NVIDIA PLAYS CATCH UP WITH UNI­FIED MEM­ORY EMORY

HWM (Malaysia) - - LAB TEST -

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.

Newspapers in English

Newspapers from Malaysia

© PressReader. All rights reserved.