AMD Powers Frontier AI Training for Zyphra
AMD (NASDAQ: AMD) announced that Zyphra trained ZAYA1, the first large-scale Mixture-of-Experts foundation model trained entirely on AMD Instinct MI300X GPUs with AMD Pensando networking and the ROCm software stack on Nov. 24, 2025.
ZAYA1-base (8.3B total, 760M active) reportedly matches or exceeds several open models across reasoning, math, and coding benchmarks and rivals Qwen3-4B and Gemma3-12B. AMD highlights the MI300X’s 192 GB high-bandwidth memory enabled simpler training (avoiding costly sharding) and Zyphra reports 10x faster model save times using AMD optimized distributed I/O. The model was trained on a jointly engineered AMD and IBM cluster using IBM Cloud fabric and storage.
AMD (NASDAQ: AMD) ha annunciato che Zyphra ha addestrato ZAYA1, il primo modello di base su larga scala Mixture-of-Experts addestrato interamente su GPU AMD Instinct MI300X con rete AMD Pensando e lo stack software ROCm il 24 novembre 2025.
ZAYA1-base (8,3 miliardi in totale, 760 milioni attivi) riferisce di eguagliare o superare diversi modelli aperti in ambiti di ragionamento, matematica e codifica e confrontarsi con Qwen3-4B e Gemma3-12B. AMD mette in evidenza la memoria ad alta velocità 192 GB del MI300X che permette un addestramento più semplice (evitando shardaggio costoso) e Zyphra riporta 10x tempi di salvataggio del modello più veloci grazie a I/O distribuito ottimizzato AMD. Il modello è stato addestrato su un cluster progettato congiuntamente da AMD e IBM utilizzando la rete IBM Cloud e lo storage.
AMD (NASDAQ: AMD) anunció que Zyphra entrenó a ZAYA1, el primer modelo de fundamento de gran escala Mixture-of-Experts entrenado completamente en GPUs AMD Instinct MI300X con redes AMD Pensando y el stack de software ROCm el 24 de noviembre de 2025.
ZAYA1-base (8,3B en total, 760M activos) supuestamente iguala o supera varios modelos abiertos en razonamiento, matemáticas y codificación y compite con Qwen3-4B y Gemma3-12B. AMD destaca la 192 GB de memoria de alta velocidad del MI300X, que facilita un entrenamiento más simple (evita un costoso sharding), y Zyphra reporta 10x más rápidos tiempos de guardado del modelo gracias a I/O distribuido optimizado por AMD. El modelo se entrenó en un clúster diseñado conjuntamente por AMD e IBM utilizando la red IBM Cloud y el almacenamiento.
AMD (NASDAQ: AMD)가 Zyphra가 ZAYA1을 훈련시켰다고 발표했습니다. ZAYA1은 AMD Instinct MI300X GPU와 AMD Pensando 네트워킹 및 ROCm 소프트웨어 스택으로 완전히 학습된 첫 대규모 Mixture-of-Experts 기초 모델이며 2025년 11월 24일에 발표되었습니다.
ZAYA1-base(총 8.3B, 활성 7.6e8) 은 추론, 수학, 코딩 벤치마크에서 여러 오픈 모델과 동등하거나 이를 상회하고 Qwen3-4B 및 Gemma3-12B와 경쟁합니다. AMD는 MI300X의 192 GB 고대역폭 메모리가 더 간단한 학습을 가능하게 했다고 강조하며(비용이 많이 드는 샤딩을 피함), Zyphra는 AMD 최적화 분산 I/O를 사용한 10배 더 빠른 모델 저장 시간을 보고합니다. 이 모델은 IBM Cloud 파브릭과 스토리지를 사용하는 IBM과 AMD가 공동 엔지니어링한 클러스터에서 학습되었습니다.
AMD (NASDAQ: AMD) a annoncé que Zyphra a entraîné ZAYA1, le premier modèle fondation Mixture-of-Experts à grande échelle entièrement entraîné sur des GPU AMD Instinct MI300X avec un réseau AMD Pensando et la pile logicielle ROCm, le 24 novembre 2025.
ZAYA1-base (8,3 Mds au total, 760M actifs) égalerait ou dépasserait plusieurs modèles publics dans des domaines tels que le raisonnement, les mathématiques et le codage, et rivalise avec Qwen3-4B et Gemma3-12B. AMD met en avant la mémoire haute vitesse 192 Go du MI300X qui permet un entraînement plus simple (évite un shardage coûteux) et Zyphra rapporte des temps d’enregistrement du modèle 10x plus rapides grâce à des E/S distribuées optimisées par AMD. Le modèle a été entraîné sur un cluster conçu conjointement par AMD et IBM en utilisant IBM Cloud Fabric et le stockage.
AMD (NASDAQ: AMD) kündigte an, dass Zyphra ZAYA1 trainiert hat, das ersten groß angelegten Mixture-of-Experts-Grundmodell, das vollständig auf AMD Instinct MI300X-GPUs mit AMD Pensando-Vernetzung und dem ROCm-Software-Stack trainiert wurde, am 24. November 2025.
ZAYA1-base (8,3B insgesamt, 760M aktiv) soll mehrere Open-Modelle in Bereichen wie reasoning, Mathematik und Codierung erreichen oder übertreffen und konkurriert mit Qwen3-4B und Gemma3-12B. AMD hebt die 192 GB hochbandbreitige Speicher des MI300X hervor, der einfacheres Training ermöglicht (vermeidet teures Sharding), und Zyphra meldet 10x schnellere Modell-Speicherzeiten durch AMD-optimierten verteilten I/O. Das Modell wurde auf einem von AMD und IBM gemeinsam entwickelten Cluster unter Verwendung des IBM Cloud-Fabrics und Speichers trainiert.
AMD (بورصة ناسداك: AMD) أعلنت أن Zyphra درّب ZAYA1، أول نموذج أساسي كبير النطاق من Mixture-of-Experts مدرّب بالكامل على وحدات AMD Instinct MI300X GPUs مع شبكات AMD Pensando وتكديس البرمجيات ROCm في 24 نوفمبر 2025.
يُزعم أن ZAYA1-base (8.3 مليار إجمالي، 760 مليون نشط) يوازي أو يتفوق على عدة نماذج مفتوحة عبر الاستنتاج والرياضيات والترميز ويتنافس مع Qwen3-4B و Gemma3-12B. تُبرز AMD أن ذاكرة الوصول العشوائي عالية النطاق الترددي 192 جيجابايت من MI300X سمحت بتدريب أبسط (وتجنب التقطيع المكلف)، وتذكر Zyphra أوقات حفظ نموذج أسرع بعشرة أضعاف باستخدام إدخال/إخراج موزع محسن من AMD. تم تدريب النموذج على عقدة مُصممة بالتعاون بين AMD وIBM باستخدام نسيج IBM Cloud والتخزين.
- First large-scale MoE trained entirely on AMD Instinct MI300X
- 192 GB high-bandwidth memory reduced sharding complexity
- 10x faster model save times with AMD optimized distributed I/O
- ZAYA1-Base (8.3B total, 760M active) matches/exceeds leading open models
- Performance claims are reported in a Zyphra technical report, not an independent third-party audit
- No financial, commercialization timeline, or deployment revenue details provided
Insights
Zyphra trained a large-scale MoE model on AMD hardware, claiming high efficiency and competitive benchmark performance.
Zyphra reports that ZAYA1 trained on AMD Instinct MI300X GPUs with AMD Pensando networking and ROCm delivered strong results. The MI300X’s 192 GB memory and AMD-optimized distributed I/O reportedly removed complex sharding, enabled 10x faster model save times, and supported a model with 8.3B total and 760M active parameters that matches several peer models.
Caveats include reliance on a Zyphra technical report and collaborative system engineering with AMD and IBM; independent reproduction and third-party benchmark validation matter. Short-term impacts center on marketing, customer proofs-of-concept, and momentum for AMD’s AI platform; watch for independent benchmarks, deployment case studies, and cluster availability over the next year, including further details in the Zyphra technical report and related blogs around
News Highlights:
- Zyphra ZAYA1 becomes the first large-scale Mixture-of-Experts model trained entirely on AMD Instinct™ MI300X GPUs, AMD Pensando™ networking and ROCm open software.
- ZAYA1-base outperforms Llama-3-8B and OLMoE across multiple benchmarks and rivals the performance of Qwen3-4B and Gemma3-12B.
- Memory capacity of AMD Instinct MI300X helped Zyphra simplify its training capabilities, while achieving 10x faster model save times.
SANTA CLARA, Calif., Nov. 24, 2025 (GLOBE NEWSWIRE) -- AMD (NASDAQ: AMD) announced that Zyphra has achieved a major 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. Using AMD Instinct™ MI300X GPUs and AMD Pensando™ networking and enabled by the AMD ROCm™ open software stack, the achievement is detailed in a Zyphra technical report published today.
Results from Zyphra show that the model delivers competitive or superior performance to leading open models across reasoning, mathematics, and coding benchmarks—demonstrating the scalability and efficiency of AMD Instinct GPUs for production-scale AI workloads.
“AMD leadership in accelerated computing is empowering innovators like Zyphra to push the boundaries of what’s possible in AI,” said Emad Barsoum, corporate vice president of AI and engineering, Artificial Intelligence Group, AMD. “This milestone showcases the power and flexibility of AMD Instinct GPUs and Pensando networking for training complex, large-scale models.”
“Efficiency has always been a core guiding principle at Zyphra. It shapes how we design model architectures, develop algorithms for training and inference, and choose the hardware with the best price-performance to deliver frontier intelligence to our customers,” said Krithik Puthalath, CEO of Zyphra. “ZAYA1 reflects this philosophy and we are thrilled to be the first company to demonstrate large-scale training on an AMD platform. Our results highlight the power of co-designing model architectures with silicon and systems, and we’re excited to deepen our collaboration with AMD and IBM as we build the next generation of advanced multimodal foundation models.”
Efficient Training at Scale, Powered by AMD Instinct GPUs
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 improving 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 further details on the results, read the Zyphra technical report, the Zyphra blog, and the AMD blog, for comprehensive overviews of the ZAYA1 model architecture, training methodology, and the AMD technologies that enabled its development.
Supporting Resources
- Follow AMD on LinkedIn
- Follow AMD on Twitter
- Read more about AMD Instinct GPUs here
- Learn more about how AMD is advancing AI innovation at www.amd.com/aiapplications
About AMD
For more than 50 years AMD has driven innovation in high-performance computing, graphics, and visualization technologies. Billions of people, leading Fortune 500 businesses, and cutting-edge scientific research institutions around the world rely on AMD technology daily to improve how they live, work, and play. AMD employees are focused on building leadership high-performance and adaptive products that push the boundaries of what is possible. For more information about how AMD is enabling today and inspiring tomorrow, visit the AMD (NASDAQ: AMD) website, blog, LinkedIn, and X pages.
Contact:
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
_________________________
1 Testing by 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.
FAQ
What did AMD announce about Zyphra's ZAYA1 model on Nov. 24, 2025 (AMD)?
How large is ZAYA1-Base and how many active parameters does it use (AMD)?
What hardware features of AMD Instinct MI300X helped Zyphra train ZAYA1 (AMD)?
What performance or efficiency gains did Zyphra report using AMD technology (AMD)?
Did AMD and Zyphra use any partner infrastructure to train ZAYA1 (AMD)?
