NVIDIA
NVIDIA B200 192GB
Curated Aggregate·2026-02-22192GB / $40k
Measured
360 tok/s
Power
1000 W
Llama 3 8B Q4 benchmark available. Open the detail page for source links, confidence, and buying advice.
Benchmark-backed recommender
Pick the model.See the GPUs that proved it.
This is not a static VRAM chart. Choose a workload and we rank GPUs by real benchmark evidence, trust status, and hardware fit.
Tested GPUs
54
Consumer, pro, and datacenter paths
Model presets
8
Mapped to actual workloads in the benchmark DB
Freshest evidence
2026-05-25
Latest tested date among ranked devices
Choose a target model
Selected workload
Llama 3 8B Q4
The mainstream local-LLM baseline. Great for chat, RAG, and lightweight agents.
Benchmark workload
Llama 3 8B Q4
VRAM guidance
8GB works, 12GB+ feels comfortable.
Category
general
Ranked GPUs
12 tested
Recommendation logic
We only rank GPUs that have an actual benchmark record for this workload. Ordering favors measured throughput first, then the device's broader composite quality score.
Search your GPU and we'll show whether this specific model target is already benchmarked on it.
Ranked hardware
Best GPUs for Llama 3 8B Q4
great
NVIDIA
NVIDIA H200 141GB
Curated Aggregate·2026-04-20141GB / $30k
Measured
215 tok/s
Power
700 W
405B-class models at FP8 with context headroom — the capacity play for frontier open-weight models
NVIDIA
NVIDIA GeForce RTX 5090 32GB
Curated Aggregate·2026-05-2432GB / $2.0k
Measured
182 tok/s
Power
575 W
70B Q4 full-GPU at 40-55 tok/s with 8-16K context — the first consumer card where 70B daily-driving is practical
NVIDIA
NVIDIA H100 SXM5 80GB
Curated Aggregate·2026-04-0980GB / $25k
Measured
178 tok/s
Power
700 W
70B FP16 with substantial context, 405B Q4 with headroom. Production multi-tenant serving at batch 8-64 via TensorRT-LLM or vLLM.
NVIDIA
NVIDIA GeForce RTX 5080 16GB
Curated Aggregate·2026-05-2316GB / $999
Measured
165 tok/s
Power
360 W
13B-class Q4 full-GPU at 70+ tok/s — excellent for coding agents on Qwen 2.5 Coder 14B or similar
NVIDIA
NVIDIA GeForce RTX 4090 24GB
Curated Aggregate·2026-05-2524GB / $1.6k
Measured
132 tok/s
Power
450 W
32B-class Q4 full-GPU at 70+ tok/s — Qwen 3 32B, Qwen 2.5 Coder 32B, QwQ 32B all fit comfortably
NVIDIA
NVIDIA GeForce RTX 5070 Ti 16GB
Curated Aggregate·2026-03-0416GB / $749
Measured
128 tok/s
Power
300 W
Llama 3 8B Q4 benchmark available. Open the detail page for source links, confidence, and buying advice.
AMD
AMD Instinct MI300X 192GB
Curated Aggregate·2026-03-19192GB / $15k
Measured
122 tok/s
Power
750 W
Large-model inference where VRAM capacity matters more than peak compute — 405B Q4 with headroom, 70B FP16 comfortably
NVIDIA
NVIDIA GeForce RTX 4080 Super 16GB
Curated Aggregate·2025-08-0916GB / $999
Measured
102 tok/s
Power
320 W
Llama 3 8B Q4 benchmark available. Open the detail page for source links, confidence, and buying advice.
Apple
Apple M3 Ultra (80c GPU, 512GB)
Curated Aggregate·2026-01-26512GB / $7.0k
Measured
102 tok/s
Power
270 W
Llama 3 8B Q4 benchmark available. Open the detail page for source links, confidence, and buying advice.
AMD
AMD Radeon RX 7900 XTX 24GB
Curated Aggregate·2026-03-1148GB / $999
Measured
95.0 tok/s
Power
355 W
24GB VRAM for Linux-native builders who want NVIDIA-tier capacity at AMD pricing and are comfortable with software tradeoffs
NVIDIA
NVIDIA GeForce RTX 5070 12GB
Curated Aggregate·2026-03-2012GB / $549
Measured
92.0 tok/s
Power
250 W
Llama 3 8B Q4 benchmark available. Open the detail page for source links, confidence, and buying advice.