- According to a new study, a new processing technique could double the performance speed of an existing computer while also improving efficiency during peak workloads.
- The software algorithm, which uses a technique known as simultaneous and heterogeneous multithreading (SHMT), uses multiple processing units and AI accelerators to avoid computational bottlenecks.
- This SHMT process isn’t coming to a device near you yet—the idea is still being tested to see if it’s worth implementing.
When Apple, Google, or Microsoft announce a fancy new smartphone or laptop, it’s packed with nice but admittedly small improvements—a slightly better processor, a few more megapixels, and slightly longer battery life. However, if one of these gadgets that have to boast twice the speed using only half the power? Well, it’s a technological advance worth the (inevitable) multi-million dollar ad campaign that follows.
A new breakthrough in processing promises just these computational gains through a simple software algorithm. The idea is called “simultaneous and heterogeneous multithreading” (SHMT), and it’s an advanced computing technique that basically uses different processors—including graphics processing units (GPUs), central processing units (CPUs), and even the relatively new hardware accelerators for AI (known as tensor processing units or TPUs)—on your smartphone or laptop in a more coordinated way.
This concept—researched by UC Riverside associate professor Hung-Wei Tseng—could potentially speed up existing hardware by 1.96 times with 51 percent utilization less energy. Tseng presented his work at the 56th Annual IEEE/ACM International Symposium on Microarchitecture held in October in Toronto, Canada, at the end of 2023.
“The landscape of modern computing is undoubtedly heterogeneous, as all computing platforms integrate multiple types of processing units and hardware accelerators,” the paper states. “However, entrenched programming models focus on using only the most efficient processing units for each region of code, underutilizing processing power within heterogeneous computers.”
Computers use different processing units to solve different tasks more efficiently. But this can sometimes lead to inefficiencies, as bottlenecks occur when data travels between different units. Some devices already use a technique known as simultaneous multithreading, which allows processors to run additional instructions on two separate hardware threads. Tseng’s research extends this idea to include multithreading across different (or heterogeneous) processing units.
To test this concept, the researchers used an ARM Cortex-A57 CPU, an Nvidia GPU, and a Google Edge TPU. The team then relied on a job-stealing scheduler (QAWS), according to the website Tom’s Hardware, which allowed the SMHT experiment to avoid “high error rates and evenly balance the workload among all components.” This means that tasks are reassigned to other processors if performance did not meet expectations.
Although the SMHT ran twice as fast with half the power, those figures were calculated only under the heaviest workloads. Lower workloads showed smaller gains, as there were few opportunities to process bottlenecks.
Currently, this speed-boosting and power-saving software algorithm is in its very early stages, so it’s likely it is not it’s just coming as an update in the app store. First, the software must be rewritten to take advantage of the technique—which is a many of work—and secondly, software designed in the lab does not yet meet the rigorous standards required by quality assurance departments.
But simultaneous and heterogeneous multithreading proves that even our modern processors have room for improvement. And it’s unlikely—especially in the era of AI-powered smartphones and laptops—that tech companies will leave that extra computing power on the table for long.
Darren lives in Portland, has a cat, and writes/edits about science fiction and the way our world works. You can find his previous stuff on Gizmodo and Paste if you search hard enough.