Alex Medick

Cut AI training costs by switching from NVIDIA to AMD GPUs - boosting performance, memory, and efficiency at scale.

How Zyphra is Cutting AI Training Costs with AMD GPUs

Training cutting-edge AI models isn’t cheap. For startups competing with well-funded tech giants, the cost of compute can make or break innovation. Enter Zyphra, a company building Maya OS, a multimodal agent system for enterprises. Their secret to slashing training costs? Moving from NVIDIA to AMD GPUs.

On this episode of Beyond CUDA, host Jeff Tatarchuk sits down with Quentin Anthony, Model Training Lead at Zyphra, to talk about distributed training at scale, porting kernels to AMD hardware, 

Training cutting-edge AI models isn’t cheap. For startups competing with well-funded tech giants, the cost of compute can make or break innovation. Enter Zyphra, a company building Maya OS, a multimodal agent system for enterprises. Their secret to slashing training costs? Moving from NVIDIA to AMD GPUs.

On this episode of Beyond CUDA, host Jeff Tatarchuk sits down with Quentin Anthony, Model Training Lead at Zyphra, to talk about distributed training at scale, porting kernels to AMD hardware, and why memory and cost advantages are pushing more teams to explore alternatives to CUDA.









Key Takeaways

1. From Quantum Labs to AI at Scale

Quentin’s path started in physics and quantum computing before shifting to high-performance computing (HPC). His background in distributed systems and networking gave him the expertise to tackle large-scale AI training challenges.

2. Frontier Supercomputer Experience

At Oak Ridge National Lab, Quentin worked on porting FlashAttention to AMD’s MI250 GPUs - navigating unique multi-level interconnect topologies and memory-sharing challenges. That early AMD experience set the stage for Zyphra’s migration strategy.

3. Building Personalizable Models for Enterprise

Zyphra’s mission with Maya OS is to create inference-efficient, multimodal models that work across heterogeneous hardware - from iPhones to large GPU clusters. This requires new hybrid architectures (Mamba + Transformer) optimized for both performance and portability.

4. Why Zyphra Chose AMD

Training on AMD MI300X GPUs offers tangible cost savings and technical advantages:

 * More VRAM per GPU reduces the need for complex parallelism.
 * Larger batch sizes speed up training and cut development time.
 * Lower cost per GPU compared to NVIDIA alternatives.

5. Overcoming the “Switching Costs” Myth

While developers worry about ecosystem maturity, Quinton argues the barriers are overstated. Once initial kernel work is complete, ROCm software is far more usable than many assume, and AMD’s developer support continues to improve rapidly.

6. The Bigger Picture

For Quinton, Beyond CUDA means pushing past monopoly thinking. AMD, Cerebras, Groq, and others all provide viable alternatives. Ignoring them isn’t just shortsighted—it could be a costly mistake for teams operating at scale.


Wrapping Up

Zyphra proves it’s possible to train at scale without burning cash. By leveraging AMD GPUs, they’re cutting costs, improving performance, and opening new doors for AI innovation.

If your team is ready to explore AI training on AMD Instinct GPUs, get started with TensorWave today and see what’s possible beyond CUDA.


About TensorWave

TensorWave is the AI GPU cloud purpose-built for performance. Powered exclusively by Instinct™ Series GPUs, we deliver high-bandwidth, memory-optimized infrastructure that scales with your most demanding models—training or inference.

Ready to get started? Connect with a Sales Engineer.

TensorWave Welcomes the AMD Instinct™ MI355X

How Zyphra is Cutting AI Training Costs with AMD GPUs

Subscribe to our Blog

Stay ahead of the curve with the latest in AI, AMD accelerators, and all things TensorWave.

Product

Resources

Company

© 2025 TensorWave Inc. - All rights reserved.