GPU

PC & Tech Authority - - NEWS -

BACK TO THE FERMI

Nvidia’s lat­est GeForce 384 se­ries driv­ers have done some­thing rather un­ex­pected, of­fer­ing GeForce 400 and 500 se­ries (Fermi based) graph­ics cards a new lease on life by giv­ing them DX12 sup­port for the first time. It is thought that this is more to pro­vide DX12_0 level com­pli­ance sim­ply so they’re fully sup­ported af­ter the Win­dows 10 Cre­ators Up­date more than any­thing else so these parts con­tinue work­ing. So don’t go ex­pect­ing your five year old GPU to get a free speed boost. Just be happy it’s still sup­ported.

NVIDIA NAVI-GATES THE FU­TURE

With large mono­lithic pro­ces­sors get­ting harder and harder to man­u­fac­ture, com­pa­nies are start­ing to look to­wards multi-chip mod­ules (MCMs) to break through the man­u­fac­tur­ing R&D wall stalling the progress of chip mak­ers try­ing to keep up with Moore’s law lev­els of per­for­mance gains ev­ery year.

The prob­lem with sin­gle large mono­lithic chips is that just a sin­gle de­fect can ruin the whole chip, and the more die aerial space used per chip the big­ger the chance there is of them oc­cur­ring, re­duc­ing yields.

The ben­e­fits of MCM be­come quickly ap­par­ent as smaller dies mean higher yields dur­ing pro­duc­tion (equals cheaper prod­ucts) and be­ing able to slap sev­eral chips to­gether al­lows you to ex­ceed die space re­stric­tions of cur­rent sin­gle chip man­u­fac­tur­ing, mean­ing more tran­sis­tors to play with for the pro­ces­sor as a whole, in­creas­ing per­for­mance.

AMD has al­ready started down the MCM path to try and off­set this man­u­fac­tur­ing risk in the form of its Ryzen CPU’s which uses its newly minted In­fin­ity Fab­ric to pro­vide the in­ter­con­nects and data sig­nalling re­quired across the phys­i­cally sep­a­rated chips on the CPU mod­ule.

AMD has long had on its roadmaps (af­ter Vega) a GPU ar­chi­tec­ture co­de­named Navi with the main de­tail about it be­ing that it brings multi GPU scal­a­bil­ity. Only now does this make sense. Essen­tially, it’ll be mul­ti­ple smaller GPU chips glued to­gether on the same pack­age us­ing AMDs In­fin­ity Fab­ric con­nect­ing them all to­gether.

Come to think of it, de­spite AMD’s Zen pro­ces­sors suc­cess and per­for­mance in­crease (over 40%! from last gen), I would say AMD’s In­fin­ity Fab­ric is ac­tu­ally its great­est achieve­ment from Ryzen’s de­vel­op­ment. It’ll be ty­ing to­gether CPU, APU and GPU MCM mod­ules from AMD for many years to come. With­out this strong in­ter­con­nect foun­da­tion, none of which would be pos­si­ble.

Well ap­par­ently Nvidia has started lay­ing the ground­work for its own MCM in­ter­con­nect. With the likes of its mono­lithic Tesla V100 com­ing in at a mas­sive 812mm² (the lim­its of cur­rent sin­gle die man­u­fac­tur­ing) Nvidia is des­per­ate for more die space, and MCM is the only way for­ward.

In a whitepa­per Nvidia dis­cuss man­u­fac­tur­ing easy to make small GPUs mod­ules (GPMs) and in­te­grat­ing them to­gether with high band­width power ef­fi­cient sig­nalling tech­nolo­gies (ah-la In­fin­ity Fab­ric), of­fer­ing a prod­uct that was 45% faster than the largest cur­rently man­u­fac­turable sin­gle chip de­sign and within 10% of a the­o­ret­i­cal (cur­rently un­man­u­fac­turable) mono­lithic de­sign with the same tran­sis­tor count. In­ter­est­ingly it is also 26% faster than a multi GPU sys­tem (read SLI) with the same tran­sis­tor counts.

Fun times ahead. It looks like graph­ics cards will for­ever change in the fu­ture with mul­ti­ple GPUs in­te­grated on the one card. It’ll re­de­fine what it means to have a multi-GPU sys­tem.

Newspapers in English

Newspapers from Australia

© PressReader. All rights reserved.