01 // Representative inference estimates
Single-stream decode · llama.cpp
Aggregated from published community sources. Test setups differ, so use close results as directional evidence.
Sources and normalization →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
ArchitectureAmpere
Process nodeSamsung 8N
Memory24 GB
Memory bandwidth936 GB/s
FP16 compute35.6 TFLOPS
INT8 compute71 TOPS
TDP350 W
PCIeGen 4 x16
Form factorTriple-slot
CoolingAxial
03 // Model fit
Weight estimate plus 15% planning headroom for runtime buffers and a modest KV cache. Long context can require more.
+ STRENGTHS
- ✓24GB clears our Qwen3 32B Q4 planning target
- ✓936 GB/s memory bandwidth · top tier in its class
- ✓Indexed formats: FP16, Q8, Q4 · verify support in your runtime
− TRADE-OFFS
- −Draws 350W under load — plan PSU and thermals accordingly
- −Limited to triple-slot chassis
- −Driver lock-in to vendor stack
related research
Research behind GeForce RTX 3090 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