LFM2.5-8B on Ryzen AI MAX+ 395 iGPU (EvoX2) — unified-memory APU backend measurement
LFM2.5-8B-A1B (Q4_K_M, 8.47B params, 4.79 GiB) on the integrated Radeon 8060S GPU of EvoX2's AMD Ryzen AI MAX+ 395 (gfx1151, ~124GB unified memory pool visible to HIP). Generates at 150.16 tok/s and processes prompts at 3661.53 tok/s via ROCm 7.2.2. This is the only unified-memory APU in the fleet - the same machine's discrete RX 7900 XT (bm-007) is ~1.8x faster on this small model, but the APU's point is not raw speed: unified memory removes the VRAM ceiling that bounds every discrete card here. Closes the unified-memory / APU platform gap and answers the 'should I buy Strix Halo / Ryzen AI MAX for local AI?' question with first-party data.
Configuration
- Model
- LFM2.5-8B-A1B
- Artifact
- lfm2.5-8b-Q4_K_M.gguf
- Checksum
- sha256:pending-capture-on-next-load
- Quantization
- q4-k-m
- Context length
- 2,048 tokens
- Backend
- rocm
- Runtime
- llama.cpp build c57607016 (ROCm/HIP 7.2.2, gfx1151)
- Settings
- -p 512 -n 256 -ngl 99 -r 3 (llama-bench, 3 reps); HIP_VISIBLE_DEVICES=1 to isolate the iGPU. Server-mode load for VRAM/behavior capture used the same flags at --port 8199.
- 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. Same workload as bm-007 (the discrete-GPU companion on this machine) and bm-009 - directly comparable across GPUs. NOT the site's workload-pack-v1 or an end-to-end app suite; this is a backend throughput measurement. The publishable insight is the cross-GPU comparison (bm-011 vs bm-007), not a quality claim.
150.2tok/s
GEN · EvoX2
3662tok/s
PROMPT
0ms
TTFT
4.8GB
VRAM USED
| MACHINE | PROMPT tok/s | GEN tok/s | TTFT ms | VRAM GB | RAM GB | POWER W |
|---|---|---|---|---|---|---|
EvoX2 EvoX2 integrated AMD Radeon 8060S on Ryzen AI MAX+ 395 (gfx1151). HIP_VISIBLE_DEVICES=1 isolated the iGPU, which HIP reports as 'Device 0: AMD Radeon Graphics, gfx1151, VRAM: 126976 MiB' - i.e. the full ~124GB unified system memory pool is exposed as device memory. pp512 = 3661.53 ± 41.22 t/s, tg256 = 150.16 ± 0.12 t/s (3 runs each, llama-bench). Model footprint 4.79 GiB fits trivially in the pool. vram_used_gb recorded as 4.79 (the model size) because rocm-smi does not cleanly report unified-memory allocations - see limitations. Build c57607016, ROCm 7.2.2. | 3661.5 | 150.2 | 0 | 4.8 | 0.0 | 0 |
Limitations
- ◆vram_used_gb is recorded as 4.79 (the model's on-disk footprint) rather than a measured runtime allocation. On a unified-memory APU, the iGPU allocates from system RAM via the HIP runtime; rocm-smi reports GPU[1] VRAM Total as 1GB (a small carve-out) and does not reflect the model's actual unified-memory allocation. This is an inherent measurement ambiguity on APU hardware, not a data error - documented here per the honesty protocol. A follow-up using rocm-smi --showmeminfo vis_vram during a sustained generation, or a HIP memory query, would yield a more precise number.
- ◆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.
- ◆Workload is llama-bench pp512+tg256, NOT the site's workload-pack-v1. context_length is recorded as 2048 (the llama-bench default for this build; the -c flag is not supported in c57607016). LFM2.5-8B's native context is 128K - the throughput here reflects short-context inference only.
- ◆This is the unified-memory half of a same-machine comparison. The discrete RX 7900 XT result (bm-007) is ~1.8x faster on generation (269.01 vs 150.16 tok/s) and ~2x faster on prompt processing (7464.85 vs 3661.53 tok/s) for this small model. The APU's advantage is NOT speed - it is that the ~124GB unified pool can hold models no 20GB discrete card can (70B+ at Q4, large-context KV caches). A follow-up benchmark running a model too large for the discrete card is needed to demonstrate that advantage directly.
- ◆gfx1151 ROCm support: ROCm 7.2.2 detected and ran gfx1151 cleanly with no errors, but it is a newer architecture than ROCm's primary targets (gfx1100 is fully mature). Stability over long runs and at extreme context lengths is unverified - the 3-run variance here was tight (±0.12 on tg256), but sustained multi-hour loads are a separate question.
- ◆model_checksum not captured at run time (noted as pending).