As the mobile world has quickly filled up with devices using the Snapdragon 801 chip, device manufacturers are now busy building the next-generation handsets and tablets around the Snapdragon 805 chip. The name doesn’t provide a real good hint, but Snapdragon 805 (MSM8084 internal codename) is a much larger jump.
You may remember that Snapdragon 801 was a “performance kicker”, or an optimized version of the Snapdragon 800 processor. As such, it was pin-compatible and easy to upgrade. As we have seen in the real world, the performance improvements were real at times, although not out of this world, except for graphics performance. Snapdragon 805 is designed to bring more performance improvements, but new features are truly its reason of being. Let’s take a closer look.
New Cores, Faster Memory
First of tall, Snapdragon 805 is built around a new CPU core design called Krait 450, which is an evolution of the Krait 400 core that has served Qualcomm very well in the recent past. This new core can be clocked as high as 2.7GHz according to Qualcomm, and has support for more than 4GB of RAM, thanks to a 36-bit Large Physical Address Extensions (LPAE).
This is a sign that upcoming handset will support more than 4GB within the next 12 months. Ultimately, a full move to a 64bit solution will solve the memory addressing concerns for a long time.
Talking about memory, Snapdragon 805 uses a 128bit LPDR3 memory interface (2x64bit POP memory, up to LPDDR3-1600) that can yield a bit more than 25 GBps of theoretical memory bandwidth. This is one of the highest bandwidth available, and no matter how you look at it, high-performance applications are often constrained by the amount of data they can move (you can’t process it before accessing it…).
The biggest difference with Snapdragon 805 is once again in the graphics processor (GPU) and its 4.8 Gigapixel/sec fillrate. Qualcomm did not make any particular noise about it, but the Adreno 420 GPU has DirectX 11.2 level features that were built by a team with many former AMD engineers. Adreno 420 also has higher memory bandwidth, which is extremely important since graphics are the single most demanding applications for that.
Among the new GPU features, Adreno 420 supports ASTC texture compression, which is key to lowering the memory footprint and bandwidth requirements of modern games. Unified Shader are extended to include pixels, vertex, compute, tessellation and geometry operations. We’ve known all that for some time, but if you want to see an 805 tessellation demo, check this out:
As usual, Adreno supports Tiled rendering and Direct rendering, and can switch on the fly. Qualcomm remains the only GPU vendor that can go from one to another liberally. Imagination is a good example of pure tiled renderer, while NVIDIA’s K1 only use direct rendering. Both techniques have pro and cons, so it’s nice to be able to switch whenever.
I honestly thought that going from DX9 to DX11 graphics would be more painful for Qualcomm, but by hiring the right people, Qualcomm has been able to get ahead of the game and be ready for OpenGL 4.x (pdf link) graphics while Android doesn’t even support it. OpenGL ES 3.x will offer partial DX11 features support at best, thanks to OpenGL Extensions.
This means that for the time being, there are no apps that can truly use the full potential of this GPU, which is why we will focus on existing games until something better comes along.
For those of you who aren’t interested in Gaming, a fast GPU also allows the fast image composition which is required to handle complex user interfaces on 4K displays. Tablets, but also smartphones will be prime targets for this.
Finally, the compatibility with OpenCL 1.2 will let developers use the GPU compute units to do things like many people do often: stitching panoramas, post-process videos, running RenderScript and many more things that aren’t 3D graphics.
Higher Resolution Video and Photography
Snapdragon 805 can now decode H.265 (also called HEVC) which is critical to handling 4K, thanks to the improved compression. Unfortunately, hardware encoding remains in H.264, which means that 4K camera recording will still occupy quite a lot of space. I’ve tried 4K recording with the Galaxy S5, and although it looks impressive, the sheer size of the files would deter me from uploading videos from the phone.
Since Snapdragon is capable of capturing 4K video at 30FPS, it can use the same horsepower to capture 1080p videos at 120FPS. This can be useful to capture super-smooth videos, but more importantly, this lets people do easy slow-motion movies with just a few taps. The 805 chip is designed so that the data stream coming from the image sensor can go through the GPU or DSP before being processed by the ISP module. This allows for exotic image processing without changing the overall data workflow.
Since we’re going into higher resolutions, the Image Signal Processor (ISP) has seen a huge performance boost (near 50%) and can now process 1.2 Billion pixels per second. There are actually two ISPs in Snapdragon 805, which is handy to do things like dual-shot, dual-recording etc… from the front and back cameras. For instance, the chip can process a 4k/30fps recording and a 16 Megapixel 15FPS burst photo shoot at the same time.
Compare that to the 640 million pixels per second of Snapdragon 800. If you are not familiar with the ISP, it’s the chip that is in-between the camera sensor and the main application processor. It is responsible for turning the raw image data into digital images and video.
Hollywood Quality Video (HQV) is a video upscaler that turns 1080p into 4K content. This works by filtering the image to smooth curves and add pixels. While this is not as good as “native” 4K content, it does look perceptibly better than 1080p content that is just magnified with a lesser algorithm. Since most of the content is still circulating at 1080p resolution, this will improve the quality right away.
Snapdragon 805 is ready to support depth cameras (like the ones for Google’s Project Tango) and stereo camera setups for the next-generation applications. The idea is to have the phone to be visually aware of your surroundings and be able to have a volumetric representation of that environment. In the near future, those handsets will be able to capture images and build relatively accurate 3D representation, including textures, precise room measurements, and other characteristics that can be derived from knowing the 3D layout of what is being looked at.
Performance & Benchmarks
OK, you probably have more than enough context, and we can get to the meat of this page, which is the raw performance benchmarks as seen on a development reference device. Keep in mind that each handset maker (OEM) can tweak this up and down (mostly down) to fit their own power and thermal requirements.
Note that these benchmarks were run on early hardware and very early drivers, so even as we tested this, the results that we got from multiple machines were quite different at times. I used an average of values taken over multiple runs and multiple machines to compile those numbers, and here they are. I’ll comment below.
Like it was the case before, the largest jump in measured performance comes from the OpenGL ES 2.x level graphics. This is a clear cut win, and this shows that existing games will get a serious boost. The good news is that the large majority of games still use this kind of graphics, so there is an immediate benefit to future users.
While GLBenchmark is a pure graphical test, 3DMark Icestorm unlimited tries to be a more balanced workload between CPU and GPU. Here, you can see that performance is much closer to what’s on the market today. Now, if you factor in that the drivers and firmware of Snapdragon 805 will improve by launch time and over its lifetime, it should end up with a small gain by the time it becomes popular. I’m not sure if the GPU is constrained by the CPU here, but this is probably a more realistic picture for certain types of games.
The Geekbench single-core CPU performance test shows what we expect: a slight bump in CPU performance which correlates to the technological improvements of Krait 450. Single-core performance remains key because many apps or tasks don’t scale very well with more cores.
Strangely enough,to see g the lead doesn’t extend to the multi-core test for which I was expecting the same kind of improvement as seen in the single-core test. Some of the systems went slightly above 3000 while we tested them, but overall, the average number had remained below some of the notable devices already shipping today.
Conclusion and looking forward
While Snapdragon 805 is designed to bring additional performance, it is primarily a vehicle to introduce new features. As such, it will be more expensive for handset makers than the Snapdragon 801 counterpart, which is why Qualcomm expects both chips to be in circulation until sometime in 2015.
It is interesting to see how graphics performance is evolving much faster than CPU performance in general. And since this is Qualcomm’s first generation DX11-level graphics hardware, expect the next revision to provide another boost due to optimizations of the same architecture.
We have seen the same thing happen in the PC world where GPUs grew in performance at a rate the industry called More’s law “cubed”. That’s because they can easily scale by having more cores since “computer graphics” is a massively parallel problem in which each pixel can be processed nearly independently from the others.
The other thing to remember is that mobile devices aren’t just built for “absolute” top performance because the most important factor remains their power consumption, which is limited. In fact, a lot of the performance improvements could be better expressed in terms of performance-per-watt, and this is something that benchmarks scores don’t always show well. But that is another story…
Next Story: Apple Is The Leading Smartphone Vendor In The US