PostHeaderSection

Symbol

Global NavBar

PostContentSection

Global Footer

Blog Detail Page

Beirit Harvey

Read how Spectral Compute used TensorWave’s AMD-native AI cloud and MI355X infrastructure to benchmark CUDA workloads on AMD GPUs with improved portability and performance.

Spectral Compute Runs Unmodified CUDA Workloads on TensorWave’s Infrastructure with Up to 25.7× Faster Performance

“The MI355X doesn't get nearly the recognition it deserves. Most of the conversation around AMD's latest hardware is about AI inference, but our benchmarks on TensorWave show it's equally formidable for simulation, HPC, and the kind of compute-intensive workloads that have been running on CUDA for decades — and our Rodinia benchmark results on TensorWave's infrastructure prove that.”

— Michael Søndergaard, CEO, Spectral Compute


About Spectral Compute

Spectral Compute is working to reduce friction between CUDA software and modern GPU hardware platforms. Its SCALE compiler toolkit allows developers to run CUDA applications for multiple GPU architectures without rewriting code, helping organizations reduce vendor dependency while preserving existing CUDA investments.

The company validates SCALE against real-world AI, HPC, and simulation workloads, ensuring production-grade compatibility and performance across modern GPU platforms including AMD Instinct™ GPUs.


The Challenge

GPU development has long lagged behind CPU development in portability, tooling, and developer experience. For organizations running CUDA-based workloads, dependence on a single hardware ecosystem creates operational and infrastructure constraints that become more significant as GPU demand scales.

Spectral Compute was founded to address that problem through SCALE, which enables developers to write CUDA code once and run it across GPU hardware from multiple vendors.

To validate SCALE against real-world workloads, the team needed consistent access to the latest AMD GPU infrastructure capable of supporting large-scale correctness validation and benchmarking. The company regularly tests over 50 real-world CUDA projects including PyTorch, llama.cpp, vLLM, GROMACS, OpenCV, Blender Cycles, FastEddy, and other HPC and AI frameworks.

Before working with TensorWave, obtaining reliable access to cutting-edge AMD GPU capacity at the scale required for reproducible benchmarking was a bottleneck.


Why TensorWave

For Spectral Compute, selecting a cloud provider was not simply about GPU availability. Their validation and benchmarking workflows required reproducible environments, low-level hardware access, and infrastructure stable enough to generate meaningful performance data across real-world CUDA workloads.

TensorWave stood out because of its deep operational focus on AMD GPU infrastructure and its ability to rapidly deploy and operationalize the latest AMD Instinct™ systems in production environments. 

While many providers announce support for new hardware generations, TensorWave has consistently been among the first to get platforms like the AMD Instinct MI355X running at scale with production-ready infrastructure behind them.

For Spectral Compute, that mattered. Their work depends on validating SCALE against the newest AMD architectures as early as possible, while maintaining the consistency and low-noise performance required for benchmarking and compiler validation.

According to Michael Søndergaard:

“When you're benchmarking SCALE against native HIP codebases, you want a compute partner who is genuinely AMD-native — one whose infrastructure, expertise, and roadmap are built around AMD GPU computing rather than bolted on alongside other offerings.”

The Spectral Compute team also required:

 * Reliable and reproducible benchmarking environments
 * Bare metal-style access for low-level validation work
 * High-memory GPU infrastructure
 * Engineers deeply familiar with the AMD GPU software stack
 * Fast access to the latest AMD hardware generations

TensorWave met those requirements while also collaborating directly with Spectral Compute on BIOS and kernel-level tuning needed to optimize their workloads. Rather than simply provisioning infrastructure, the TensorWave team worked alongside Spectral Compute to help validate and tune performance on MI355X systems. The TensorWave team provided hands-on engineering collaboration at a level rarely available from traditional cloud providers.


The Solution

Spectral Compute deployed its SCALE validation and benchmarking workflows on TensorWave’s AMD Instinct MI355X infrastructure.

Their validation pipeline compiles and tests dozens of unmodified CUDA applications directly on AMD GPUs using SCALE. This allows Spectral Compute to verify correctness across AI frameworks, HPC applications, computer vision libraries, and specialized GPU workloads without requiring developers to rewrite code for AMD hardware.

The second component of the deployment focused on performance benchmarking. Spectral Compute used industry-standard benchmark suites including Rodinia and CUDA Handbook benchmarks to compare SCALE against native HIP implementations on the same hardware.

One particularly important area of focus was CUDA managed memory. Managed memory enables CPU and GPU memory access without explicit memory transfer management by developers. While NVIDIA’s implementation has historically delivered strong performance, HIP-based managed memory migration on AMD hardware previously introduced major first-access latency penalties that made some workloads impractical.

Using TensorWave’s MI355X infrastructure, Spectral Compute optimized this workflow significantly with SCALE 1.6.1.

In benchmark testing involving a 4 GiB managed allocation:

 * HIP achieved initial migration throughput as low as ~7 MB/s
 * SCALE improved throughput by roughly 100×
 * Initial migration time dropped from more than 10 minutes to approximately 6 seconds

This improvement made managed memory a viable feature for CUDA portability on AMD GPUs and represented a major milestone toward reducing migration friction for CUDA-based workloads.


Results

Spectral Compute’s benchmarking results on TensorWave infrastructure demonstrated both portability and performance improvements for CUDA workloads running on AMD GPUs.

Using the Rodinia benchmark suite on TensorWave MI355X instances:

 * SCALE outperformed native HIP implementations in 10 out of 14 workloads
 * Average performance improvement reached 5.94×
 * Peak improvement reached 25.7× faster performance for dwt2d

The results reinforced one of SCALE’s core goals: enabling organizations to run existing CUDA code on AMD hardware without modification while maintaining, and in many cases improving, workload performance.

TensorWave also accelerated Spectral Compute’s engineering workflows.

Reliable access to MI355X infrastructure allowed the company to:

 * Shorten validation cycles
 * Expand workload coverage across real-world CUDA projects
 * Catch correctness issues earlier in development
 * Produce reproducible benchmark data for customer validation
 * Benchmark continuously against the latest AMD hardware generations

Michael Søndergaard notes:

“TensorWave has been a fantastic partner. Their team was genuinely engaged from the start — willing to dig into the kind of low-level system configuration work that most cloud providers wouldn't entertain.”


Looking Ahead

As SCALE expands its CUDA API coverage and support for increasingly complex workloads, Spectral Compute expects its infrastructure requirements to grow alongside it.

Future validation efforts will require:

 * Broader library and framework coverage
 * Larger-scale performance benchmarking
 * Continuous testing against new AMD GPU generations
 * Validation across increasingly demanding HPC and AI workloads

TensorWave’s continued investment in AMD Instinct infrastructure positions the partnership to scale alongside those needs.

For Spectral Compute, the long-term direction is clear: enabling organizations to run CUDA workloads beyond a single-vendor ecosystem while maintaining the performance, tooling, and developer experience they expect.


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. 



Spectral Compute Runs Unmodified CUDA Workloads on TensorWave’s Infrastructure with Up to 25.7× Faster Performance

Product

Solutions

Resources

Company

Product

Resources

Solutions

Company