NeuroSpark Expands AI Compute Choice with AMD

NeuroSpark pairs AMD Instinct ™ GPUs on OCI with ROCm™ software to serve inference workloads and give customers more compute options.

NeuroSpark AI provides training and inference infrastructurefor AI labs, AI-native companies, and enterprises, running everything from large language models to video generation, text-to-speech, and embedding models. NeuroSpark handles the entire path for its customers, from model to silicon, on a variety of hardware, including AMD Instinct™ GPUs.

NeuroSpark’s customers are seeing their token demand explode, and for many, compute has become their highest cost. Simply obtaining necessary GPU capacity on some platforms can be difficult. New models often need more performance quality than a customer’s current GPUs can deliver. And once models reach production, they must stay reliable.

NeuroSpark helps customers overcome those limits by each model to the hardware that best fits its unique size, latency, and performance requirements. That adaptability matters because models can perform differently from one GPU to the next depending on input patterns, workload, and concurrency. Adding AMD Instinct GPUs on Oracle Cloud Infrastructure (OCI) bare metal has expanded the capacity NeuroSpark can offer customers, along with a lower-cost option for many of their most advanced models.

Adapting with open AMD ROCm Software

NeuroSpark’s agentic platform runs an iterative loop: it benchmarks inference and training frameworks against live workloads, learns from the results, and keeps converging on the best fit. The open-source nature of AMD ROCm software gives NeuroSpark flexibility to adapt the software for customer-specific model requirements. After the team establishes a working baseline for a given model, AMD ROCm lets them fix workload-specific bugs, tune kernels, and add features until the model meets the customer's performance goals. Neurospark’s framework options include ATOM (ROCm-native inference engine), AITER (ROCm-native kernel tuning & kernel library), mono-kernel, FlyDSL, composable kernel (variation of kernel selections to best fit for customer use-case) and integrated ROCm into mainstream inference engine. “AMD ROCm gives us an environment where we can develop and that we can rely on to support our customers,” says Thomas Ning, founder of NeuroSpark AI.

That openness runs in both directions. “We are collaborating with AMD to enhance ROCm and provide new software features,” says Ning. The NeuroSpark team has made many contributions to AMD ROCm, including support for new models performance improvements, and helps the integration to SGLang, a mainstream open-source inference engine. “AMD ROCm is the bedrock for us. We are able to build upon that bedrock and forge it into different forms to overcome customer challenges.”

Deploying on AMD Instinct GPUs

NeuroSpark runs AMD Instinct GPUs on OCI bare metal to serve its customers’ language, video generation, and text-to-speech models. When Z.ai released GLM-5.3 in August 2026, NeuroSpark's agentic system ported its support for GLM-5.2 to GLM-5.3 using AMD ROCm software. “GLM-5.3 is a state-of-the-art model, and we supported it on AMD from day zero, shortly after model launch,” Ning says. “With the right software orchestration built on AMD ROCm, we quickly brought frontier open-source models such as Kimi K3, MiniMax M3, and Reflection’s Beam into production serving on AMD. ”

OCI supplies the bare-metal infrastructure, capacity, and support NeuroSpark needs. “We couldn't achieve this without the reliability and support from OCI that help us to really leverage AMD GPUs in a stable and friendly environment,” says Ning. 

Empowering AI Customers

“There’s a lot of untapped potential with AMD chips, and we can help companies figure out how to best use them to achieve better performance and better economics,” Ning says. NeuroSpark is now moving toward collaborating with OCI on using various AMD Instinct Series GPUs.

By pairing AMD Instinct GPUs with AMD ROCm expertise, NeuroSpark is expanding options for its customers. “By being smart with hardware choice, we're giving power back to our customers,” Ning says.

Learn more about NeuroSpark AI at neuro-spark.ai.