In the ARM ecosystem, Apple has emerged as one of the only companies building its own high-performance custom CPUs. Other onetime rivals, like Qualcomm, have taken to licensing parts from ARM instead of building their own architectures. The field, however, isn’t empty — Samsung also has a custom CPU designed according to its own criteria. The reason you don’t hear as much about their chips is that previous efforts haven’t been all that great at competing with what Apple and ARM have brought to the table. The company is taking another crack at the topic, however, this time with a new CPU core, new cluster design, and a custom NPU (Neural Processing Unit) baked into the chip.
We don’t have a lot of the salient details, but here’s what we know so far. The Exynos 9820 will integrate a brace of Samsung M4 chips (these are its own custom architecture), two Cortex-A75s, and a quad-core of Cortex-A55s. This type of Big.Little.littlest configuration is more common than it used to be, as companies experiment with deploying multiple types of cores optimized to very specific frequency ranges. The theoretical upside of this approach is that you can optimize each core more narrowly and specifically, trading die size for efficiency. This is also a potentially effective way to offset the dark silicon trend, in which manufacturers can put more transistors on a die but lack the power budget to run the entire chip at full power simultaneously. The downside, at least in theory, is that moving workloads back and forth across cores or attempting to spread a task across a larger group of heterogeneous cores could cost you additional power best spent elsewhere.
The Samsung Exynos 9820 is reportedly targeting a 20 percent uplift in single-core performance at the same power or a 40 percent reduction in power consumption. That’s critical for the chip, considering its predecessor ran into real problems with power efficiency. With the M3, Samsung built a core that couldn’t quite compete with Qualcomm when you cranked up the clock (while completely blowing the power budget) or an SoC that delivered real improvements compared with the M2 in terms of power efficiency and performance but wasn’t competitive with other SoCs. The PR statement from Samsung appears to contain a syntax error stating the opposite of what Samsung meant, specifically:
With an enhanced architecture design, the Exynos 9820’s new fourth-generation custom core delivers around 20-percent improvement in single core performance or 40-percent in power efficiency when compared to its predecessor which can load data or switch between apps much faster.
Presumably it’s the Exynos 9820 that can load data or switch between apps much more quickly than the Exynos 9810 rather than the other way around. Anandtech notes that the Exynos 9810 put its Cortex-A55 cores on a separate cluster from the M3 cores, which meant cache coherency ran through the interconnect, in a manner similar to early big.Little designs. The 9820 fixes this, which should yield the improvements mistakenly attributed to the 9810 above.
Whether these gains will be enough to put the chip in better competition with Qualcomm or Apple remains to be seen. Samsung only predicts a 15 percent improvement in multi-threaded performance; it didn’t give any data on frequency, caches, or L2 size; and we don’t yet know what kind of performance improvements the NPU will bring to the table in real-world scenarios (Samsung claims a 7x performance improvement for the NPU compared with the CPU for AI tasks). Still, the company is clearly dedicated to improving its own custom foundry efforts and ensuring there’ll continue to be custom CPU designs in mobile on both Android and iOS devices. Samsung hasn’t been doing this nearly as long as Apple has — hopefully its CPUs continue to improve.
People Are Using a Neural Network App to Create Fake Celebrity Porn
Machine learning has become so advanced that a handful of developers have created a tool called FakeApp that can create convincing "face swap" videos. And of course, they're using it to make porn.
MIT Neural Network Processor Cuts Power Consumption by 95 Percent
MIT has developed new neural network processing methods that could cut the power consumption of existing solutions by up to 95 percent. It's a sea change that could effectively reinvent the nature of AI — and where such workloads are performed.
Google Neural Network Can Isolate Individual Voices in Videos
In a new feat of machine learning magic, Google Research has developed a system that can replicate the "cocktail party effect," where your brain focuses on a single audio source in a crowded room.
IBM Aims To Reduce Power Needed For Neural Net Training By 100x
Custom silicon for speeding up AI inferencing is here, but IBM wants to go further and use a hybrid computing architecture and elements of resistive computing to also improve the efficiency of training neural networks.