Samsung to sample nearly four times faster graphics memory
Next February at the industry's leading symposium on semiconductor technology, Samsung is preparing to present a dissertation and perhaps a sample on a new class of graphics memory with 375% the data rate of today's GDDR3.
Along the way, someone at some point decided to skip a number: It will be called GDDR5, and unless someone else steps forward really soon, Samsung may be the only manufacturer (unless Rambus steps up to the plate) to present not only a technical explanation but also working samples to a solid-state circuits conference in San Francisco in February.
GDDR5's maximum data rate is expected to be 6 gigabits per second, and its perceived throughput rate -- the actual raw data it can produce for a display -- is now projected at 24 GB/sec (gigabytes per second).
That's nearly four times the highest data rate of GDDR3, which isn't even ubiquitous today. The most common first-generation GDDR memory has a 1 Gbps data rate.
Samsung said today it plans to accomplish this while at the same time reducing signaling voltage, and thus power consumption. GDDR5 will operate at 1.5 volts, versus GDDR3 at 1.8 V and first-generation GDDR at 2.0 V.
While the best that marketers can do is promise faster games -- as if games need to be sped up even more -- breakthrough accelerations in graphics memory could actually make feasible newer and better ways of rendering scenes. Since real-time 3D ray tracing is still a distant goal even with today's tremendously more capable graphics cards, engineers have been working on interim means such as global illumination that make it appear to ordinary eyes like ray tracing is actually taking place.
If everyday graphics processing truly were four times faster in 2010 than it is today -- a goal which Samsung today said was truly feasible -- programmers could develop more realistic rendering systems, including some that involve limited or partial ray tracing for select areas. The results could include real-time prismatic effects for surfaces such as semi-transparent or translucent objects, raindrops, and the most difficult single prismatic object of all to model, the living human eye.
Chasing Samsung to this goal is Rambus, which last week at an event in Japan kicked off what it's calling its Terabyte Bandwidth Initiative. Its objective is to drive one terabyte per second of bandwidth in everyday systems by 2010.
But Rambus' secret for accomplishing this involves a clever kind of parallelism: a memory controller that can simultaneously connect to 16 DRAM modules, each of which provides 16 Gbps of bandwidth, and each module producing 4 bytes per cycle. And all this assumes that instead of the double-data-rate processors we're accustomed to today (the "DDR" in "GDDR"), we have modules capable of 32x data rates.
It will be interesting to see whether Rambus submits a paper on this next February.