AMD unveils 2016 GPU roadmap; Polaris gives 2.5x performance consistent with watt, may additionally make use of GDDR5

At its Capsaicin event the day prior to this, AMD didn’t simply unveil its $1,500 Fiji GPU — it additionally showed off its subsequent-technology Polaris GPU and gave an replace on its longer-time period GPU roadmap. the brand new roadmap isn’t just full-size for what it says, however what it doesn’t mention as properly.

First, a chunk of background. 14nm is predicted to be a main new node for each AMD and Nvidia, however it’s particularly vital to AMD’s scenario. now not simplest is team purple preventing to regain market proportion lost to Nvidia over the last few years, its GPUs have been much less power-green and noisier than crew inexperienced’s counterparts. (The Radeon Nano, it must be cited, become an exception to those tendencies, with markedly higher energy performance than any previous AMD GCN GPU.) that is especially complicated in cell, where every watt and decibel remember.

previously, AMD had claimed that it might supply a 2x performance-consistent with-watt development with Polaris as compared with its 28nm-class hardware. Now, the employer has bumped that prediction to two.5x improved overall performance-according to-watt. Readers must take into account that like any metrics, performance-in keeping with-watt isn't an absolute. due to the fact silicon power consumption is not a easy linear curve, these figures are in all likelihood pleasant-case estimates based totally on midrange hardware, no longer the worst-case state of affairs when comparing top-end elements clocked at maximum frequency. in spite of this, a 2.5x performance bump is a noteworthy generational leap.

AMD’s GPU roadmap

AMD’s GPU roadmap for the subsequent 18 months is proven beneath:

There are numerous interesting bits of information we’d like to talk about. First, AMD’s subsequent-era GPU own family, codenamed Vega, is now forecast to reach hard at the heels of Polaris. After measuring the purple boundary containers and Polaris’ relative function in comparison to the summer time 2016 release time body, the placement of the Vega container suggests a launch at the tail cease of 2016 or very early 2017 time-frame.

The reality that Vega is mainly categorised as the use of HBM2 also indicates that this is the major excessive-end structure revamp that group crimson enthusiasts were waiting for, with a good deal larger body buffers and, if rumors are actual, a real next-technology GPU architecture from AMD (Polaris remains based on GCN, albeit a heavily stepped forward model).

The truth that Vega is categorised as HBM2 while Polaris isn’t categorized with any kind of memory in any respect indicates that Polaris will both use GDDR5 or GDDR5X. I’m inclined to think GDDR5 is much more likely — GDDR5X nonetheless requires considerably greater strength than GDDR5, and isn’t expected to be ready for a quantity launch in a matter of months. laptop perspective believes Polaris will still use HBM1, which is also feasible.

This slide suggests the hassle of stacking second RAM configurations indefinitely. Polaris could nonetheless offer 8GB of RAM in 8x incorporated circuits, but, rather than the sixteen seen right here.

If AMD is using conventional GDDR5 for Polaris, it’s probable for sensible reasons. while the organization debuted Fiji remaining 12 months, it took pains to notice that GDDR5 made step by step much less experience as body buffer sizes and clock speeds ramped up. AMD chose to undertake HBM in 2015, partly because it desired to supply an widespread amount of reminiscence bandwidth and was inclined to simply accept a smaller most VRAM body buffer to do it. The organization has always maintained that HBM made extra experience than GDDR5 for the Fury X and Fury Nano, but by no means claimed that it would adopt HBM across the entirety of its product stack.

one of the rumors we’ve heard approximately Polaris is that AMD will goal the mainstream and upper-midrange of the GPU marketplace. that is wherein the cash is in fanatic gaming; Steam’s hardware Survey is presently broken and not displaying any GPUs from NV at all, but archival facts from September confirmed the GTX 970 because the single-maximum popular GPU from AMD and Nvidia, with three.97% of the whole install base. The Steam hardware survey is inexact at excellent — it frequently takes months to be updated with new cards and at present all the Nvidia data seems to had been rolled into Intel’s facts — but it nevertheless gives goal evidence that this is the sweet spot for AMD to goal if it wants most sales for its efforts. If AMD intends to deliver a GPU with hugely stepped forward performance according to greenback, overall performance-in line with-watt, and preserve open the option for 8GB playing cards, GDDR5 may also still be the higher choice. HBM2 might not be geared up for release and HBM turned into never extra than a niche release, intended to be fee-effective above a excessive fee factor but no longer to take the technology mainstream.

Whither Greenland?

if you’ve paid interest to the diverse AMD “roadmaps” handed around in fanatic circles, you’ve possibly visible references to Greenland sprinkled throughout them inside the beyond year. It’s in no way been clear exactly what the specifications of Greenland were, and the numerous claims made approximately it transformed over time. both way, it’s not indexed in this roadmap.

If we expect that Greenland existed in any respect, this statistics indicates AMD canceled the part in want of pulling Vega in and launching an all-new structure greater fast. Alternately, the code call could still have some bearing on future Zen APUs. We know AMD is working on those parts, but the enterprise has stated not anything about a release date or what functions and skills they’ll offer.

I’ve speculated in several stories that AMD might use HBM2 reminiscence in APUs to eventually smash the bandwidth barriers which have held included snap shots returned for nearly two decades. I still suppose that is likely ultimately, however whether or not or now not AMD rolls the generation out will rely on both HBM/HBM2 charges and Zen’s ability to command favorable margins and higher absolute fees compared to Carrizo and Kaveri/Godavari.

As of this writing, the cheapest DDR4-2133 on Newegg is $fifty nine, even as DDR4-3200 is $85. for the reason that AMD’s included GPUs have frequently been totally memory bandwidth-bound, it’s completely possible that the primary wave of Zen APUs will lean on the 50% improved memory bandwidth presented by means of excessive-velocity DDR4-3200 — and fifty one.2GB/s of memory bandwidth is nothing to sneeze at in a dual-channel gadget.