Since everyone seems to have an “AI processor” in their SoC, why didn’t Qualcomm add one to the 845? If you know what AI is and what it needs, SemiAccurate thinks you already know the answer to this conundrum.
As we said a few days ago, Qualcomm’s Snapdragon 845 has new and updated CPUs, GPUs, and ISPs. The changes Qualcomm made to the ARM A75/A55 cores hint at that integration, as do the integration between the GPU and ISP. This is shown off in the computational photography enhancements we talked about earlier but there is a lot more to the paradigm as well. Let’s look at the DSP, secure core, and audio this time around.
The Hexagon 685 DSP is one of those things a lot of people don’t understand or care about, mainly due to the aforementioned lack of understanding. In a modern system the DSP is critical and makes the difference between a good SoC and a bad one. This shows up mainly in energy use so it is hard for the user to tell. One way you can see the power of a DSP today is the quality of photographs in marginal situations, AI on the DSP can make a visible difference. In the future you will see this a lot more in voice recognition, AR, and other more obvious AI uses.
Yes this is the long way of saying that others have “AI cores” that they are calling out and Qualcomm doesn’t. Have a name that is, they are on their 3rd generation of DSP and tight integration for other functional units for AI uses, among other things, they just never slapped a silly name on it. If you look at teardowns and micrographs of Apple’s A11 CPU, their “Neural Engine” is pretty obviously licensed Ceva DSP cores, something confirmed by moles working in a large powered donut factory in Cupertino. Qualcomm’s DSP is both vastly more powerful and better integrated than a licensed DSP, not that Ceva’s product is bad, Qualcomm’s is just really good, well thought out, and not added on afterwards.
The important bit isn’t taking a licensed DSP and slapping it on die, it is the integration and the software, something that Qualcomm has been doing for years now. Remember Zeroth? Remember the fast paths between units on the 800-series SoCs? Qualcomm uses their Hexagon 685 DSP in conjunction with the CPU cores and GPU to do the AI jobs at hand. More important is that the Zeroth and related APIs can, in theory, shuffle work around to different units if one is busy.
This matters to users because AI is rarely a task done by itself if you are not training on a rack of servers. If you want a clean picture of something, you need AI to figure out the settings before the snap, parse the image for faces, objects that need attention, and the rest. Clearly this is a job for the DSP and GPU but what if they are doing other things like, oh, taking the picture itself? This is where flexibility and the ability to forward between units with as little latency and energy use as possible matter. You can do all of this on a CPU but not in real time and not necessarily well.
So in the end Qualcomm has a Hexagon 685, Adreno 630, and Kyro 385 but no “Neural Engine” or other specific AI core. SemiAccurate would argue that the others have a licensed DSP with nowhere near the level of integration as Qualcomm but in the end, the proof is in the results. Unfortunately those results are pretty hard to quantify at the moment but in the near future things will be much more clear. Names don’t matter in this space. OK they do for people who don’t have a clue about tech but in the real world…
On the technical front there have been changes to all three of the mentioned units to support AI better. The Hexagon 685 HVX ISA now supports INT8 instructions, the Adreno 630 has FP32 and FP16 as you would expect from a GPU, and the Kryo 385 now does FP32 and INT8. As you probably know INT8 is a very popular format for AI workloads and it’s addition to the two units above is not by chance. All in all Qualcomm is claiming the Snapdragon 845 has 3x the AI performance of the 835 and it is all done with a snappy name.S|A
Latest posts by Charlie Demerjian (see all)
- Intel kills off the 10nm process - Oct 22, 2018
- HyperX ships it’s 60 millionth enthusiast memory module - Oct 15, 2018
- Bittware/Nallatech water cools 300W of Xilinx FPGA - Oct 12, 2018
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018