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
01b // Performance across quantization
vs. nearest competitors
How tok/s scales from FP16 → Q8 → Q4 compared to GPUs in a similar price/VRAM range.
02 // Hardware specs
ArchitectureAda Lovelace
Process nodeTSMC 4N
Memory24 GB
Memory bandwidth1,008 GB/s
FP16 compute82 TFLOPS
INT8 compute165 TOPS
TDP450 W
PCIeGen 4 x16
Form factorTriple-slot
CoolingAxial
03 // Model fit
Approximate VRAM required to load weights + 4096 ctx KV cache.
+ STRENGTHS
- ✓24GB VRAM is enough for 32B-class models at Q4
- ✓1008 GB/s memory bandwidth · top tier in its class
- ✓Strong tooling: FP16, FP8, Q8, Q4 all officially supported
− TRADE-OFFS
- −Draws 450W under load — plan PSU and thermals accordingly
- −Limited to triple-slot chassis
- −Driver lock-in to vendor stack
related research
Research behind GeForce RTX 4090 inference tradeoffs
These papers explain the quantization, cache, bandwidth, and runtime constraints that matter before buying this GPU for local AI.
Local AI inference papers
llama.cpp, Apple Silicon, constrained GPUs, offload, and one-box inference.
Open LLM quantization research
GPTQ, AWQ, GGUF, FP4, NF4, and what low-bit formats mean for VRAM fit.
Open KV cache optimization papers
Cache quantization, compression, reuse, and long-context memory pressure.
Open