The Radeon HD 4850 & 4870: AMD Wins at $199 and $299

AMD's RV770 vs. NVIDIA's GT200: Which one is More Efficient?

It is one thing to be able to sustain high levels of performance and altogether another to do it efficiently. AMD's architecture is clearly the more area efficient compared to NVIDIA.

Alright now, don't start yelling that RV770 is manufactured at 55nm while GT200 is a 65nm part: we're taking that into account. The die size of GT200 is 576mm^2, but if we look at scaling the core down to 55nm, we would end up with a 412mm^2 part with perfect scaling. This is being incredibly generous though, as we understand that TSMC's 55nm half-node process scales down die size much less efficiently one would expect. But lets go with this and give NVIDIA the benefit of the doubt.

First we'll look at area efficiency in terms of peak theoretical performance using GFLOPS/mm^2 (performance per area). Remember, these are just ratios of design and performance aspects; please don't ask me what an (operation / (s * mm * mm)) really is :)

This shows us that NVIDIA's architecture requires more than 2x the die area of AMD's in order to achieve the same level of peak theoretical performance. Of course theoretical performance doesn't mean everything, especially in light of our previous discussion on extracting parallelism. So let's take a look at real performance per area and see what we get in terms of some of our benchmarks, specifically Bioshock, Crysis, and Oblivion. We chose these titles because relative performance of RV770 is best compared to GT200 in Bioshock and worst in Oblivion (RV770 actually leads the GT200 in bioshock performance while the GT200 crushes RV770 in Oblivion). We included Crysis because it's engine is quite a popular and stressful benchmark that falls somewhere near the middle of the range in performance difference between RV770 and GT200 in the tests we looked at.

These numbers look at performance per cm^2 (because the numbers look prettier when multiplied by 100). Again, this doesn't really show something that is a thing -- it's just a ratio we can use to compare the architectures.

While it doesn't tell the whole story, it's clear that AMD does have higher area efficiency relative to the performance they are able attain. Please note that comparing these numbers directly doesn't yield anything that can be easily explained (the percent difference in frames per second per millimeter per millimeter doesn't really make much sense as a concept), which is part of why these numbers aren't in a graph but are in a table. So while higher numbers show that AMD is more area efficient, this data really doesn't show how much of an advantage AMD really has. Especially since we are normalizing sizes and looking at game performance rather than microbenches.

Some of this efficiency may come from architectural design, while some may stem from time spent optimizing the layout. AMD said that some time was spent doing area optimization on their hardware, and that this is part of the reason they could get more than double the SPs in there without more than doubling the transistor count or building a ridiculously huge die. We could try to look at transistor density, but transistor counts from AMD and NVIDIA are both just estimates that are likely done very differently and it might not reflect anything useful.

We can talk about another kind of efficiency though. Power efficiency. This is becoming more important as power costs rise, as computers become more power hungry, and as there is a global push towards conservation. The proper way to look at power efficiency is to look at the amount of energy it takes to render a frame. This is a particularly easy concept to grasp unlike the previous monstrosities. It turns out that this isn't a tough thing to calculate.

To get this data we recorded both frame rate and watts for a benchmark run. Then we look at average frame rate (frames per second) and average watts (joules per second). We can then divide average watts by average frame rate and we end up with: average joules / frames. This is exactly what we need to see energy per frame for a given benchmark. And here's a look at Bioshock, Crysis and Oblivion.

This is where things get interesting. AMD and NVIDIA trade off on power efficiency when it comes to the tests we showed here. Under Bioshock RV770 requires less energy to render a frame on average in our benchmark. The opposite is true for Oblivion, and NVIDIA does lead in terms of power efficiency under Crysis. Yes, RV770 uses less power to achieve it's lower performance in Crysis and Oblivion, but for the power you use NVIDIA gives you more. But RV770 leads GT200 in performance under Bioshock while drawing less power, which is quite telling about the potential of RV770.

The fact that this small subset of tests shows the potential of both architectures to have a performance per watt advantage under different circumstances means that as time goes on and games come out, optimizing for both architectures will be very important. Bioshock shows that we can achieve great performance per watt (and performance for that matter) on both platforms. The fact that Crysis is both forward looking in terms of graphics features and shows power efficiency less divergent than Bioshock and Oblivion is a good sign for (but not a guarantee of) consistent performance and power efficiency.