RMBS- Is Rambus an 800 LB Gorilla or an 8 LB Monkey ?
The need for faster memory was felt when Intel and other PCI controller manufacturer's introduced the 100MHz bus to fully utilize the speed advantages provided by faster CPUs. PC100 Sync Dram had sufficient speed to address this 100Mhz bus (the memory is actually 125Hz or 8.5 ns). The theory is that as CPU's get faster, the system bus need to get faster and hence the memory needs to get faster - this is where RMBS is supposed to fit in. Thus in theory, one would expect the high end systems to adopt faster memory technologies first. In reality the opposite is the case. Let's examine why-
A high end system (such as a high end server), usually has multiple memory slots (and multiple memory banks). A crucial element that affects a system's performance is what is called "memory refreshing". Due to the architectural characteristics of DRAM (and memory modules),each memory cell needs to be periodically refreshed, else it looses track of the data. During this refresh process, the memory cannot be read from or written to thus slowing the system. So what designers try to do is to hide these periodic refresh cycles from the host. They do this with a process called interleaving. Typically, this process requires multiple memory banks and is set up such that if one memory bank is accessed another is refreshed and vice versa, thus hiding the refresh cycles from the host. Using this technique alone, one might obtain "SDRAM like" performance via plain EDO chips. In fact if one looks at the PCI controller roadmap for high end servers, you may find that the servers are the last adopters of new technology (High end Servers were predominantly EDO even when SDRAM was the sweet spot in the market. My theory is that the servers shifted to SDRAM primarily for cost and not performance reasons). The design of a high end server main board with the associated memory sub-system is complex. Thus this segment may not opt to go with RMBS technology as an early to mid adopter, and may rely on mature memory technologies such as SDRAM.
Now comes desktops (which are basically free). A desktop is much smaller than Server and has fewer memory slots. So the advanced interleaving techniques cannot be applied here due to the paucity of memory banks. However, Rambus memory has shown very little system level performance improvement over PC100 SDRAM (3%-5% improvement). This is because over 80% of the memory requests are handled by L1 cache, another 16% by L2 cache, only 4% of the access actually hits the memory. Further, it is my opinion that while RMBS technology provides superior sequential access speeds it does not provide superior random access speeds. Will the price sensitive desktop market be willing to pay 15% premium for a memory technology that provides very little performance gain - probably not.
Now the laptop segment. Not so price sensitive, only one memory slot (may be two). This is a good fit for RMBS. They have recently shown their power consumption to be attractive with respect to SDRAMs.
So effectively from a technology perspective, the TAM for RMBS is limited. One has to obviously consider the influence of Intel. Sure, they can force RMBS on to the consumer by designing PCI controllers capable of addressing RMBS technology only - but they are not likely to do that. First my theory on why Intel supported RMBS. Intel wants to control the price of everything in a box (including memory) and use the CPU as a differntiator. It would be tough for Intel to control some one like Samsung, Micron etc who are Multi Billion $ Corps. So they turn to RMBS - a relatively small company that can be easily controlled by Intel. Thus by supporting RMBS, Intel would indirectly control the memory industry (via RMBS licensing). Unfortunately for Intel, , the relative price premiums of RMBS over traditional memory technologies prompted Intel's PCI controller rivals (such as VIA) to make controllers that addressed the faster but proven SDRAM chips such as the SDRAM PC133. Initially Intel never officially supported it (PC133). There are rumors that they may do so in the future. If official, this could be the kiss of death for RMBS. If Intel does not make PC133 and beyond controllers, they stand to loose a huge share of the PCI controller business to their rivals due to the extreme pricing pressures faced by desktop OEMs.
So where does this leave RMBS - extremely over-valued (even at $60), with an unrealistic TAM. The stock is valued on the basis that RMBS will garner most of desktop market. It is my opinion that a serious re-evaluation is in order. |
