AMD Radeon Vega

AMD’s lat­est cards bring graph­ics and mem­ory closer to­gether, re­veals Brad Cha­cos

PC Advisor - - CONTENTS -

Wait for Vega.” For the past six months that’s been the mes­sage from the Radeon faith­ful, as Nvidia’s beastly GeForce GTX 1070 and GTX 1080 stomped above AMD’s Radeon RX 400-se­ries graph­ics cards.

While Nvidia’s pow­er­ful new 16nm Pas­cal GPU ar­chi­tec­ture scales all the way from the lowly £150 GTX 1050 to the mighty £1,300 GTX Ti­tan X, AMD’s 14nm Po­laris graph­ics are de­signed for more main­stream video cards, and the flag­ship Radeon RX 480 is no match for Nvidia’s higher-end brawlers. Thus ‘Wait for Vega’ has be­come the ral­ly­ing cry for AMD sup­port­ers with a thirst for face-melting game­play – Vega be­ing the code name of the new en­thu­si­ast-class 14nm Radeon graph­ics ar­chi­tec­ture teased on AMD road maps for early 2017.

Un­for­tu­nately, the wait will con­tinue, as the new ar­chi­tec­ture won’t ap­pear in ship­ping cards un­til some­time later in the first half of 2017. But at CES, Vega is be­com­ing more than a mere co­de­name: AMD is fi­nally re­veal­ing some tech­ni­cal teases for Radeon’s per­for­mance-fo­cused re­sponse to Nvidia’s ti­tans, in­clud­ing how the new GPU in­ter­twines graph­ics per­for­mance and mem­ory ar­chi­tec­tures in rad­i­cal new ways.

Be­fore we dive in too deeply, here’s a high-level over­view of the Vega tech­ni­cal ar­chi­tec­ture pre­view. All those words will be­come mean­ing­ful in time. Let’s start with what you want to hear about first.

1. Speed de­mon

In a pre­view shown to jour­nal­ists and an­a­lysts in De­cem­ber, AMD played 2016’s sub­lime Doom on an early Radeon Vega 10 graph­ics card with ev­ery­thing cranked to Ul­tra at 4K res­o­lu­tion. The game scales like a champ, but that’s hell on any graph­ics card: even the GTX 1080 can’t hit a 60 frames per sec­ond (fps) av­er­age at those

set­tings, per Techspot tinyurl.com/gvt9kxe. Radeon Vega, meanwhile, floated be­tween 60- and 70fps. Sure, it was run­ning Vulkan – a graph­ics API that favours Radeon cards in Doom – rather than DirectX 11. But, hot damn, the demo was im­pres­sive.

A cou­ple of other sight­ings in re­cent months con­firm Vega’s speed. At the New Hori­zon livestream that in­tro­duced AMD’s Ryzen CPU to the world, the com­pany showed Star Wars: Bat­tle­front run­ning on a PC that pairs Ryzen with Vega. The duo maxed out the 4K mon­i­tor’s 60Hz speed with ev­ery­thing cranked to Ul­tra. The GTX 1080, on the other hand, hits just shy of 50fps, Techspot’s test­ing shows (tinyurl.com/hbf6­fad).

Meanwhile, a since-deleted leak in the Ashes of the Sin­gu­lar­ity data­base in early De­cem­ber showed a GPU with the De­vice ID ‘687F:C1’ sur­pass­ing many GTX 1080s in bench­mark re­sults. Here’s the twist: the De­vice ID shown in the frame rate over­lay dur­ing AMD’s re­cent Vega pre­view with Doom con­firmed that Vega 10 is in­deed 687F:C1. Th­ese num­bers come with all sorts of caveats: Vega 10 isn’t in its fi­nal form yet, we don’t know whether the graph­ics card AMD teased is Vega’s beefi­est in­car­na­tion, all three of those bench­marked games heav­ily favour Radeon, and so on.

But all that said, Vega cer­tainly looks com­pet­i­tive on the graph­ics per­for­mance front, partly be­cause AMD de­signed Vega to work smarter, not just harder. “Mov­ing the right data at the right time and work­ing on it the right way,” was a ma­jor goal for the team, ac­cord­ing to Mike Man­tor, an AMD cor­po­rate fel­low fo­cused on graph­ics and par­al­lel com­pute ar­chi­tec­ture – and a large part of that stems from ty­ing graph­ics pro­cess­ing more closely with Vega’s rad­i­cal mem­ory de­sign.

2. All about mem­ory

When it comes to on­board mem­ory, Vega is down­right rev­o­lu­tion­ary – just like its pre­de­ces­sor. AMD’s cur­rent high-end graph­ics cards, the Radeon Fury se­ries, brought cut­ting-edge high-band­width mem­ory to the world. Vega car­ries on the torch with im­proved next-gen HBM2, bol­stered by a new ‘high-band­width cache con­troller’ in­tro­duced by AMD.

Tech­ni­cal lim­i­ta­tions lim­ited the first gen­er­a­tion of HBM to a mere 4GB of ca­pac­ity, which in turn lim­ited the Fury se­ries to 4GB of on­board RAM. Thank­fully, HBM’s raw speed hid that flaw in the vast ma­jor­ity of games, but now HBM2 tosses those shack­les aside. AMD hasn’t of­fi­cially con­firmed Vega’s ca­pac­ity, but the over­lay dur­ing the Doom demo re­vealed that par­tic­u­lar graph­ics card packed 8GB of RAM. And that su­per-fast RAM is get­ting even faster, with AMD’s Joe Macri stat­ing that HBM2 of­fers twice the band­width per pin of HBM1.

But as it turns out, HBM was just the be­gin­ning. “It’s an evo­lu­tion­ary technology we can take through time, make it big­ger, faster, make all th­ese key im­prove­ments,” said Macri, a driv­ing force be­hind HBM’s cre­ation. Vega builds on HBM’s shoul­ders with the in­tro­duc­tion of a new high-band­width cache and high-band­width cache con­troller, which com­bine to form what Radeon boss Raja Ko­duri calls “the world’s most scal­able GPU mem­ory ar­chi­tec­ture”.

AMD crafted Vega’s high-band­width mem­ory ar­chi­tec­ture to help pro­pel mem­ory de­sign for­ward in a world where sheer graph­ics per­for­mance keeps im­prov­ing by leaps and bounds, but mem­ory ca­pac­i­ties and ca­pa­bil­i­ties have re­mained rel­a­tively

static. The HB cache re­places the graph­ics card’s tra­di­tional frame buf­fer, while the HB cache con­troller pro­vides fine-grained con­trol over data and sup­ports a whop­ping 512TB – not gi­ga­bytes, ter­abytes – of vir­tual ad­dress space. Vega’s HBM de­sign can ex­pand graph­ics mem­ory be­yond on­board RAM to a more het­ero­ge­neous mem­ory sys­tem ca­pa­ble of manag­ing sev­eral mem­ory sources at once.

That’s likely to make its big­gest im­pact in pro­fes­sional ap­pli­ca­tions, such as the new Radeon In­stinct line-up or the cut­ting-edge Radeon Pro SSG card that graft high-ca­pac­ity NAND mem­ory di­rectly to its graph­ics pro­ces­sor. “This will al­low us to con­nect ter­abytes of mem­ory to the GPU,” David Wat­ters, AMD’s head of In­dus­try Al­liances, told our col­leagues at PC-World when the Radeon Pro SSG was re­vealed, and this new cache and con­troller ar­chi­tec­ture de­signed for HBM’s high speeds should su­per­charge those ca­pa­bil­i­ties even more.

To drive the po­ten­tial ben­e­fits home, AMD re­vealed a pho­to­re­al­is­tic re­cre­ation of Macri’s home liv­ing room. The 600GB scene nor­mally takes hours to ren­der, but the com­bi­na­tion of Vega’s prow­ess and the new HBM2 ar­chi­tec­ture pumps it out in mere min­utes. AMD even al­lowed jour­nal­ists to move the cam­era around the room in real-time, al­beit some­what slug­gishly. It was an eye-open­ing demo.

Ko­duri stressed that games can also ben­e­fit from the high-band­width cache con­troller’s fine-grained, dy­namic data man­age­ment, cit­ing Witcher 3 and Fall­out 4, each of which ac­tu­ally use less than half the mem­ory al­lo­cated by the games when they’re run­ning at 4K res­o­lu­tion. “And those are well-op­ti­mised games,” he said. Mem­ory de­mands are only get­ting greed­ier in high­pro­file games, and dou­bly so at bleedingedge res­o­lu­tions. Here’s hop­ing that the HB cache’s finer con­trols paired with HBM’s sheer speed al­le­vi­ates that some­what.

AMD also says that fu­ture gen­er­a­tions of games could take ad­van­tage of high-band­width mem­ory de­sign to up­load large data sets di­rectly to the graph­ics pro­ces­sor, rather than han­dling it with a more hands-on ap­proach as done to­day.

3. Ef­fi­cient pipe­line man­age­ment

The way graph­ics cards ren­der games isn’t very ef­fi­cient. Case in point: the scene (shown top right) from Deus Ex: Mankind Di­vided. It packs in a whop­ping 220 mil­lion poly­gons, ac­cord­ing to Ko­duri, but only two mil­lion or so are ac­tu­ally vis­i­ble to the player. En­ter Vega’s new pro­gram­mable ge­om­e­try pipe­line.

Ren­der­ing a scene is a multi-step process, with graph­ics cards pro­cess­ing ver­tex shaders be­fore pass­ing the in­for­ma­tion on to the ge­om­e­try en­gine for ad­di­tional work. Vega speeds things up with the help of prim­i­tive shaders that iden­tify the poly­gons that aren’t vis­i­ble to play­ers so the ge­om­e­try en­gine doesn’t waste time on them. Vega also blazes through in­for­ma­tion at over twice the peak through­put of its pre­de­ces­sors, and in­cludes a new ‘In­tel­li­gent Work­group Dis­trib­u­tor’ to im­prove task load bal­anc­ing from the very be­gin­ning of the pipe­line.

The scene from Deus Ex: Mankind Di­vided. It packs in a whop­ping 220 mil­lion poly­gons, ac­cord­ing to Ko­duri, but only two mil­lion or so are ac­tu­ally vis­i­ble to the player

Th­ese tweaks drive home how AMD’s in­fil­tra­tion in con­soles can ben­e­fit PC gamers, too. The in­spi­ra­tion for the load bal­anc­ing tweaks comes from con­sole de­vel­op­ers used to work­ing “closer to the

metal” than PC de­vel­op­ers, who high­lighted it as a po­ten­tial area for im­prove­ment for AMD, Raja Ko­duri ex­plained.

4. Right task, right time

AMD de­signed Vega to “smartly sched­ule past the work that doesn’t have to be done,” ac­cord­ing to Mike Man­tor. The fi­nal tid­bits made pub­lic by the com­pany drive that home. Vega con­tin­ues AMD’s mul­ti­year push to re­duce mem­ory band­width con­sump­tion. Its next-gen pixel en­gine in­cludes a ‘draw stream bin­ning ras­ter­izer’ that im­proves per­for­mance and saves power by team­ing with the high-band­width cache con­troller to more ef­fi­ciently process a scene. After the ge­om­e­try en­gine per­forms its (al­ready re­duced amount of) work, Vega iden­ti­fies over­lap­ping pix­els that won’t be seen by the user and thus don’t need to be ren­dered. The GPU then dis­cards those pix­els rather than wast­ing time ren­der­ing them. The draw stream bin­ning ras­ter­izer’s de­sign “lets us visit a pixel to be ren­dered only once,” re­veals Man­tor.

The re­vamped Vega ar­chi­tec­ture also now feeds ren­der back-ends from the pixel en­gine into the larger, shared L2 cache, rather than pump­ing them di­rectly into the mem­ory con­troller. AMD says that should help im­prove per­for­mance in GPU com­pute ap­pli­ca­tions that rely on de­ferred shad­ing.

5. Next-gen com­pute en­gine

Fi­nally, AMD teased Vega’s ‘next-gen com­pute en­gine’, which is ca­pa­ble of 512 8-bit op­er­a­tions per clock, 256 16-bit op­er­a­tions per clock, or 128 32-bit op­er­a­tions per clock. The 8- and 16-bit ops mostly mat­ter for ma­chine learn­ing, com­puter vi­sion, and other GPU com­pute tasks, though Ko­duri says the 16-bit ops can come in handy for cer­tain gam­ing tasks that re­quire less strin­gent ac­cu­racy as well. (The AMD-pow­ered PlaySta­tion 4 Pro also sup­ports 256 16-bit op­er­a­tions per clock.)

Co­in­ci­den­tally enough, the Vega NCU can per­form two 16-bit ops si­mul­ta­ne­ously, dou­bled up and sched­uled to­gether. This wasn’t pos­si­ble in pre­vi­ous AMD GPUs, Ko­duri says. Vega’s next-gen com­pute unit has been op­ti­mised for the GPU’s higher clock speeds and higher in­struc­tions-per­cy­cle, though AMD de­clined to dis­close the core clock speeds for Vega just yet.

Wait­ing for Vega

The wait for Vega con­tin­ues, but now we have some idea of the ace hid­den up the Radeon Tech­nolo­gies Group’s sleeve. Th­ese tech­ni­cal teases pro­vide just enough of a glimpse to whet the whis­tle of graph­ics en­thu­si­asts while re­veal­ing tan­ta­lis­ingly lit­tle in the way of hard news re­lat­ing to con­sumer-fo­cused Vega graph­ics cards. (AMD doesn’t want to show its hand to Nvidia too much, after all.) It’s clear that AMD’s at­tempt­ing some nifty new tricks to im­prove the ef­fi­ciency and po­ten­tial of Vega both in games and pro­fes­sional uses. De­tails are sure to drip-drop out over the com­ing months.

Fin­gers crossed that Vega comes sooner rather than later, how­ever. AMD teased its 14nm Po­laris GPU ar­chi­tec­ture at CES 2016 but failed to ac­tu­ally launch the Radeon RX 480 un­til the very end of June. Vega has been slapped with a re­lease win­dow some­time in the first half of 2017, so if AMD waits un­til E3 to launch this new gen­er­a­tion of en­thu­si­ast-class graph­ics cards, Nvidia’s beastly GTX 1080 will have al­ready been on the streets for a full year.

Vega looks in­cred­i­bly in­trigu­ing, but even the most diehard Radeon loy­al­ists can only wait for so long to build a new rig, es­pe­cially with AMD’s much-hyped Ryzen pro­ces­sors launch­ing now.

A tech­ni­cal pre­view of AMD’s Radeon Vega graph­ics ar­chi­tec­ture

Vega’s Prim­i­tive Shaders

Vega’s New Com­pute Unit can per­form two 16-bit ops at once

Newspapers in English

Newspapers from UK

© PressReader. All rights reserved.