Thoughts About AMD Vega
Coming off the release of the Ryzen CPU lineup, RTG has a lot to prove with the up coming launch of the new Vega GPU series. While Ryzen has been a bit of a let down in the gaming department (although it’s no slouch), the R7 series has been running away with compute performance for workflows involving editing, rendering, basically any kind of CPU horsepower, either demolishing the similarly priced chips or costing a fraction of what the Intel performance equivalent runs for. RTG has been out of the high end GPU game for quite a while now, and really needs to hit a home run with Vega.
If everything lines up, they could do it.
*Note, information is simply speculative at this point. We have little information available, and what little we have is either directly from AMD, or through leaks. These leaks while suggestive of potential performance, should not be considered the final say in performance, as it is quite possible the performance could increase due to software improvements, driver improvements, or simply due to optimizations made to the card.
Let’s talk about what we know about Vega. We know that Vega will feature up to 4096 stream processors, as well as up to 8gb of HBM2 memory. Beyond that, little is known about the hardware under the hood. However, Vega also represents one of the largest jumps in architecture design for AMD in years. Specifically, the improvements, or addition, of the High-Bandwidth Cache Controller, Rapid Packed Math, Next Generation Geometry Pipeline, Next Generation Compute Engine, and finally Advanced Pixel engine. All of these promise to do a whole lot to help Vega succeed in its mission, but before we discuss the potential for these improvements, let’s examine GCN. Specifically, the latest version of GCN utilized in the Polaris architecture.
Without digging too much into the history of GCN, the latest generation brought mostly efficiency changes to the table, rather than raw performance increases. Changes were made to increase tessellation performance, however these were slight. A large majority of the changes went towards optimizing the architecture for finfet designs. The previous generation, GCN3 made improvements to the memory, specifically in improving bandwidth and lowering latency, however in GCN4 we do not see many apparent changes to the architecture in this regard. For this reason or others, the RX400 series while very efficient compared to previous generations suffers a loss of performance in regard to memory.
Thats right! On Polaris GPUs overclocking the memory can lead to nearly linear improvements, 10% bump to speed, 10% in performance. However, it’s a bit more complicated than that. In the first few weeks following launch, BIOS modding was not available on the new RX400 series, leaving overclockers with little to do but bump memory speeds and hope for the best. After such mods were developed, it became apparent that as clock speeds increase, so do timings. This is nothing new, however the rate at which the latency increases is nearly the same as the increase in bandwidth, offsetting some of the performance in all but a few cases, thus limiting the impact that overclocking memory could have on polaris. Through BIOS mods, these timings can be changed and “tightened”, therefore allowing the memory to scale quite well. Check out the link at the end of the article for an example of memory scaling, courtesy of the wonderful BIOS modding thread over at the overclock.net forums. What this tells us is that there is plenty left in the current architecture in terms of GPU power, and that memory is, to at least some degree, holding performance back. One of the key components of Vega addresses this issue directly.
HBM2 has an incredible amount bandwidth. With what we know about Vega right now, we’re looking at a card with around 512GB/s of bandwidth, more than double the bandwidth offered on the RX480 GPU. That said, by removing the memory bottleneck, Vega is already looking at an increase in performance in the range of 10-20%, all things held equal, not too shabby. The story gets better though, as Vega is reported to have up to 4096 stream processors, slightly less than double the count on the RX480. Assuming scaling performance and a 15% boost, this puts the theoretical performance of a 4096 SP Vega GPU at or above GTX 980 levels of performance*. Not that impressive, all said and done. At this rate, it doesn’t look like AMD will be crushing the Titan Xp any time soon. However, AMD has demoed the new Vega cards performing similarly to a GTX 1080 before, which crushes the GTX 980, so what gives?
The secret lies somewhere in the improvements made in the new GCN5 architecture. There is no way to tell for sure what improvement is making the difference, or really what exactly is going on under the hood. All we have to go on right now is what AMD has shared with the public as mentioned before, specifically the HBM Cache, RPM, and the Next Generation Geometry Pipeline. How the HBM Cache works, we don't yet know. According to AMD, this alone can contribute up to a 100% boost in minimum framerate, as well as improving performance an average of 50%. RPM appears to be part of the Next Generation Compute Engine, or perhaps a subset of the Next Generation Geometry Pipeline. RPM is said to increase the amount of calculations that the Vega GPU can process, and as demonstrated in the TressFX demo at an AMD event earlier this year the calculations more than doubled. Finally, looking at the Next Generation Geometry Pipeline, AMD claims to have achieved more than double the performance throughput per clock over GCN4. While this has yet to be seen demonstrated, this would suggest that Vega will be much better at leveraging the GPU horsepower than GCN4, allowing the full potential of this chip to shine through.
This all boils down to performance. Between smarter caching, more efficient tessellation computation, double the throughput, smarter and more efficient rendering, removing the memory bottleneck present in GCN4, as well as other architectural improvements, Vega is shaping up nicely. On top of all of this the 14nm manufacturing process is now far more mature, hopefully allowing for higher clockspeeds and overclocking. The most recent performance metric we have to measure Vega with is a SiSandra performance test placing Vega Roughly 35% ahead of the 1080**, or right in line with the 1080ti, very competitive indeed.
For AMD to hit a home run here, they need to be more than striking distance with Nvidia’s big bad boy, the 1080ti. We need to see a repeat of the 290x, not the Fury x. By this I mean, we need nearly identical performance to the titan, or in this case the 1080ti, for a lower cost. AMD doesn’t have to beat the titan, just come within 2-5% and offer the card at a lower price, try 500-600 range. This puts pressure on the the 1080ti as well, further undercutting Nvidia and driving the value of these cards up up up. From the looks of things it seems that AMD could have a winner on their hands, and one with the potential to grow and become an even better card with updates and optimizations, as well as laying the groundwork for an even more refined and improved GCN later on. All that’s left to see is the real world performance, and where AMD decides to price their new king.
*Based on very rough averages obtained from scrolling through the top 200 scores of the RX480 and GTX 980 when paired with an Intel i7-7700k on 3dMark Timespy
RX480 - 4300-5000
GTX 980 - 4400-5200
Link to OCN BOIS modding thread: http://www.overclock.net/t/1604567/polaris-bios-editing-rx480-rx470-rx460
Vega looks like typical AMD - making futuristic products which aren't as relevant today. This is like Fury X all over again, I don't think it'll beat 1080 Ti. My guess it'll end up somewhere between 1080 and 1080 Ti. I have no doubt that two years down the line it'll beat even Titan Xp, but not today. The only thing they can do is price it well and move on to Navi by the time Volta arrives next year. They can't afford another year long lapse in the high-end market again.