NVIDIA PLAYS CATCH UP WITH UNIFIED MEMORY EMORY
A signi cant performance bottleneck in GPU computing has alwaysays been the act of moving data between main memory systems and graphics memory. N IDIA’s existing GPU technology, including the powerful eplerer architecture found on the GTX Titan, doesn’t allow the graphics processing unitnit to access system memory directly. Instead, the CPU and GPU store data onn their own respective (and exclusive) memory systems. During computationalal operations, data transfer between the two processing units is carried out via an interconnect channel. This transfer of data involves copying of information as such, this introduces both redundancy as well as performance latency.
Enter Heterogeneous System Architecture (HSA), an initiative driven by powerhouse players such as AMD, ARM, Texas Instruments, Samsung Electronics and ualcomm, which aims to integrate both CPU and GPU memory, as well as eliminate the ine ciencies caused by the
current procedures of sharing data. A quick look at the speci cations of both the Xbox One and the PlayStation 4 show that the devices use something called uni ed memory .
Uni ed virtual memory is a key component of HSA it’s essentially common memory space which can be accessed by both the CPU and GPU. HSA won’t be the sole domain of AMD, as it will also be coming to N IDIA’s product lineup with the upcoming Maxwell GPU, a successor to epler. On the other hand, N IDIA’s roadmap for Maxwell only indicates arrival sometime in 2014.
With both new consoles from Microsoft and Sony arriving by the holiday season of 2013, AMD will have the head start, at least in the technological arena. Of course, by the time N IDIA makes their next-generation graphics cards available, things may change.
Industry chatter seems to indicate that the Maxwell GPU will be based on the newer 14nm fabrication process, obviously leading to more transistors per square millimeter of silicon wafer versus epler. This also means that the Maxwell GPU may be cast on a smaller die size, delivering the same performance as epler, but at much lower power consumption. In short, Maxwell could well be N IDIA’s answer to AMD’s custom x 6 chips for both the Microsoft Xbox One and Sony PlayStation 4.
Not content to rest on its laurels, N IDIA is geared to push the envelope further with its olta GPU. The GPUs will feature a technology called Stacked DRAM, which enables video memory chips to co-exist on the same silicon strata as the olta GPU, thus reducing design complexity and overall graphics card footprints.
N IDIA CEO, Jen-Hsun Huang, speaking at the 2013 GPU Technology Conference in San Jose earlier this year, indicated that, such an arrangement will increase memory bandwidth of the GPU by up to 1TB s. In comparison, olta’s suggested memory bandwidth would be about three and a half times that of the GTX Titan. olta would mean speedier access coupledoupled with more video RAM. Game developers will be able to throw in bigger and more detailediled textures and larger game levels, with quicker load times. However, olta is some ways out,, presumably arriving in 2016 if N DIA’s typical two-year architecture refresh cycle remains thehe same.