Why you can trust Tom's Hardware Our expert reviewers spend hours testing and comparing products and services so you can choose the best for you. Find out more about how we test.
Gauging AI performance, especially for LLMs, is a broad discipline with a dizzying number of permutations thanks to the vast range of models, model runners, and possible workloads available.
We can’t possibly cover every one of those permutations in this article, but after much research and experimentation, we are at least confident that we’re able to provide some useful perspective on what the Spark can do.
We’ll start with LLM inference performance, since that’s where most enthusiasts dipping their toes into local AI are focused nowadays. We chose llama.cpp as our inference platform since it’s broadly compatible across platforms and offers a well-documented and tunable benchmarking interface.
Other model runners, like vLLM, SGLang, and Nvidia’s own Triton inference server, have their own pros and cons and their own performance characteristics, but we’re not benchmarking them against each other today.
We’ll start off our benchmarks using Meta’s llama-3.1-8B, a dense LLM (i.e., one for which all parameters are activated per token) whose behavior is well-understood in the current AI landscape.
We extended our dense model testing using Google’s Gemma 3 12B and 27B.
We’ll also look at performance with OpenAI’s gpt-oss 120b and Qwen3-30B-A3B models, which use mixture-of-experts (MoE) architectures that are closer to the state of the art in LLM research.
MoE models can have massive parameter counts, but only a portion of those parameters are activated for any given token. That allows MoEs to balance both capability and performance, although they can still require large amounts of memory. GPT-OSS 120b, for example, requires about 60 GB of RAM.
Speaking generally, the Spark’s higher raw compute horsepower makes it much faster for prompt processing than Strix Halo, meaning that your time to first token will generally be lower on the Spark, especially as context lengths grow. (For reference, a context length of 32,768 tokens is about 25,000 words.)
Once the tokens start rolling, the Spark is generally faster at churning them out than Strix Halo, too—especially at long context lengths.
The closest results between the platforms come from the dense Gemma 3 27B model, where both GB10 and Strix Halo seem to struggle with the sheer number of active parameters per token. But on every other test, the Spark holds the lead, and at every context depth.
LLMs are just one usage for AI compute, of course. Nvidia also positions the Spark as a companion accelerator for creative users who primarily work on laptops or Macs but need to offload tasks like image generation to a more powerful or versatile GPU without upgrading their entire system or migrating to an unfamiliar platform.
We tested the latest Flux.2 Klein 9B image generation workflow in ComfyUI using the same random seed as a base. After the first load for the workflow, the Spark generated iterations on our desired image in just over 60 seconds, while the Strix Halo platform took a whopping four times as long.
If iteration on generative imagery matters to you, the Spark can do the job with practically any image model thanks to its generous VRAM pool. Discrete GPUs can be much faster still, but you have to match the amount of available VRAM to your workload. And have you tried to buy anything more powerful than an RTX 5070 recently?
The brand-new LTX-2 video generation model also worked flawlessly on the Spark via ComfyUI as soon as I was able to load it up. After the first load, I was able to generate iterations on short 720p videos locally in just over three minutes.
Even with the latest ROCm 7.1.1 stack, however, trying to run the same LTX-2 workflow on our Strix Halo system resulted in HIP errors that would hang ComfyUI or crash the entire GNOME desktop environment.
That experience emphasizes a major advantage of the Spark as of this writing: stuff just works. If you need to try a particular workflow, you can generally rest assured that the Spark will run it on day one.
AMD has repeatedly and publicly committed to making ROCm a better AI platform from the desktop to the data center, but that improvement will take time, and the company clearly has its work cut out for it in this department.
MORE: Best Graphics Cards
MORE: GPU Benchmarks and Hierarchy
MORE: All Graphics Content
