Benchmark HiFive Premier P550

[geerlingguy]

See: geerlingguy/sbc-reviews#65

To install Ollama on RISC-V, I followed this process:

# Install cmake
sudo apt install -y cmake

# Install Go (download link is under 'Unstable versions')
wget https://go.dev/dl/go1.24rc3.linux-riscv64.tar.gz
sudo rm -rf /usr/local/go && sudo tar -C /usr/local -xzf go1.24rc3.linux-riscv64.tar.gz
export PATH=$PATH:/usr/local/go/bin

$ go version
go version go1.24rc3 linux/riscv64

# Build ollama
git clone --recurse-submodules https://github.com/mengzhuo/ollama.git
cd ollama
go build .
sudo ln -s `pwd`/ollama /usr/local/bin/ollama

# Run ollama and test a tiny model (approx. 400 MB)
ollama serve > /dev/null 2>&1 &
ollama run qwen:0.5b

More details: geerlingguy/sbc-reviews#65 (comment) — and fully documented in my blog post: How to build Ollama to run LLMs on RISC-V Linux.

[geerlingguy]

Running benchmark 3 times using model: llama3.2:3b (13.5W)

Run	Eval Rate (Tokens/Second)
1	0.24 tokens/s
2	0.24 tokens/s
3	0.24 tokens/s
Average Eval Rate	.24 tokens/second

Running benchmark 3 times using model: llama3.1:8b (13.6W)

Run	Eval Rate (Tokens/Second)
1	0.10 tokens/s
2	0.10 tokens/s
3	0.10 tokens/s
Average Eval Rate	.10 tokens/second

Running benchmark 3 times using model: llama2:13b (13.6W)

Run	Eval Rate (Tokens/Second)
1	0.10 tokens/s
2	0.12 tokens/s
3	0.12 tokens/s
Average Eval Rate	.11 tokens/second

[geerlingguy]

On the first run-through, Deepseek wouldn't load (got Error: pull model manifest: 500: {"errors":[{"code":"INTERNAL_ERROR","message":"internal error"}), but this morning I am running it again, and the models are pulling successfully.

[geerlingguy]

Running benchmark 3 times using model: deepseek-r1:1.5b (13.5W)

Run	Eval Rate (Tokens/Second)
1	0.49 tokens/s
2	0.49 tokens/s
3	0.49 tokens/s
Average Eval Rate	.49 tokens/second

...

After four hours I gave up trying to run deepseek-r1:8b or 14b :)