AMD Radeon Vega - The Next Big Leap in GPU Technology
It feels like years since AMD has had a truly next generation GPU architecture, and while the Fiji architecture was a nice departure from the 'same old same old' architectures we've come to know and love, it didn't rattle the cage all that much.
That all changes with Vega, in more ways than one. But before we get into the juicy stuff, let's talk about where AMD has come in 18-24 months. I took over as the lead GPU editor for TweakTown towards the tail end of 2014, so I've been knee deep in GPU technologies, events, architectures, graphics card reviews, monitor technology and so much more. I've been doing this for 20 years now, but writing professionally for a major tech site is a big difference.
For the first time in a very long time, I'm beyond excited about the future of GPU technology. We've been on GDDR5 memory technology for over 5 years now, we were on the 28nm node for around the same time - and only shifted to 14/16nm with AMD's newish Polaris GPU architecture, and NVIDIA's Pascal GPU architectures, respectively.
At the end of 2015, NVIDIA held over 80% of the discrete GPU market - a massive strangehold and dominance that seemed like an impossible mountain that AMD had to climb. I remember sitting down with the key staff of the Radeon team during GDC 2016 and being asked "what should we do, what should AMD do to be bigger and better over the next 12-18 months?" and I remember telling AMD to "not fight NVIDIA in the high-end, and hit them in the low/mid-range market".
My response was met with a sly smile, and then Polaris hit the mainstream sub $300 market in a big way. Fast forward to the end of 2016 and AMD had secured itself another 10% of the GPU market, bumping them up to 30% or so of the dGPU market. This is all with mid and even lower-end graphics cards like the Radeon RX 470 and RX 460, a market that NVIDIA had to respond to with the GeForce GTX 1050 and GTX 1050 Ti.
Enter Vega... AMD's Next Big Thing
Whenever I say 'AMD' next to 'Vega' it always feels weird - as Radeon Technologies Group is the team behind Polaris and Vega, and you can feel it in more ways than one. Prior to that, the Hawaii and Fiji (and everything in between) were part of the old AMD.
AMD spun off its GPU division into Radeon Technologies Group in late 2015, and the team has been working around the clock since. Has it paid off? You can be damn sure it has, and they're only getting started.
Enter Vega... where my excitement begins to swirl into a mess of enthusiasm and excitement. I can't believe I'm writing this as a 'professional', as I'm just as excited as you guys and girls are about RTG's new technology - so let's jump right into the middle of it.
There are some really interesting things to go through, so prepare yourself for a nice read. But I have to tease you as we go into the next page... RTG explained at its Tech Summit 2016 event that "conventional architectures are not scaling to meet needs", and they are 100% right.
Radeon Vega - World's Most Scalable GPU Memory Architecture
See. I told you - look at that headline: "the world's most scalable GPU memory architecture". Say what?! How is this so? Is it just HBM2? No - it's not just that, as AMD have always been a company that strives towards next generation technologies, and that's what Radeon Vega introduces.
Introducing High-Bandwidth Cache
Everyone until now has thought that AMD has just worked on a refreshed GPU architecture and slapped HBM2 technology inside, but it's not quite that easy - a large part of Vega is the new High-Bandwidth Cache.
We also have HBM2 technology, which provides up to 2x bandwidth per pin - a huge 1024GB/sec (1TB/sec) of VRAM, ready for your current, and next-gen games.
HBM2 provides some massive savings on the PCB of graphics cards, with a 50% smaller footprint when compared to GDDR5.
Next up, we have the High-Bandwidth Cache Controller, which has up to 512TB (yes, that's terabytes) of virtual address space.
This allows for what AMD explains as "adaptive, fine-grained data movement" - remember, Vega isn't here for gamers only, but for supercomputers and massive systems that work on super-complex dataloads.
If we look at 4K gaming on Ultra settings with titles like Fallout 4 and The Witcher 3, the difference between 'allocated' and 'accessed' VRAM - there's a large difference. You don't need as much VRAM as you think, because the game will simply chew up a bunch of it - but only uses around half of that in real-world usage.
Radeon Vega Technology Deep Dive, Continued
AMD has included a new programmable geometry pipeline with the Vega GPU architecture, which has over 2x peak throughput per clock.
We all remember Vertex and Geometry shaders, but now we have a Primitive shader - which, isn't - well, primitive at all.
The new Geometry engine is super efficient, as it works with an 'intelligent workgroup distributor' that spreads the load across the GPU much more efficiently.
I did say that there is a lot of changes with Vega, and the next-generation compute engine is another big part of the new GPU architecture from AMD.
Introducing the Vega NCU
Vega GPU, you mean - Anthony? Nope, the Vega NCU - AMD's next-generation compute unit.
We have 512 8-bit operations per clock, 256 16-bit operations per clock, and 128 32-bit operations per clock. AMD also adds that double precision rate is configurable, for compute-intensive applications and systems.
The massive workloads that some users will have will benefit from the 'rapid packed math' that the Vega NCU is capable of.
Better yet, the Vega NCU is optimized for higher clock speeds and higher IPC - music to my ears.
AMD has also added that they've been "working on reducing memory bandwidth consumption for many years".
Next-Generation Pixel Engine
AMD has also baked in a new pixel engine into the Vega NCU, which is helped by the draw stream binning rasterizer.
The new draw stream binning rasterizer is "designed to improve performance and save power" by only fetching once, thanks to the new on-chip bin cache.
Final Thought on Radeon Vega
It's hard to say what Radeon Vega will perform like apart from the few demos we've seen so far, with Vega 10 powered by 8GB of HBM2 running DOOM at 4K Ultra settings at 60FPS+ as well as Star Wars: Battlefront at the same resolution and 60FPS+.
I'd love to have Vega 10 in my hands before I talk performance, but just from the technology side of things - I'm vibrating with excitement. AMD will be the first with High-Bandwidth Cache, HBM2 on a consumer graphics card, and will finally join NVIDIA in the high-end space for the first time in over 18 months.
We have to wait a few more months to see if HBM2 and High-Bandwidth Cache will improve gaming performance at lower resolutions like 1080p and 1440p, but I'm sure at 4K and beyond - HBM2 and HBC will be things of wonder.
VR gaming and multi-monitor gaming should heavily benefit by the massive bandwidth of HBM2, but again - we'll have to wait and see.
The Radeon Rebellion continues.