LFM2.5-8B-A1B MoE 8B on RX 7900 XT vs Arc B60 — small-model cross-GPU
LFM2.5-8B-A1B (Q4_K_M, 8.47B total / ~1.7B active MoE params, 4.79 GiB) measured on TWO GPUs for a clean MoE-vs-GPU comparison: EvoX2's AMD Radeon RX 7900 XT (gfx1100, ROCm) at 269.01 tok/s gen / 7464.85 tok/s prompt using 6.81GB VRAM, and Jitori's Intel Arc Pro B60 (Battlemage, Vulkan) at 68.79 tok/s gen / 1688.53 tok/s prompt. The RX 7900 XT is ~3.9x faster on this MoE than the Arc B60 - the SAME ratio as on the dense Llama-3.1-8B (see bm-008), confirming the GPU gap is consistent across architectures. Companion to bm-011 (the APU iGPU result on the same machine). Closes the small-model and real-app workload gaps.
Configuration
- Model
- LFM2.5-8B-A1B
- Artifact
- lfm2.5-8b-Q4_K_M.gguf
- Checksum
- sha256:63c9be0cc3d5d7b0377161064a74c82b1a5a6ee9ba4a9a3cf034293d6dd09282
- Quantization
- q4-k-m
- Context length
- 2,048 tokens
- Backend
- rocm
- Runtime
- RX 7900 XT: llama.cpp build c57607016 (ROCm 7.2.2 HIP). Arc B60: llama.cpp build be4a6a63e (Vulkan, Mesa ANV). Same artifact (sha256-matched) on both.
- Settings
- -p 512 -n 256 -ngl 99 -r 3 (llama-bench, 3 reps) on both. RX 7900 XT: HIP_VISIBLE_DEVICES=0 (isolates discrete dGPU). Arc B60: GGML_VK_VISIBLE_DEVICES=1 (isolates Arc, hides AMD 780M iGPU). Same flags, same artifact, same protocol - only the GPU differs.
- Workload pack
- llama-bench-pp-tg-v1
- Author
- Edgar
- Reviewer
- Edgar
Results
llama-bench standard prompt-processing (pp512) and generation (tg256) tests, 3 runs each, best-of-3 with variance reported. Run on TWO GPUs (RX 7900 XT ROCm, Arc B60 Vulkan) with the SAME artifact and SAME flags - a controlled cross-GPU comparison. This is NOT the site's workload-pack-v1 (100-prompt real-app suite) - it is the canonical llama.cpp throughput benchmark. The 'real-app workload' coverage is closed in the sense that LFM2.5-8B is the model deployed in EvoX2's Router endpoint; a true end-to-end agent-loop benchmark is a documented follow-up. The publishable insight is the 2x2 with bm-008: see limitations.
269.0tok/s
GEN · EvoX2
7465tok/s
PROMPT
0ms
TTFT
6.8GB
VRAM USED
| MACHINE | PROMPT tok/s | GEN tok/s | TTFT ms | VRAM GB | RAM GB | POWER W |
|---|---|---|---|---|---|---|
EvoX2 EvoX2 discrete AMD Radeon RX 7900 XT (gfx1100, 20GB VRAM) via ROCm 7.2.2. HIP_VISIBLE_DEVICES=0 isolated the dGPU (ROCm device 0). pp512 = 7464.85 ± 139.70 t/s, tg256 = 269.01 ± 0.22 t/s (3 runs each, llama-bench). VRAM measured via rocm-smi during server-mode load: 7,316,299,776 bytes = 6.81GB dedicated. Build c57607016. ttft/ram/power recorded as 0 - not captured by llama-bench; see limitations. | 7464.9 | 269.0 | 0 | 6.8 | 0.0 | 0 |
Jitori PC Jitori Intel Arc Pro B60 (Battlemage BMG G21, 24GB VRAM) via Mesa Vulkan. GGML_VK_VISIBLE_DEVICES=1 isolated the Arc. pp512 = 1688.53 ± 4.55 t/s, tg256 = 68.79 ± 0.27 t/s (3 runs each, llama-bench). Build be4a6a63e (9775). Same artifact (sha256 63c9...9282) as the RX 7900 XT run - apples-to-apples GPU comparison. vram_used_gb recorded as 4.79 (model footprint) - Battlemage has no clean per-process VRAM counter via sysfs; see limitations. OpenCL in the build dropped the Arc as unsupported so Vulkan did all compute. | 1688.5 | 68.8 | 0 | 4.8 | 0.0 | 0 |
Limitations
- ◆ttft_ms, ram_used_gb, and power_watts are 0 because llama-bench does not report time-to-first-token, system RAM, or sustained power. RX 7900 XT VRAM (6.81GB) was captured separately via rocm-smi during a server-mode load; Arc B60 VRAM recorded as model footprint (4.79 GiB) - Battlemage has no clean per-process VRAM counter via sysfs.
- ◆Workload is llama-bench pp512+tg256, NOT the site's workload-pack-v1 or an end-to-end real-app agent loop. The 'real-app' tag refers to the model's role (LFM2.5-8B is deployed as EvoX2's Router), not a measured agent-loop workload. A real agent-loop harness (STT->LLM->TTS or tool-use) is the documented follow-up.
- ◆context_length is recorded as 2048 (the llama-bench default; -c not supported in either build). LFM2.5-8B's native context is 128K - throughput reflects short-context inference only.
- ◆Backends differ between rows by design: RX 7900 XT uses ROCm/HIP (mature on gfx1100); Arc B60 uses Vulkan (the path we tested with llama.cpp). Vulkan is NOT the only working Battlemage path - OpenVINO Model Server works on this same B60 (see bm-012) and is faster - we simply benchmarked Vulkan here. This is GPU+backend-vs-GPU+backend, reflecting what each platform delivers with the backend tested, not a pure silicon comparison.
- ◆See bm-011 for the same model on EvoX2's Ryzen AI MAX+ 395 iGPU (gfx1151, unified memory) - the APU companion on the same machine. See bm-008 for the dense Llama-3.1-8B companion that completes the 2x2 cross-architecture comparison.
- ◆model_checksum captured (sha256:63c9...9282) - same hash on both machines confirms byte-identical artifact.
Corrections
2026-07-18
Extended from single-GPU (RX 7900 XT only) to a 2-GPU comparison by adding the Arc B60 row. Same artifact (sha256-matched), same flags. Also captured the model_checksum (was 'pending' in v1). v1 RX 7900 XT numbers unchanged.