The first reviews of AMD’s new Trinity line of APUs are coming in. The successor of AMD’s Llano replaces the Stars-based CPU logic with a Piledriver-based CPU. Piledriver, as you may know, is an update of the Bulldozer architecture.
Bulldozer was not exactly a strong performer in the performance-per-watt area. Since Trinity, like Llano is mainly aimed at mobile platforms, it needs to do a lot better than Bulldozer here. Battery life is all important for mobile systems.
And AMD managed to improve on the power consumption issues. One big improvement is that Bulldozer used soft-edge flip-flops everywhere (which aren’t very sensitive to clock jitter). Piledriver has replaced these with hard-edge flip-flops where possible, which reduces power considerably. Sounds a bit like what ex-AMD engineer Cliff A. Maier was talking about: Bulldozer was not hand-optimized, but just a ‘bruteforce’ automated design.
AMD also silently introduced the F16C instructionset extension. This extension offers instructions to convert between 16-bit (half precision) floating point and 32-bit (single precision) floating point. They could be useful for interoperation with the GPU, which can read and write to 16-bit float buffers. And that is what an APU should be all about: heterogeneous computing.
So how is the improved Piledriver not an improvement then? Well, Bulldozer was a very large and powerhungry chip. In order to stay within the same power envelope as Llano, Trinity can only use a 2-module (4-core) CPU at most. Since Llano also used a quadcore CPU, and the Stars architecture was more efficient than Bulldozer, the new Trinity isn’t all that spectacular.
The power consumption in general more or less the same as it was with Llano. Power management has improved for some scenarios, so you will get slightly better battery life when idle (but when are you ever completely idle for a long period of time?) On the other hand, under workloads such as h264 playback, battery life is still worse than Llano. It is certainly not a leap forward, and AMD was already trailing well behind Intel here.
And performance? Well, it is faster than Llano, just not by a whole lot. And ironically enough, for a CPU that was designed more for multithreaded performance than for single-threaded tasks, it even loses to Llano in the Cinebench multithreaded benchmark:
So Piledriver is cutting it pretty close on the CPU part. Which is probably what lets it down in the next part: GPU performance.
On paper, the GPU is much faster than Llano’s. Up to 50%. And in some tests, you may get close to these results, such as 3DMark Vantage in Performance mode:
However, in other tests it is a different story. As I already pointed out in the comments when Ivy Bridge launched: it’s not as simple as just strapping on a faster GPU. The CPU needs to be fast enough to drive that GPU. In fact, 3DMark Vantage already shows us the elephant in the room: The ASUS N56VM scores about the same as Trinity, with Intel’s HD4000 GPU. Now it goes without saying that the HD4000 GPU is much slower than AMD’s Trinity GPU. But since the Core i7-3720QM CPU is much faster than Trinity’s, it makes up for it.
We see the same in some actual games, such as Batman: Arkham City:
Or Dirt 3:
The Intel system is faster, despite having a slower GPU. Note also the very small differences between Trinity and Llano here. It looks like the CPU is holding Trinity back. The GPU may be faster, but the CPU cannot leverage it.
It seems that both Intel and AMD have very much unbalanced APUs. AMD could do with a better CPU, and Intel could do with a better GPU. At the very least it shows that games are not all about GPU performance. The CPU is equally important. For pretty much anything else, Intel is still the winner hands-down. They offer much better performance coupled with much better battery life. Llano and Piledriver were supposed to offer value and better gaming performance mostly. But Intel has already caught up on the gaming performance enough to be competitive in at least some titles.