DGX Spark vs EVO-X3: Flash-Next same-prompt bench
DGX Spark (NVFP4/SGLang) vs EVO-X3 (ROCmFP4/Vulkan) on Flash-Next with the same Angry Birds and 3D architecture HTML prompts. Decode ~37 vs ~27.5 tok/s.
Bottom line. Compare boxes, not model tabs: DGX Spark ↔ EVO-X3. Today: one filled cell — Flash-Next, same prompts. Decode: Spark ~37 tok/s · EVO ~27.5 tok/s (~75%). Prefill TTFT on arch3d: EVO 1.69s vs Spark 8.84s. Stacks differ (NVFP4+SGLang vs ROCmFP4+Vulkan) — read as hardware tendency, not identical-binary A/B.
Two boxes
NVIDIA DGX Spark · GB10 · ~128GB UMA
Flash-Next NVFP4 · SGLang. Same protocol as the prior Spark Flash cell.
arch3d 37.3 tok/s · angry 36.7 tok/s.
GMKtec EVO-X3 · Ryzen AI MAX+ 395 · Radeon 8060S · ~120GB GTT
ROCmFP4-FAST GGUF · Laurent llama.cpp Vulkan RADV · :8081
arch3d 27.5 tok/s · angry 27.6 tok/s · TTFT arch 1.69s.
Same prompts · Flash-Next speed
| task | Spark TTFT | EVO TTFT | Spark decode | EVO decode | EVO/Spark |
|---|---|---|---|---|---|
| arch3d (≤12k) | 8.84s | 1.69s | 37.3 | 27.5 | ~74% |
| angry (≤7k) | 1.06s | 1.19s | 36.7 | 27.6 | ~75% |
3D architecture HTML
Angry Birds HTML
What to pick (Flash-Next only)
| Goal | Pick |
|---|---|
| decode tok/s | DGX Spark |
| AMD Strix Halo–class local + ROCmFP4 path | EVO-X3 |
| Model choice on Spark | three-model post |
Empty cells
- EVO Qwen3.8-27B ROCmFP4 + DFlash2 (redownloading)
- EVO DeepSeek-V4-Flash Q3-ROCmFP4 (split tensor count mismatch)