Qualcomm S4 (Krait) Hardware breakdown

Kevin OQuinn

AC Team Emeritus
May 17, 2010
9,267
496
0
Visit site
There's quite a bit of talk about the new chip from Qualcomm, the S4. Let's get this out of the way right now, this isn't the same as what's in your Evo 3D, Sensation, Amaze, Skyrocket, or Vivid (or any other Snapdragon-based device). This is an entirely new architecture from Qualcomm, with major advancements in every area.

First we need to understand the difference between how Qualcomm develops a chip vs Samsung/TI/nVidia?etc. Qualcomm is an ARM licensee. What that means is that they have a license to use the ARM instruction set, but they make their own chip design. So while a chip from Samsung and TI will look very similar, a Qualcomm chip will look very different. There are a number of advantages, the main one being that Qualcomm can focus the design for the intended device, in our case phones and tablets. They can make changes to improve battery life, performance, etc that cater to the needs of the device. The disadvantage is higher cost to manufacture and develop the chip. This means they generally have less architecture introductions than other companies. So while they were decidedly midrange in performance last year, this year they take the crown back. But how do they do it? Let's explore a bit further to find out.

First some general information about the S4 SoC (and the Krait CPU in general).

S4 is the SoC, and Krait is the CPU. The SoC encompasses things like the CPU, GPU, baseband, memory controller, image processor, and most other things that allow our device to function. The CPU is, well, the CPU. It does what a CPU does and is where the major jump in performance comes from.

Krait is the first mobile CPU to be based on a 28nm production process. This alone gives a pretty substantial increase in power consumption and performance. This is no small feat, either, and is bleeding edge for a mobile CPU. Later this year we'll see other CPU's based on this same process, but for now Qualcomm is exclusive.

Also of note is the power management features Qualcomm built-in. Each core can dynamically change its voltage based on the load it sees. So the less you do, the lower it clocks. The more you do the higher it clocks. If one core isn't needed it can shut off completely. This all happens without you knowing, and with no noticeable hit in performance. What this effectively does is allow the CPU to shut off when it's not needed. Qualcomm expects a huge increase in battery life vs Scorpion and A9 based solutions. All the signs are pointing at this being absolutely correct.

The S4 is also the first to incorporate a baseband that covers the majority of frequencies being used today, including LTE. This is important for a few reasons. The first one should be obvious to anyone that knows how previous LTE phones have worked. They have all had a separate chip for LTE connectivity. This hurts battery life, because it needs a separate controller to talk to the network and communicate with the phone. That's a separate processor that needs to draw from your battery. The S4 includes the LTE connectivity, which means it communicates directly with the CPU. No separate processor means more battery power to do other things. This also allows phones to packaged better, which leads to thinner phones, and at the very least gives manufacturers more flexibility with designs.

Right now Qualcomm is the only SoC maker to integrate LTE onto the SoC, so this gives them a competitive advantage over its competitors, especially in markets with a significant LTE footprint like the US.

Support for wifi and bluetooth are also built-in. 802.11 a/b/g/n and bluetooth 4.0 to be specific.

Digging a little deeper into how the CPU operates, Qualcomm has made significant increases to the number of instructions and processes that can be completed at any given time. They have lengthened and widened the "pipe". The pipe is the path the instruction takes through the various stages in the CPU. Think of it as a firehose. If you make the firehose wider more water can flow at any given time. Same concept here. These changes should give a theoretical 50% boost in performance versus a Scorpion core at the same speed. (note: everything before Krait is Scorpion) This also puts it ~25-30% faster than an A9 based core (think Exynos or OMAP 4).

Another area that sees substantial improvement is in NEON instruction performance. Throughput is increased 50% here. NEON is important for all you gamers out there, because it's a set of instructions designed specifically for gaming and video (there are other things, too, but those are the ones we care about).

The memory interface has been improved too, as it now supports a true dual-channel LPDDR2 setup. On Scorpion it could be setup as dual-channel, but was honestly a hassle and most implementations stuck with a single-channel interface, which hindered performance. Now that the memory controller is capable of a true dual-channel memory setup I suspect we'll see more of this, and that will help with performance. More memory bandwidth is always better. We don't want our really fast CPU waiting for memory to give it what it wants.

These are the areas that have seen the most improvement over Scorpion. It's a formidable architecture and should hold the performance crown at least until we see A15 based chips from TI and Samsung. Qualcomm also has a higher clocked dual-core and quad-core models in the works for release at a later time.

Now on to the GPU...
 
Last edited:
  • Like
Reactions: InTheRough

Kevin OQuinn

AC Team Emeritus
May 17, 2010
9,267
496
0
Visit site
Re: NO COMMENTS PLEASE Qualcomm S4 (Krait) Hardware breakdown

Adreno 225

Not much to talk about architecturally. It doesn't differ at from Adreno 220 in that respect. It does improve upon it in certain areas, though.

Moving to the 28nm process allows it to be clocked higher, up to 400mhz (the 220 was capped at 266mhz). Raw compute performance is actually really high, on par with the PowerVR SGX543MP2, but other architectural limitation prevent it from flexing all of it's muscle. Memory bandwidth and texture fill seem to be the two most limiting factors. Qualcomm has said that they have improved the drivers to wring out more performance and better battery life, too. According to Qualcomm these changes will result in a 50% improvement in performance.

Later this year we'll see the Adreno 3xx which should be a significant improvement over the 225.
 

jeffreii

CrackBerry Convert
Apr 20, 2012
798
53
0
Visit site
Thanks Kevin. Good stuff here.

I find it interesting how down people are on the S4 (which I find to be an annoying moniker when paired with the S3 (not the older Snapdragon S3 but the Galaxy S3!)).

If you look up information on the S4 you can find articles from only 2 months ago about this powerhouse of a SoC being released and changing the game, so to speak, with the 28nm process and A15 instruction set etc.. Yet here we are acting like this is the old tech and not anything special!

Sure I'd rather have something brand new (A15 Exynos)...but 2 months isn't all that old!
 

UCLA 15

Well-known member
Apr 18, 2012
74
11
0
Visit site
Good read. The dynamic change of voltage is one reason I would not go with Tegra 3 (which, if I remember, cannot do that). Plus, I'll take the S4's aSMP over vSMP any day. I feel that aSMP wasn't implemented very well in the S3 with my Sensation, but I would imagine they have sorted out the kinks by now (apparently the 28nm process greatly helps with this). I'm definitely a fan of going with the more advanced dual-core over the less advanced quad-core. Still hoping that Adreno 3xx will bring significant improvements, though, as I feel that the GPU is often the thing that is holding these devices back the most.
 

Kevin OQuinn

AC Team Emeritus
May 17, 2010
9,267
496
0
Visit site
Good read. The dynamic change of voltage is one reason I would not go with Tegra 3 (which, if I remember, cannot do that). Plus, I'll take the S4's aSMP over vSMP any day. I feel that aSMP wasn't implemented very well in the S3 with my Sensation, but I would imagine they have sorted out the kinks by now (apparently the 28nm process greatly helps with this). I'm definitely a fan of going with the more advanced dual-core over the less advanced quad-core. Still hoping that Adreno 3xx will bring significant improvements, though, as I feel that the GPU is often the thing that is holding these devices back the most.

I believe that the Tegra does have the ability to dynamically manage cores. I know that it shuts them off if they aren't needed. The fifth companion core lets all of them shut off when the horsepower isn't needed.
 

UCLA 15

Well-known member
Apr 18, 2012
74
11
0
Visit site
I believe that the Tegra does have the ability to dynamically manage cores. I know that it shuts them off if they aren't needed. The fifth companion core lets all of them shut off when the horsepower isn't needed.

Krait can dynamically adjust voltage, which Tegra 3 can't do, which is a positive thing for battery life.
 

Kevin OQuinn

AC Team Emeritus
May 17, 2010
9,267
496
0
Visit site
Krait can dynamically adjust voltage, which Tegra 3 can't do, which is a positive thing for battery life.

Yep. It can't scale the voltage of individual cores, but it does shut them down when not needed. That's the V in the vSMP. There's a video that nVidia did that shows this in action.

From the Kal-El white paper

[NOTE]NVIDIA?s Project Kal-El processor implements a novel new Variable Symmetric
Multiprocessing (vSMP) technology. Not previously disclosed publicly, vSMP includes a fifth
CPU core (the ?Companion? core) built using a special low power silicon process that executes
tasks at low frequency for active standby mode, music playback, and even video playback. The
four main ?quad? cores are built using a standard silicon process to reach higher frequencies,
while consuming lower power than dual core solutions for many tasks. All five CPU cores are
identical ARM Cortex A9 CPUs, and are individually enabled and disabled (via aggressive
power gating) based on the work load. The ?Companion? core is OS transparent, unlike current
Asynchronous SMP architectures, meaning the OS and applications are not aware of this core,
but automatically take advantage of it. This strategy saves significant software efforts and new
coding requirements[/NOTE]
 

Trending Posts

Forum statistics

Threads
942,925
Messages
6,916,511
Members
3,158,738
Latest member
Jan