Beirit Harvey

At this year’s Beyond Summit, Spectral CEO Michael Søndergaard made a point that cut through a lot of the usual infrastructure debate: You shouldn’t have to pick a side.

Run CUDA Across AMD and NVIDIA: What Spectral Compute Announced at Beyond Summit

At this year’s Beyond Summit, Spectral Compute CEO Michael Søndergaard made a point that cut through a lot of the usual infrastructure debate: You shouldn’t have to pick a side.

For years, teams have been forced into a simple decision. Build on CUDA and stay within NVIDIA’s ecosystem, or move to alternatives and absorb the cost of migration, rewrites, and performance uncertainty.

In a single update, Spectral Compute introduced:

 * Native support for both AMD and NVIDIA GPUs
 * ~6× performance

At this year’s Beyond Summit, Spectral Compute CEO Michael Søndergaard made a point that cut through a lot of the usual infrastructure debate: You shouldn’t have to pick a side.

For years, teams have been forced into a simple decision. Build on CUDA and stay within NVIDIA’s ecosystem, or move to alternatives and absorb the cost of migration, rewrites, and performance uncertainty.

In a single update, Spectral Compute introduced:

 * Native support for both AMD and NVIDIA GPUs
 * ~6× performance gains on AMD Instinct MI355X
 * A new language server for better GPU debugging
 * A roadmap into AI frameworks like PyTorch, vLLM, and SGLang

The message was clear. What Spectral Compute is building challenges that assumption at the compiler level.


The Shift: From Translation to Compilation

Spectral Compute's Scale toolkit started with a clear goal: run unmodified CUDA code on AMD GPUs. That alone addressed one of the biggest sources of lock-in. But the latest update moves beyond that. Scale now supports NVIDIA as well.

At first glance, that sounds circular. A CUDA-based toolkit targeting CUDA hardware. But the architecture matters. This is not a translation layer. It is a full compiler stack, from CUDA C++ down to the target GPU. That distinction changes the outcome.

Instead of maintaining separate toolchains for different vendors, Scale applies optimizations once and lets them compound across both AMD and NVIDIA environments. The result is not portability with compromise, but portability with shared performance gains.


Performance That Carries Across Environments

Using the Rodinia benchmark suite, Spectral Compute reported:

 * ~6× average performance improvement on MI355X
 * Early parity with NVIDIA’s own compiler

These improvements are not tied to a single vendor path. They come from compiler-level optimizations that apply across architectures. That means performance scales with the toolchain, not just the hardware.

Over time, that shifts where teams look for leverage.


Optionality Without Rewrites

The core idea is simple. Write CUDA once, run it anywhere, and improve it over time. Instead of forcing teams to migrate stacks or maintain parallel codebases, Scale keeps the CUDA path intact while removing the dependency on a single vendor.

That turns multi-vendor from a strategy into something operational.


Moving Into AI

Until now, Spectral Compute has focused on HPC workloads. In 2026, the focus shifts to AI.

Support for PyTorch is coming first, followed by vLLM and SGLang. The goal is to extend the same model, one code path across vendors, into both training and inference.

There is a broader implication here. For years, performance gains in AI have been tied to hardware cycles. New GPUs, new architectures, incremental improvements in frameworks.

What Spectral Compute is pointing to is a different layer of leverage.

Compiler-driven performance that compounds over time.Portability that does not require rewrites.Optionality that exists without operational overhead.

It does not remove the importance of hardware. But it reduces how much teams are constrained by it. At a time when infrastructure decisions can define long-term cost, flexibility, and velocity, that shift matters.


What Actually Changes

This is not about replacing CUDA, it’s about decoupling it from a single ecosystem.

When performance improvements come from the compiler and portability comes without rewrites, the constraint shifts. Teams can choose infrastructure based on cost, availability, and fit, not just compatibility. That is a meaningful change in how AI systems get built.

Watch the full Beyond Summit 2026 session with Michael Sondergaard here:


TensorWave Cloud: Engineered for AI Velocity

TensorWave is the industry’s premier AMD-exclusive cloud provider, delivering memory optimized AI infrastructure for training, fine-tuning and inference. 

Our platform combines the raw computational power of AMD's latest GPU architectures with innovative liquid cooling technology and ultra-high-speed networking, creating an optimized environment for the most demanding AI workloads. From large language model training to real-time inference, TensorWave provides the performance, scalability, and cost-efficiency that modern AI initiatives demand.

Ready to get started? Connect with a sales engineer. 

AMD Instinct™ MI355X GPUs Now Available

Run CUDA Across AMD and NVIDIA: What Spectral Compute Announced at Beyond Summit

Subscribe to our Blog

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

Product

Resources

Company

© 2026 TensorWave Inc. - All rights reserved.