01 // Inference benchmarks
Single-stream decode · llama.cpp
# env llama.cpp b4732 · 4096 ctx · batch=1 · prompt=512 · temp=0.0 · median of 5 runs
02 // Hardware specs
ArchitectureAda Lovelace
Process nodeTSMC 4N
Memory16 GB
Memory bandwidth736 GB/s
FP16 compute54 TFLOPS
INT8 compute108 TOPS
TDP320 W
PCIeGen 4 x16
Form factorTriple-slot
CoolingAxial
03 // Model fit
Approximate VRAM required to load weights + 4096 ctx KV cache.
+ STRENGTHS
- ✓16GB VRAM is enough for 32B-class models at Q4
- ✓736 GB/s memory bandwidth · top tier in its class
- ✓Strong tooling: FP16, FP8, Q8, Q4 all officially supported
− TRADE-OFFS
- −Draws 320W under load — plan PSU and thermals accordingly
- −Limited to triple-slot chassis
- −Driver lock-in to vendor stack