As AMD was in the process of ramping up for the Polaris launch last year, one of the unexpected but much appreciated measures they took was to release a bare-bones GPU architecture roadmap for the next few years. AMD has traditionally held their cards very close to their proverbial chest on what they’re working on next, typically only announcing a new architecture weeks before it’s set to launch in retail products. Polaris itself was a departure from that, as it was announced a good 5 months in advance, but last year’s roadmap was the first honest-to-goodness look we’ve had at AMD’s long-term plans in a long time.

What did that map show us? After 2016’s Polaris would come Vega, and after that would be Navi. As a high-level roadmap it didn’t show us much – in fact other than a timeframe, the only detail attached to Vega was “HBM2”  - but it was enough to help understand one of the things AMD would be doing architecturally to set Vega apart from Polaris. As for the timeframe itself, that was ambiguous at best in AMD’s roadmap. But now as we draw closer to the launch of Vega, the picture has become clearer. AMD will be hitting a yearly cadence with Vega. The first chip, which tapped out last year, will be launching in the first half of this year (H1’17).

To that end, with Vega’s launch not too far over the horizon, AMD is ready to start talking about what will be their next GPU architecture. Last year at this time we got our first real glimpse into Polaris and what would become the Radeon RX 480/470/460, and this year AMD is back again with a teaser of things to come with Vega.

Setting The Stage: Expectations Management; Less How & More Why

Before we dive into any architectural details, perhaps it’s best we first set the stage. This goes for both what to expect of today’s announcement, and to better understand what AMD is doing and why.

First and foremost, today’s detail release is a teaser, not a deep dive, or even a preview. AMD is only releasing a few details about Vega, and those are being kept at a high level. In fact it’s fair to say that there’s just enough information to answer little and raise even more questions; just what a proper teaser should be.

Why? Well part of the reason is that we’re still months off from the launch of Vega. I believe it’s fair to say that by announcing a first-half of the year launch date when we’re already in 2017 is a strong indicator that Vega will not launch until later in that window, likely some time in Q2. So we’re still a good three to five months out from the launch of Vega, which means AMD doesn’t want to (or need to) release too many details this far out. Rather they can trickle out chosen details for maximum impact.

At the same time the AMD of 2017 has more they can focus on in the high-performance space than just GPUs. Ryzen launches soon, and they also have other products on the horizon such as the Radeon Instinct accelerators. Polaris received as much detail as it did because it was all AMD really had to talk about, and they needed to recover from a rough 2015 where AMD’s at-the-time power efficiency woes were brought into full focus. But now Vega can share the stage with Ryzen and other products, and that lets AMD be more selective about what they say.

All of which is something I would argue is a good thing. At the end of the day Polaris was an optimized version of the GCN 1.2 (aka GCN 3) architecture for GlobalFoundries’ 14nm FinFET process. The resulting GPUs were solid competitors in the mainstream and value markets, improving on AMD’s power efficiency in a way they badly needed. But they weren’t high-end parts; they didn’t excite like those parts did, and for technology enthusiasts they didn’t significantly change the architecture itself (in fact GCN 4 was ISA compatible with GCN 3, something that doesn’t happen a lot in the GPU space). AMD talked big about Polaris – perhaps too big – and I do think it hurt them in some circles once it became clearer that this was AMD catching up. Which is not to say that AMD’s marketing arm won’t talk big about Vega as well, but they need not ride the technology angle so hard. Vega is a launch that can be more natural and more subdued, especially as at this point we know AMD is aiming big with a much-needed new generation of high-end parts.

In any case, as AMD isn’t riding the technology angle quite as hard in this year’s teaser, they are spending a bit more time explaining the market and some of the logic behind Vega’s design. For its teasing debut, Vega is little less discussion of “how,” and a little more conversation of “why”.

So what is AMD looking to do with Vega? Besides aiming for the high-end of the market, AMD is looking at how the market for GPUs has changed in the last half-decade, and what they need to do to address it. Machine learning is one part of that, being a market that has practically sprung up overnight to become a big source of revenue for GPUs. This is where the previously announced Radeon Instinct will fit in.

But more than that, it’s about fundamental shifts in how workloads are structured. GPU performance growth has far outpaced GPU memory capacity. Scene geometry complexity has continued to grow. Newer rendering methods have significantly changed GPU memory access patterns.

To that end, AMD is looking to address all of these factors with Vega. Which is not to say that this is everything – this is a teaser, after all – but this is where AMD is starting. Where they are going to be with their next generation architecture and how they believe it will address the changes in the market. So without further ado, let’s take a teasing look at what the future has in store for AMD’s GPUs.

Vega’s NCU: Packed Math, Higher IPC, & Higher Clocks
Comments Locked


View All Comments

  • silverblue - Thursday, January 5, 2017 - link

    Formerly Videologic, changed in 1999 to Imagination Technologies.

    I had a Kyro II. The first Kyro was clocked at a mere 115MHz for both core and memory, but Kyro II was a whole 60MHz faster. It really worked well in UT when paired with my XP 1700+. :) There was a Kyro II SE with 200MHz clocks but was as rare as rocking horse excrement. I was also very excited about its supposed TnL, DDR-powered successor, but it never launched, which meant bye-bye on the desktop for PowerVR, rather than the bloody nose that ATi and NVIDIA deserved at that time.
  • wumpus - Friday, January 6, 2017 - link

    If my poor memory is remotely accurate (don't bet on it), the chip to run Unreal Tournament was the S3 chip (probably the last they made). Not sure if the "TnL" it supposedly had was fraudulent or defective, but the texture compression (probably only used in Unreal Tournament before S3 died and it became a standard feature) gave the S3 the best visuals in the game.

    I know I had one of those boards (simply because it was cheap, long after its day) but can't remember how it performed on Unreal Tournament (I may not have bothered putting windows on that box. In those days the best way to connect windows to the internet was to connect a Linux machine to a DSL modem, then connect windows to Linux. Connecting Windows to DSL was asking to get it to blindly reset all your configuration files at random intervals).
  • BrokenCrayons - Friday, January 6, 2017 - link

    You're probably thinking about S3's Savage video card series with the S3 MeTaL API. It's been ages, but I'm pretty sure UT supported MeTaL out of the box. Epic was pretty good about that back when there was a plethora of competing APIs before we settled into DirectX and OpenGL. When I tried to diversify my PC shop by adding an after-hours LAN arcade, we used those graphics cards because they were good for the multiplayer stuff we wanted to run. I do recall they were a bit glitchy under Descent Freespace. There were artifacts around ship engine glow effects and I think it was due to poor DirectX suport, but don't quote me on that, it's been years.
  • silverblue - Friday, January 6, 2017 - link

    Yeah, MeTaL was there at retail, as long as you installed the texture pack from the second disc. I had a Savage 4 (Diamond S540) which worked flawlessly... until I flashed the BIOS one day, thereafter it would hang during every session at a random point, thus requiring a reboot. That was a stupid idea, especially considering it looked excellent. Luckily, somebody got the OpenGL mode to work with those textures.

    I believe the T&L on the Savage 2000 was unfit for purpose; it was broken in hardware so you couldn't coax its true performance out.
  • silverblue - Friday, January 6, 2017 - link

    Sorry, to clarify, you could use MeTaL without the texture pack, but what would be the point? :)
  • Threska - Friday, January 6, 2017 - link

    And for the time it was a very nice card. Shame Kyro didn't take off
  • eldakka - Friday, January 6, 2017 - link

    I always liked the infinite planes concept (no polygons!) of the PowerVR 1 series. I remember being disappointed at the time, and still am, that infinite planes never caught on. The entire paradigm gave many operations 'for free' that today require a lot of hardware and software support to implement.
  • Threska - Saturday, January 7, 2017 - link

    Could have been a patent issue.
  • tarqsharq - Friday, January 6, 2017 - link

    I remembering excitedly reading about TBR with an article about the Kyro II on this very website oh so many years ago like you're saying.

    Funny enough, it didn't do well in the market because the APIs at the time weren't very flexible and required a lot of tweaking to it working in each engine pretty much.

    The games that it did work well in, the card punched WAY above its cost bracket if I remember correctly.
  • Jtaylor1986 - Thursday, January 5, 2017 - link

    This design has probably been in the works for 3+ years. Nvidia just beat them to market it had nothing to do with David Kanter.

Log in

Don't have an account? Sign up now