NeuroSpark adds AMD Instinct GPUs on OCI for AI inference

News

NeuroSpark said it is using AMD Instinct GPUs on Oracle Cloud Infrastructure with ROCm software to widen hardware choices for AI workloads.

NeuroSpark AI said it has expanded its compute options by pairing AMD Instinct GPUs on Oracle Cloud Infrastructure with AMD ROCm software to serve inference workloads, a move the company said gives customers more hardware choice and a lower-cost option for some advanced models, according to AMD in a blog post published Oct. 23, 2026.

Evidence and context

In the company account relayed by AMD, NeuroSpark AI runs infrastructure for AI labs, AI-native companies and enterprises across training and inference use cases, including large language models, video generation, text-to-speech and embedding models. NeuroSpark said it manages the path from model to silicon across multiple hardware types, and that adding AMD Instinct GPUs on OCI bare metal has expanded the capacity it can offer customers.

Thomas Ning, founder of NeuroSpark AI, said in the AMD post that the company uses AMD ROCm as an open-source software base to adapt workloads to customer-specific model requirements. He said the team can fix workload-specific bugs, tune kernels and add features after establishing a baseline for a given model. The post also says NeuroSpark’s framework options include ATOM, AITER, mono-kernel, FlyDSL, composable kernel and an integration of ROCm into a mainstream inference engine.

NeuroSpark said its platform benchmarks inference and training frameworks against live workloads and iterates toward the best fit for each model. The company said model behavior can vary across GPUs depending on input patterns, workload and concurrency, which it described as a reason for matching hardware to size, latency and performance needs.

The AMD post also says NeuroSpark has ported support for several models to AMD infrastructure, including GLM-5.3 after Z.ai released the model in August 2026, and that it brought frontier open-source models such as Kimi K3, MiniMax M3 and Reflection’s Beam into production on AMD. Those claims are company statements in an AMD-published profile and were not independently verified in the material provided.

Oracle Cloud Infrastructure is described in the post as supplying the bare-metal infrastructure, capacity and support NeuroSpark needs to deploy AMD GPUs in a stable environment. The material does not provide independent performance data, customer counts, pricing, revenue impact or a verified timeline for deployment beyond the Oct. 23, 2026 publication date. Research notes also say entity identification for NeuroSpark is provisional and that the report focuses on NeuroSpark AI, the AI infrastructure company identified on its own website.