Breaking the GPU-Audio Barrier: This Startup is Empowering Developers to Shape the Future of Sound

Graphics Processing Units, or GPUs, have become one of the hottest topics in tech, with industry giants like NVIDIA and AMD — along with a rising wave of startups — releasing more powerful models rapidly. While their starring role in artificial intelligence often dominates the headlines, GPUs are, in fact, much more. Born to accelerate graphics rendering for video games and 3D applications, they have evolved into massively parallel computing processors that power everything from game consoles to the post-processing of complex data captured by CT and MRI scanners.

Yet until recently, one frontier remained largely untouched — audio. For decades, professional sound processing has depended on digital signal processors (DSPs) and CPUs. These specialized chips have been the backbone of the industry, but they’re now straining under the demands of modern real-time workloads. The concept of harnessing GPUs for sound processing has floated around for years, with notable attempts from companies like Acustica Audio and various indie developers. None, however, managed to deliver the performance and stability the market demanded. GPU Audio became the first to crack the code.

Alexander Talashov, CEO of GPU Audio, set out to tackle what many saw as a dead end: running highly sequential audio workloads on hardware designed for exactly the opposite kind of computation. A side project has evolved into a full-stack solution already used in automotive and music production, with far-reaching applications still to come. GPU Audio has reimagined how real-time audio — or any dependency-aware time-series data — can be processed, making it possible to handle complex, multi-threaded audio environments on GPUs for the first time ever.

Early Fascination with Parallel Computing

Alexander Talashov’s path to this breakthrough began at one of Russia’s most prestigious technical universities, renowned for its contributions to avionics, telemetry, and telecommunications. While officially focused on Digital Signal Processing to develop next-generation Wi-Fi and 5G hardware, he found himself increasingly drawn to the less-charted territory of GPU programming.
“Around 2011, GPUs were just starting to move beyond graphics into general-purpose computing,” he recalls. “It was still this underground thing — few outside research labs knew you could program them for other tasks.” Talashov dedicated both his diploma and postgraduate theses to GPU development, gaining a foundation in the unique constraints — and possibilities — of parallel architectures.

After graduation, he took a detour into finance, accepting a risk management position at the Russian branch of one of Germany’s largest banks. While working there, he began coding high-frequency trading bots as a side project. “The solution I developed leveraged GPU acceleration for algorithm training and inference, optimizing decision-making on a daily trading cycle,” he explains. The outcome was a functional prototype that not only worked under real market pressure but also marked the start of his specialization in GPU programming.

The Question That Started It All

The turning point came when Talashov met his future co-founder, a veteran sound engineer. One afternoon, the latter asked a question that would define the next decade of Talashov’s life: why can’t GPUs be used for audio computing? Two cofounders didn’t have an answer — but were eager to find it. From 2012 to 2015, they embarked on a series of experiments to see if GPU acceleration could be applied to real-time sound. “We made countless prototypes,” Talashov recalls. “Most failed, teaching us where the real barriers were.”

During this period, Alexander Talashov also took on engineering roles in other industries. At Pickpoint, a logistics company with a network of automated parcel lockers, he created the first automated update system that could handle varying hardware, software, and firmware configurations. It slashed update times from several hours to under one hour for hundreds of terminals. “I wasn’t assigned to any project yet, but I saw this problem and thought: why not fix it?” Talashov says. It was the first time he realized he could build something that meaningfully improved systems at scale—and that realization would shape everything that came next.

Another role involved building a Secure Communication Platform for military-grade encrypted mobile messaging. Talashov co-developed the key exchange and message integrity components, meeting strict regulatory and security standards. “That project gave me confidence that I could work with high-stakes systems where failure isn’t an option,” he says.

Why Audio Was a GPU’s Worst Nightmare — Until Recently

Audio processing is fundamentally different from graphics rendering, and that difference is precisely why GPUs have historically struggled with it. Graphics processors are built to run a small number of operations across vast datasets — think millions of pixels in an image or vertices in a 3D scene.

Instead of relying on a single large dataset, real-time audio requires managing hundreds of lightweight, interdependent tasks that each operate on small datasets. Talashov explains: “When you run a game, the GPU is fully dedicated to rendering a single frame. But in audio processing, things are much more fragmented. A single song might be broken down into 30 individual tasks, each with separate left and right channels. Then, when you add GPU-powered effects to each channel, the processing load quickly scales—now you're handling 60 to 100 audio channels, with each one applying its own subset of effects."

The problem is not just architectural — it’s also mathematical. About 95% of the algorithms in audio products are based on traditional Digital Signal Processing theory. A classic example is the Infinite Impulse Response (IIR) filter, widely used in professional audio for its implementation simplicity and efficiency. However, IIR filters rely on continuous feedback loops, which means each output depends on the previous sample. That infinite feedback makes true parallelization impossible, if implemented naively.

To address this, the GPU Audio team reimagined the underlying math. They replaced traditional IIR filters with a state-space representation — recasting them as sets of first-order differential equations in matrix form, and further optimized for execution on the device within a managed environment. This representation not only emulates IIR behavior but also enables parallelization, allowing multiple computations to run simultaneously on the GPU using different execution patterns.

Breaking the GPU-Audio Mismatch

This mathematical shift was just one element of a broader overhaul. The team engineered and optimized a full stack of DSP and machine learning algorithms, purpose-built for maximum efficiency on GPUs.

This breakthrough would have been unlikely without GPU Audio’s first research scientist — an award-winning expert who joined the company in 2017. His main specialty, ray tracing, simulates light physics in 3D environments and shares a core challenge with audio: resolving chains of dependent operations in real time. Drawing on this expertise, he helped the team adapt ray-tracing scheduling logic to sound processing.

The result was GPU Audio’s patented Scheduler — “a tiny operating system for audio on the GPU,” as Talashov calls it. The Scheduler collects tasks from multiple client plugins, determines execution priorities, dispatches them across GPU cores, and returns results within sub-millisecond time. “It lets you run many independent chains of effects or graphs simultaneously,” Alexander says.

The Scheduler uses a dual-architecture design, splitting orchestration and execution between two components: the Host Scheduler on the CPU and the Device Scheduler on the GPU. By cleanly separating planning from action, this technology enables real-time, ultra-low-latency GPU audio processing — even for complex DSP chains and neural audio models once considered impossible.

Why Automotive Is the First Big Target

By 2020, the team’s prototypes for professional audio industry were ready for prime time. GPU Audio unveiled the world’s first DSP framework enabling real-time audio algorithms to run on GPUs, shared demos on YouTube and Reddit and attracted a community of tens of thousands of enthusiasts. And in 2022, GPU Audio made its public debut at NVIDIA’s GTC conference, showcasing the Scheduler. The company also managed to attract significant investments from RTP Global, other prominent funds, and former executives from Airbnb, SoftBank, Disney, Amazon, and Google.

Despite having clear potential in multiple industries, GPU Audio decided to focus on the automotive sector — and for good reason. “Car audio systems haven’t kept pace with everything else in modern vehicles,” Alexander Talashov says. “Most still run on DSPs, even as the rest of the infotainment stack has evolved dramatically.”

The limitations become obvious in everyday use. Modern cars are increasingly “rolling living rooms”. As vehicles shift toward higher levels of autonomy, driver engagement diminishes — and media consumption in the car rises. That often means multiple passengers want to listen to different content at the same time: a movie in the back, a phone call in the front, navigation guidance for the driver, and voice assistant prompts throughout.

But today’s in-car sound systems aren’t built for that complexity. If multiple streams are played simultaneously, they compete in the same physical space, bleed into each zone and create a chaotic, distracting environment.

“To enable personalized, simultaneous in-cabin audio playback across multiple zones, the system must support concurrent streams targeted at different seating areas. This requires Zone Compensation — the ability to actively suppress audio from one zone at the positions of other listeners, ensuring each passenger receives a separate, uninterrupted experience without crosstalk or interference,” Talashov explains.

GPU Audio’s brand solution, Soundscape Zone, is a GPU-powered, real-time spatial filtering technique that creates distinct, isolated audio zones inside the same cabin. The system measures the impulse responses (IRs) for a given car model, building a precise acoustic profile that accounts for cabin geometry, materials, speaker placement, and seat positions. It then generates a spatial audio model of compensation IRs and uses a spatial matrix convolver algorithm to apply it in real time, suppressing unwanted audio in each zone.

Crucially, the process requires only about 5% of available GPU compute power, which means it can run alongside other in-vehicle applications without performance loss. And because it’s a software-based solution, it can be deployed to compatible vehicles via over-the-air (OTA) updates without physical access to the device. Car OEMs now can deliver truly premium, personalized soundscapes using existing compute resources — no additional hardware required.

Not Just Selling Tools, but Powering the Products Built with Them

GPU Audio has already secured its first major customer in the automotive industry. The manufacturer plans to integrate the technology into between 100,000 and 1 million vehicles annually as it moves toward mass production. Discussions with other large-scale partners are also progressing. Still, for Talashov, the vision goes far beyond a single deal: every vehicle will need to solve the challenge of delivering a personalized audio experience — and GPU/NPU-powered solutions have far more to offer in the car audio space.

Instead of focusing solely on end-user applications, GPU Audio has begun offering a software development kit (SDK). By handling the complexity of GPU scheduling, load balancing, and memory management under the hood, the SDK enables ultra-low latency, high-channel-count audio processing for next-generation automotive systems.

“However, a car as a media capsule is just a start,” Talashov shares. “We’re creating a platform designed to power the future of audio across industries.” With C++ on both host and device sides, the SDK is designed to work across multiple GPU vendors. To bridge the differences between GPU “dialects,” the platform uses a combination of templates and a context object, standardizing development while preserving maximum performance. “We’ve made it so developers don’t have to wrestle with the quirks of each vendor’s environment — they can just focus on building the best audio experiences possible,” emphasizes Alexander Talashov.

When he reflects on the early days, the GPU Audio founder remembers the skepticism he faced: audio is sequential. It can’t be parallelized. It’s not efficient to run on a GPU. “I thought: it’s just a set of constraints. If you solve for them, you unlock something no one else has.” In doing so, he believes GPU Audio could have the same transformative impact that Unreal Engine had on graphics or Dolby Atmos had on immersive audio, fundamentally changing how people create, deliver, and in the end — experience sound.