PostHeaderSection

Symbol

Global NavBar

PostContentSection

Global Footer

Blog Detail Page

Beirit Harvey

For years, the highest public inference results for frontier open-source models have largely been dominated by NVIDIA-based deployments. Read how Wafer is changing the status-quo. 

Wafer Reached #1 Inference Performance for Qwen3.5-397B on AMD

For years, the highest public inference results for frontier open-source models have largely been dominated by NVIDIA-based deployments.

Wafer just changed that.

Last week, Wafer shared a major inference milestone for the AMD ecosystem, stating that they were able to optimize “Qwen3.6-35B-A3B on ATOM and found a surprising result: end-to-end, on a single 8×MI355X node, ATOM delivers 430 tok/s of single-stream decode, the highest of any public Qwen3.6-35B-A3B provider listed on Artificial Analysis, 1.07× the next best at 402 tok/s, and 2–4× the rest of the public field.” These are amazing results and findings not only for Wafer, but for the broader AMD AI ecosystem.

It shows that frontier-scale inference performance on AMD is no longer theoretical. With the right optimizations, the right infrastructure, and teams willing to work deeply at the systems level, AMD can compete at the very top end of inference serving performance. More importantly, it positions AMD as a serious contender. 


How Wafer Achieved It

Wafer’s results did not come from a single tuning pass or benchmark shortcut.

The biggest challenge with AMD inference wasn’t the hardware, it was the software stack around it. That’s where ATOM comes in.

ATOM is AMD’s inference engine built for ROCm-based AI workloads. It includes a range of optimizations designed to improve inference performance and efficiency on AMD GPUs. While AMD inference support was once seen as inconsistent and difficult to productionize, ATOM represents a significant step toward a more reliable software foundation.

The performance gains in this test did not come from custom kernel development or framework modifications by Wafer. Instead, the team focused on finding the most effective configuration of ATOM’s existing capabilities for the MI355X and this specific model. Wafer’s optimization agent explored different tuning paths within ATOM to identify the setup that delivered the strongest performance.

Wafer findings showed, “MI355x can compete and outperform Blackwells. Though the raw specs have always been up to par, it’s very promising to finally see SOTA performance achieved on the MI355x for a model as popular as Qwen3.5 397B through the use of custom kernels, speculative decoding, and other optimization techniques." 


A Strong Signal for the AMD Ecosystem

What makes this achievement important is not just the ranking itself, but what it represents.

The gap between AMD hardware capability and production inference performance is closing quickly. As optimization tooling, inference frameworks, and low-level kernel work continue improving, we expect to see more frontier workloads achieve top-tier performance on AMD infrastructure.

Wafer was able to prove that with AMD GPU infrastructure still being offered by fewer cloud providers compared to NVIDIA, the market is less saturated today. MI355X instances are still significantly less expensive than comparable NVIDIA B200 and B300 deployments. At the time of testing, MI355X bare metal pricing reached up to roughly $2.95 per GPU/hour, well below equivalent Blackwell-class systems.

By collaborating with TensorWave, whose expertise complements Wafer’s, we’ve been able to accelerate the process of unlocking the full potential of the underlying hardware across the entire system stack. CEO of Wafer, Emilio Andere, describes its optimization philosophy as pursuing ‘speed of light’ performance: the idea that with a verifiable benchmarking environment and enough iterative search, AI agents can converge toward the true theoretical limits of hardware performance. The company argues that many published kernel benchmarks are misleading because weak testing environments allow models to ‘cheat,’ producing results that appear faster than the hardware should theoretically allow.

We’re proud that Wafer’s work was powered on TensorWave’s AMD-native cloud infrastructure and early MI355X deployments. This is exactly the kind of engineering work that moves the ecosystem forward: focused optimization, deep systems expertise, and real-world performance gains on open infrastructure.

Read Wafer’s full technical breakdown to learn more about how they achieved these results: https://www.wafer.ai/blog/qwen36-mi355x

Watch CEO of Wafer, Emlio Andere’s presentation from Beyond Summit: "Serving Cheap Tokens on Non-NVIDIA Hardware." 


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. 

Wafer Reached #1 Inference Performance for Qwen3.5-397B on AMD

Product

Solutions

Resources

Company

Product

Resources

Solutions

Company