Scale by Spectral Compute: Run Your CUDA Workloads Faster on Affordable AMD GPUs

When it comes to GPU computing, one name has dominated the landscape for over a decade: NVIDIA and its CUDA platform. But what if you could break free from vendor lock-in — without rewriting a single line of code?

At the 2025 Beyond CUDA Summit, Spectral Compute made waves by unveiling Scale, a revolutionary new technology that allows developers to run native CUDA applications on AMD GPUs — faster, cheaper, and without compromise.

By compiling directly to AMD hardware, Scale unlocks a future where AI builders, HPC engineers, and researchers can double their GPU options, cut costs, and accelerate performance — all while staying inside the familiar CUDA ecosystem they've relied on for years.

Here’s what you need to know about this groundbreaking announcement — and why it’s set to change the future of GPU computing.

🔧 The Problem: CUDA Lock-In and Limited Hardware Options

Michael opened by addressing the elephant in the room: CUDA is amazing, but it’s also created a hardware monopoly.

Today, if you’re building in CUDA, you’re locked to NVIDIA GPUs — or forced to juggle multiple codebases to support alternatives like AMD.

This bottleneck has kept developers from fully tapping into new, faster, and more affordable hardware options.

Scale changes all that.

⚡ Introducing Scale: Native CUDA on AMD Without Emulation

Unlike translation layers or emulation systems, Scale compiles CUDA directly to AMD’s architecture — meaning:

• Zero emulation overhead
• Direct performance on MI300X, MI325X and other AMD GPUs
• Stay inside the CUDA ecosystem you already know

Instead of rewriting thousands of lines of code, developers can recompile and run — just like switching between x86 and ARM on CPUs.

🛠️ Full Compatibility Goals: CUDA Math, Runtime, and Driver APIs

Spectral Compute isn’t just stopping at basic functionality.

Their roadmap includes:

• ~90% support for CUDA C Core Math APIs
• ~70% support for CUDA Runtime and Driver APIs
• Full NVCC C++ semantic support for seamless compilation

Even inline PTX assembly is supported — and in some cases, runs faster on AMD than NVIDIA hardware itself.

📊 Performance Gains: Scale vs HIP on AMD MI300X

Michael shared fresh benchmarks using the Rodinia Benchmark Suite:

• Scale achieved up to 2x faster performance compared to HIP
• Tests were run on RunPod.io using AMD’s MI300X GPUs
• Performance wasn’t even the focus yet — initial priority was compatibility!

💡 Once optimization becomes the focus later this year, even bigger speedups are expected.

🧠 Warp Sizes, Inline Assembly, and Real Optimizations

One big technical hurdle: NVIDIA and AMD handle warp sizes differently.

• NVIDIA: Warp size = 32
• AMD: Warp size = 64

Scale smartly maps CUDA’s 32-thread assumptions onto AMD’s 64-thread warps, allowing most code to run without modification.

Even better?
Spectral Compute is making improvements to inline PTX — sometimes producing better machine code for AMD than NVIDIA itself provides for CUDA.

🌎 Open Collaboration: How You Can Get Involved

Spectral Compute is building this technology openly and wants the community involved:

• Help test Scale and report issues
• Contribute to open-source reimplementations of CUDA libraries
• Extend the ROCm ecosystem together

👉 Join their Discord to collaborate, share feedback, or even send “hate mail” if something isn’t working. (Yes, they literally asked for it.)

💬 Michael's message: "We want to target all the hardware — not just some of it."

It’s a game-changer for AI, HPC, and scientific computing — making high-performance compute more accessible, flexible, and cost-effective than ever before.

📺 Watch the Full Talk 👉 Watch Michael Søndergaard’s Presentation at Beyond CUDA Summit

🚀 Deploy Faster with AMD GPUs 👉 Build, train, and infer at scale using AMD-powered AI infrastructure through TensorWave.

About TensorWave

TensorWave is the AI and HPC cloud purpose-built for performance. Powered exclusively by AMD 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.