SANTA CLARA, Calif. — November 24, 2025 — Leads & Copy — AMD (NASDAQ: AMD) announced that Zyphra has achieved a significant milestone in large-scale AI model training with the development of ZAYA1, the first large-scale Mixture-of-Experts (MoE) foundation model trained using an AMD GPU and networking platform.
Zyphra’s technical report, published today, details the achievement using AMD Instinct™ MI300X GPUs, AMD Pensando™ networking, and the AMD ROCm™ open software stack.
Results from Zyphra indicate that the model delivers competitive or superior performance compared to leading open models across reasoning, mathematics, and coding benchmarks, showcasing the scalability and efficiency of AMD Instinct GPUs for production-scale AI workloads.
Emad Barsoum, corporate vice president of AI and engineering, Artificial Intelligence Group, AMD, stated that AMD’s leadership in accelerated computing empowers innovators like Zyphra to push the boundaries of AI capabilities, highlighting the power and flexibility of AMD Instinct GPUs and Pensando networking for training complex, large-scale models.
Krithik Puthalath, CEO of Zyphra, emphasized that efficiency has always been a core principle at Zyphra. It shapes how they design model architectures, develop algorithms for training and inference, and choose hardware with the best price-performance to deliver frontier intelligence to customers.
Puthalath added that ZAYA1 reflects this philosophy, and they are thrilled to be the first company to demonstrate large-scale training on an AMD platform. The results highlight the power of co-designing model architectures with silicon and systems, and they are excited to deepen their collaboration with AMD and IBM as they build the next generation of advanced multimodal foundation models.
The AMD Instinct MI300X GPU’s 192 GB of high-bandwidth memory enabled efficient large-scale training, avoiding costly expert or tensor sharding, which reduced complexity and improved throughput across the full model stack. Zyphra also reported more than 10x faster model save times using AMD optimized distributed I/O, further enhancing training reliability and efficiency. With only a fraction of the active parameters, ZAYA1-Base (8.3B total, 760M active) matches or exceeds the performance of models such as Qwen3-4B (Alibaba), Gemma3-12B (Google), Llama-3-8B (Meta), and OLMoE.1
Building on prior collaborative work, Zyphra worked closely with AMD and IBM to design and deploy a large-scale training cluster powered by AMD Instinct™ GPUs with AMD Pensando™ networking interconnect. The jointly engineered AMD and IBM system, announced earlier this quarter, combines AMD Instinct™ MI300X GPUs with IBM Cloud’s high-performance fabric and storage architecture, providing the foundation for ZAYA1’s large-scale pretraining.
For more information about how AMD is advancing AI innovation, visit www.amd.com/aiapplications
Testing details from Zyphra as of November 14, 2025, measuring the aggregate throughput of training iterations across the full Zyphra cluster measured in quadrillion floating point operations per second (PFLOPs). The workload was training a model comprised of a set of subsequent MLPs in BFLOAT16 across the full cluster of (128) compute nodes, each containing (8) AMD Instinct™ MI300X GPUs and (8) Pensando™ Pollara 400 Interconnects running a proprietary training stack created by Zyphra. Server manufacturers may vary configurations, yielding different results. Performance may vary based on use of the latest drivers and optimizations. This benchmark was collected with AMD ROCm 6.4.
David Szabados
AMD Communications
+1 408-472-2439
david.szabados@amd.com
Liz Stine
AMD Investor Relations
+1 720-652-3965
liz.stine@amd.com
Source: AMD
