Count the devices and identify the split
Treat each card-and-count combination as a separate setup. Two R9700s are distinct from one R9700, but the count alone does not explain how a model uses them. Record the parallelization method and any split proportions reported by the author. A machine containing multiple cards does not establish that every card participated in a particular run.
For example, llama.cpp distinguishes layer placement from tensor splitting. Layer mode assigns groups of layers to devices; tensor mode divides work within layers and requires communication between GPUs. The project's documentation also notes that performance depends on the split and interconnect. Preserve the source's engine version and terminology, and do not transfer a scaling claim from another backend to AMD PRO hardware.
Match the work and software
Before calculating a speed-up, match the exact model artifact, quantization, prompt and output lengths, context depth, KV-cache format and model placement. Check whether either run offloaded work to the CPU. Record the OS, driver or runtime, inference engine, backend and build version when the report provides them. The harness is the tool measuring the run; it can be different from the engine serving the model.
Keep batch size and request concurrency separate. A serving report can summarize work across simultaneous requests, while an interactive test may time one response. The vLLM benchmark client reports both throughput and latency statistics, together with concurrency. Dividing a server's total tokens per second by its card count does not recover a measured single-user generation rate.
Limit the conclusion to the evidence
Start with the two-GPU R9700 page and open the full records and source reports. Prefer a controlled comparison that changes only the GPU configuration. Check repetitions and variability before interpreting a small difference.
If the one-card setup uses a different model or CPU offload, describe the practical configurations that were measured. That can still be useful, but it does not isolate the benefit of an extra GPU. If the GPU count, split mode or workload is missing, retain that uncertainty instead of manufacturing a scaling factor.