs and using SRAM is expensive, and you can only use so much. Ultimately, the DRAM used for system memory needs to get significantly faster.
Most systems today use synchronous dynamic RAM (SDRAM). The main difference between SDRAM and its predecessors is that it is synchronized with the system clock (the same one that governs the speed of the processor). This helps reduce the time it takes to retrieve data from memory by eliminating wait states (which occur when the processor is ready for data but the memory hasn't yet supplied it).
The good news is that improvements in system memory are just around the corner. Soon, you will witness two rounds of change--one around the middle of this year and another about 12 months after that.
One improvement that's already occurred is the move from SIMMs (Single In-Line Memory Modules) to DIMMs (Dual In-Line Memory Modules). Most machines shipping today use DIMMs exclusively, though a few motherboards are designed to let you mix and match the two. The main difference between the two types of modules is that SIMMs are 32 bits wide, so a pair is required to interface with a 64-bit processor like the Pentium Pro or Pentium II. DIMMs are 64 bits wide, so you can add them to a system one at a time.
Because 66 MHz is the fastest bus speed typically used to transfer data between system memory and the CPU, virtually every major chip set supports 66-MHz SDRAM. (The type of memory a system supports is a function of the memory controller, which is one of the features of the chip set.) All current Intel processors and most of their competitors' products have 66-MHz bus speeds.
The Price of Progress
In the future, improving system performance by adding RAM could be much more difficult than it is today. Over the last few years, during the transitions from Fast Page Mode DRAM to EDO, and then from EDO to today's version of SDRAM, it was possible for a system to support more than one type of RAM. This was a boon to users, because it offered flexibility. Putting in the newest type of memory always brought better performance, and if you had older memory lying around, chances are you could use it, too (if your system came with 16MB of SDRAM, but you had a spare 48MB of EDO, for example).
That flexibility, unfortunately, is coming to an end. The 100-MHz SDRAM interface, although physically similar, is electrically different from current DIMMs. Therefore, SDRAM designed for 100-MHz systems won't work in 66-MHz machines, and vice versa. The same is true for the Rambus interface--Rambus memory modules (a.k.a. RIMMs) will not be backward-compatible with earlier systems.
How will this impact you? Well from now on you'll have to pay much closer attention to the type of memory you buy, because you'll have to live with your choice for years to come. Compatibility won't be a given.
Speeding the Bus
As you've read in the section on processors, most CPU architectures will be transitioning to 100 MHz this year. This will necessitate a jump to 100-MHz SDRAM. As you might imagine, older 66-MHz SDRAM will be incapable of running at the higher speed, adding yet one more variable to consider when buying memory upgrade modules (see the sidebar "The Price of Progress," earlier in this section).
The move to 100-MHz SDRAM will probably be short-lived, though. That's because even 100-MHz SDRAM, which can move about 100 MBps, won't be able to provide enough bandwidth to satiate the voracious appetite of the new generation of processors.
That's why various new memory architectures are under development. One is SDRAM-DDR (the suffix stands for double data rate). As the name implies, SDRAM-DDR doubles the available bandwidth to 200 MBps. It does this by employing more advanced synchronization and signaling techniques. The SDRAM-DDR design is being championed by JEDEC (Joint Electronic Device Engineering Council), an industry group made up of most major memory manufacturers.
Another industry consortium is proposing a different memory design, called SyncLink. It extends the SDRAM design to quadruple available bandwidth to 400 MBps.
Graphics Memory Moves onto the Chip
As we've noted, memory performance has not kept pace with processor technology, which often leaves the latter idling. One way to increase memory bandwidth is simply to add more of it--going from a 64- to a 128-bit data path, for example. The problem is this method increases both the cost and complexity of the processor and the amount of memory you have to use. One way to solve this problem is to use a narrower, but much faster, interface between the two, as Rambus has done.
But what if you could combine processor and memory? There are already some examples of this concept at work, the most notable of which is the MagicGraph line of mobile graphics controllers from NeoMagic. MagicGraph chips have found their way into a substantial number of notebooks, including models from Dell and Quantex. It's no wonder why: With a 2D/3D graphics accelerator and 1.2MB of display memory integrated on the same piece of silicon and connected by a 128-bit interface, the chip offers greater performance than most conventional solutions. It also takes up less space and consumes far less power.
The drawback is that you can only fit so much memory on an integrated device--although the amount is growing (the MagicGraph used to contain a mere 768K)--so when you need a lot of memory, external RAM chips are still the way to go.
Power consumption, space, and cost are not as much of a concern on desktops. But in devices where those factors count, such as portables, you can expect to see more integration in the future.
The 800-MHz Bus
Notwithstanding these developments, it appears certain that the type of memory that will replace 100-MHz SDRAM in PC systems will be Direct RDRAM. This shift will occur in mid-to-late 1999.
RDRAM memory is not new, though. Designed by Rambus, it's currently used in devices ranging from video games and graphics cards to workstations and Ethernet switches. Direct RDRAM is an enhancement of the original RDRAM specification developed by Rambus and Intel for use as PC main memory.
Why is Direct RDRAM the heir apparent to 100-MHz SDRAM? First, it's the one Intel has chosen. And given the company's increasing dominance as the manufacturer of the building blocks of the PC platform, Intel can certainly influence the type of memory the industry will use going forward. But practically speaking, Intel's support is almost incidental, because Direct RDRAM appears to be the only architecture that can provide the performance necessary for the next generation of processors. Equally important, it offers enough headroom to support several more years of processor innovation. Here's how.
Direct RDRAM uses a relatively narrow interface, only 8 bits wide, called the Rambus Channel. However, the interface runs at the extremely high clock frequency of 800 MHz. Direct RDRAM uses two of these channels, yielding a whopping available bandwidth of 1.6 GBps. Adding more channels can provide still more bandwidth.
So that the narrow Rambus memory can interface with the wide 64-bit data path of PC processors, a new type of memory module is under development--RIMM, or Rambus In-Line Memory Module.
RIMMs have the same physical dimensions as DIMMs, which means vendors can more easily update motherboard design to support Direct RDRAM memory. However, the two devices are not compatible, and vendors will likely prevent users from interchanging the two (probably by means of a notch key).
Intel says that the chip sets it will ship in 1999 will support Direct RDRAM memory. What's not yet clear is whether the chip set will support both RDRAM and SDRAM memory in the same system--which would let vendors build systems that include one or the other type of memory. At press time, Intel would not say. Our guess is that Intel's chip set won't support both memory types, because it would make the chip sets too large and too expensive.
|
