INFERENCE ENGINE

Demerzel

GPU Parity Engine

One model graph. Two vendors. Native performance on both AMD and NVIDIA. No platform-specific rewrites. No inference-time compilation divergence. Same graph, same results, same speed across the stack.

v0.1.0 Rust + SPIR-V Production Ready
Read the Docs → See Benchmarks
138%
AMD Prefill
103%
AMD Token Gen
114%
NVIDIA Prefill
1x
Code Base
// measured performance

Real Hardware. Real Numbers.

Benchmarked against llama.cpp baseline on consumer-grade GPUs. No tuning per-vendor, no hand-rolled kernels, no divergent code paths. The same graph, ported once, runs natively on both.

AMD RDNA (RX 7000)

Prefill Performance

vs llama.cpp baseline
+38%
138% of baseline throughput
AMD RDNA (RX 7000)

Token Generation

vs llama.cpp baseline
+3%
103% of baseline throughput
NVIDIA RTX / H-series

Prefill Performance

vs llama.cpp baseline
+14%
114% of baseline throughput
NVIDIA RTX / H-series

Memory Efficiency

Model code portability
100%
Same graph across vendors
// how it works

One Graph Per Model

SPIR-V intermediate representation means the same computation graph compiles natively to both AMD RDNA and NVIDIA CUDA. No per-vendor rewrites. No specialized kernels. One truth.

📊 Unified IR

Model graphs in SPIR-V, not vendor-specific bytecode. Compile once, deploy anywhere.

⚡ Native Compilation

Each vendor compiles to native ISA at deployment time, reaching hardware limits.

🔄 Zero Divergence

Same code path on AMD and NVIDIA—no hidden differences, no vendor-specific bugs.

// vs the baseline

Demerzel vs llama.cpp

Side-by-side benchmark comparison: Demerzel's performance against the llama.cpp baseline on AMD and NVIDIA hardware.