Llama 3 70B (Q4_K_M)
Meta Llama 3 70B Instruct, Q4_K_M GGUF, batch 1, 4K context, with offloading when VRAM-bound.
Primary metric: Tokens / sec (tok/s)
Reference prompts
Representative prompts for this workload. Exact prompts and harness settings still depend on the cited source for each record.
- Prompt 1
A train leaves station A at 9:00 going 60 mph. Another leaves station B (180 miles away) at 9:30 going 80 mph toward A. When do they meet? Show your work.
- Prompt 2
If a 70B-parameter model uses ~140GB at FP16, what is its likely memory footprint at Q4_K_M? Show the arithmetic.
- Prompt 3
Walk through which of these is cheaper for 10M tokens/day at 30 days: H100 cloud at $2/hr vs RTX 4090 owned at $1600 + $0.10/kWh.
- Prompt 4
Explain step-by-step why FP8 inference can be 2x faster than FP16 on H100 but not on A100.
- Prompt 5
Given a 4-bit quantized Llama 3 70B model at ~40GB, how much VRAM headroom is needed for KV cache at 8K context, batch 1?
Reference runtime command
A representative invocation for reproducing this workload class. Source-specific runs may use adjacent runtimes unless the record says otherwise.
llama-server -m llama-3-70b-instruct.Q4_K_M.gguf -c 4096 -ngl 999 --seed 42Single batch. Q4_K_M (~40GB). When VRAM-bound we note `ngl` value explicitly. Median of 5 runs.
Full leaderboard
Every record for Llama 3 70B Q4.
Top 10 leaderboard
Llama 3 70B (Q4_K_M) · sorted by tokens / sec
Value frontier, MSRP vs tokens / sec
Each dot is a device. Top-left is best value (cheap + fast).
Full leaderboard
Click any row for detailed breakdown. Click column headers to sort.
| # | Device | Verif | Buy | |||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | NVIDIA H200 141GBDatacenter GPU NVIDIA·Q4_K_M·4K ctx | 6 | 540tok/s | 141 GB | 700 W | $30k | 0.77 tok/s/W | $0.59/M | Amazon | |
| 2 | Cerebras WSE-3 (CS-3)ASIC Cerebras·FP16·8K ctx | 6 | 450tok/s | 44 GB | 23000 W | $2.50M | 0.02 tok/s/W | $0.0587/k | Amazon | |
| 3 | 8× NVIDIA H100 SXM5 80GBDatacenter GPU NVIDIA·Q4_K_M·8K ctx | 6 | 412tok/s | 640 GB | 5600 W | $200k | 0.07 tok/s/W | $0.0051/k | Amazon | |
| #4 | NVIDIA H100 SXM5 80GBDatacenter GPU NVIDIA·Q4_K_M·4K ctx | 6 | 410tok/s | 80 GB | 700 W | $25k | 0.59 tok/s/W | $0.64/M | Amazon | |
| #5 | AMD Instinct MI300X 192GBDatacenter GPU AMD·Q4_K_M·4K ctx | 6 | 380tok/s | 192 GB | 750 W | $15k | 0.51 tok/s/W | $0.42/M | Amazon | |
| #6 | Groq LPU (8-chip rack)ASIC Groq·FP16·8K ctx | 6 | 285tok/s | 1.84 GB | 1720 W | $160k | 0.17 tok/s/W | $0.0059/k | Amazon | |
| #7 | NVIDIA B200 192GBDatacenter GPU NVIDIA·Q4_K_M·8K ctx | 5 | 138tok/s σ 12.4 | 192 GB | 1000 W | $40k | 0.14 tok/s/W | $0.0031/k | Amazon | |
| #8 | NVIDIA GH200 Grace Hopper 480GBDatacenter GPU NVIDIA·Q4_K_M·8K ctx | 6 | 95tok/s | 96 GB | 1000 W | $44k | 0.10 tok/s/W | $0.0049/k | Amazon | |
| #9 | NVIDIA H200 141GBDatacenter GPU NVIDIA·Q4_K_M·8K ctx | 5 | 84tok/s σ 8.9 | 141 GB | 700 W | $30k | 0.12 tok/s/W | $0.0038/k | Amazon | |
| #10 | NVIDIA GeForce RTX 5090 32GBConsumer GPU NVIDIA·Q4_K_M·4K ctx | 6 | 78tok/s | 32 GB | 575 W | $2.0k | 0.14 tok/s/W | $0.27/M | Amazon | |
| #11 | AMD Instinct MI300X 192GBDatacenter GPU AMD·Q4_K_M·8K ctx | 6 | 72tok/s σ 5.8 | 192 GB | 750 W | $18k | 0.10 tok/s/W | $0.0026/k | Amazon | |
| #12 | NVIDIA H100 SXM5 80GBDatacenter GPU NVIDIA·Q4_K_M·4K ctx | 5 | 66tok/s σ 7.2 | 80 GB | 700 W | $25k | 0.09 tok/s/W | $0.0040/k | Amazon | |
| #13 | AWS Trainium2ASIC AWS·Q4_K_M·8K ctx | 6 | 58tok/s | 96 GB | 500 W | $28k | 0.12 tok/s/W | $0.0051/k | Amazon | |
| #14 | AMD Instinct MI250X 128GBDatacenter GPU AMD·Q4_K_M·8K ctx | 6 | 52tok/s | 128 GB | 560 W | $14k | 0.09 tok/s/W | $0.0028/k | Amazon | |
| #15 | NVIDIA A100 80GB SXMDatacenter GPU NVIDIA·Q4_K_M·4K ctx | 6 | 48tok/s | 80 GB | 400 W | $15k | 0.12 tok/s/W | $0.0033/k | Amazon | |
| #16 | NVIDIA A100 40GBDatacenter GPU NVIDIA·Q4_K_M·4K ctx | 6 | 42tok/s | 40 GB | 400 W | $9.0k | 0.10 tok/s/W | $0.0023/k | Amazon | |
| #17 | NVIDIA DGX Spark (Project DIGITS, 128GB)Datacenter GPU NVIDIA·Q4_K_M·8K ctx | 6 | 32tok/s | 128 GB | 240 W | $3.0k | 0.13 tok/s/W | $0.99/M | Amazon | |
| #18 | NVIDIA RTX 6000 Ada 48GBPro GPU NVIDIA·Q4_K_M·8K ctx | 6 | 30tok/s | 48 GB | 300 W | $6.8k | 0.10 tok/s/W | $0.0024/k | Amazon | |
| #19 | NVIDIA GeForce RTX 5090 32GBConsumer GPU NVIDIA·Q4_K_M·4K ctx | 6 | 28tok/s σ 2.1 | 32 GB | 575 W | $2.0k | 0.05 tok/s/W | $0.76/M | Amazon | |
| #20 | AMD Instinct MI210 64GBDatacenter GPU AMD·Q4_K_M·4K ctx | 6 | 28tok/s | 64 GB | 300 W | $8.0k | 0.09 tok/s/W | $0.0030/k | Amazon | |
| #21 | 2× NVIDIA RTX 4090 24GBConsumer GPU NVIDIA·Q4_K_M·8K ctx | 6 | 28tok/s | 48 GB | 900 W | $3.2k | 0.03 tok/s/W | $0.0012/k | Amazon | |
| #22 | AMD Radeon Pro W7900 48GBPro GPU AMD·Q4_K_M·4K ctx | 6 | 24tok/s | 48 GB | 295 W | $4.0k | 0.08 tok/s/W | $0.0018/k | Amazon | |
| #23 | 4× NVIDIA RTX 3090 24GBConsumer GPU NVIDIA·Q4_K_M·8K ctx | 6 | 24tok/s | 96 GB | 1400 W | $4.0k | 0.02 tok/s/W | $0.0018/k | Amazon | |
| #24 | NVIDIA L40S 48GBDatacenter GPU NVIDIA·Q4_K_M·4K ctx | 6 | 22tok/s | 48 GB | 350 W | $7.8k | 0.06 tok/s/W | $0.0037/k | Amazon | |
| #25 | NVIDIA A40 48GBPro GPU NVIDIA·Q4_K_M·4K ctx | 6 | 18tok/s | 48 GB | 300 W | $5.5k | 0.06 tok/s/W | $0.0032/k | Amazon | |
| #26 | NVIDIA GeForce RTX 4090 24GBConsumer GPU NVIDIA·Q4_K_M·4K ctx | 6 | 14tok/s | 24 GB | 450 W | $1.6k | 0.03 tok/s/W | $0.0012/k | Amazon | |
| #27 | Apple M3 Ultra (80c GPU, 512GB)Apple Silicon Apple·Q4_K_M·8K ctx | 6 | 14tok/s | 512 GB | 270 W | $9.5k | 0.05 tok/s/W | $0.0072/k | Amazon | |
| #28 | AMD Radeon RX 7900 XTX 24GBConsumer GPU AMD·Q4_K_M·4K ctx | 6 | 11tok/s | 24 GB | 355 W | $999 | 0.03 tok/s/W | $0.96/M | Amazon | |
| #29 | NVIDIA GeForce RTX 3090 24GBConsumer GPU NVIDIA·Q4_K_M·4K ctx | 6 | 9tok/s | 24 GB | 350 W | $1.5k | 0.03 tok/s/W | $0.0018/k | Amazon | |
| #30 | Apple M4 Max (40c GPU, 128GB)Apple Silicon Apple·Q4_K_M·8K ctx | 6 | 9tok/s | 128 GB | 65 W | $5.0k | 0.13 tok/s/W | $0.0062/k | Amazon | |
| #31 | NVIDIA GeForce RTX 5080 16GBConsumer GPU NVIDIA·Q4_K_M·4K ctx | 6 | 7tok/s | 16 GB | 360 W | $999 | 0.02 tok/s/W | $0.0015/k | Amazon | |
| #32 | Apple Mac mini M4 Pro 48GBApple Silicon Apple·Q4_K_M·4K ctx | 6 | 7tok/s | 48 GB | 35 W | $2.0k | 0.19 tok/s/W | $0.0033/k | Amazon | |
| #33 | Apple M4 Max (40c GPU, 128GB)Apple Silicon Apple·Q4_K_M·8K ctx | 6 | 6tok/s | 128 GB | 70 W | $4.7k | 0.09 tok/s/W | $0.0083/k | Amazon | |
| #34 | Apple M3 Max (40c GPU, 64GB)Apple Silicon Apple·Q4_K_M·4K ctx | 6 | 6tok/s | 64 GB | 65 W | $3.5k | 0.09 tok/s/W | $0.0067/k | Amazon | |
| #35 | AMD Ryzen AI Max+ 395 (Strix Halo, 96GB)NPU AMD·Q4_K_M·4K ctx | 6 | 5tok/s | 96 GB | 120 W | $2.2k | 0.04 tok/s/W | $0.0046/k | Amazon | |
| #36 | NVIDIA Jetson AGX Orin 64GBEdge NVIDIA·Q4_K_M·4K ctx | 6 | 4tok/s | 64 GB | 60 W | $2.0k | 0.07 tok/s/W | $0.0050/k | Amazon | |
| #37 | Intel Xeon 6980P (128c Granite Rapids)CPU-only Intel·Q4_K_M·4K ctx | 6 | 3tok/s | , | 500 W | $18k | 0.01 tok/s/W | $0.0672/k | Amazon | |
| #38 | AMD Threadripper PRO 7995WX (96-core)CPU-only AMD·Q4_K_M·4K ctx | 6 | 2tok/s | , | 350 W | $10.0k | 0.01 tok/s/W | $0.0440/k | Amazon |
Cite this benchmark
Use this in your paper, blog post, or comparison table.
@misc{myaihardware_llama3-70b-q4_2026,
title = {MyAI Bench: Llama 3 70B (Q4_K_M)},
author = {{MyAIHardware Contributors}},
year = {2026},
url = {https://www.myaihardware.com/benchmarks/workload/llama3-70b-q4},
note = {Version 1.3, accessed 2026-08-27}
}MyAIHardware Contributors. (2026). MyAI Bench: Llama 3 70B (Q4_K_M). MyAIHardware. Retrieved 2026-08-27, from https://www.myaihardware.com/benchmarks/workload/llama3-70b-q4
MyAIHardware Contributors. "MyAI Bench: Llama 3 70B (Q4_K_M)." MyAIHardware, 2026, https://www.myaihardware.com/benchmarks/workload/llama3-70b-q4. Accessed 2026-08-27.
MyAI Bench, Llama 3 70B (Q4_K_M). MyAIHardware Contributors, 2026. Version 1.3. https://www.myaihardware.com/benchmarks/workload/llama3-70b-q4 (accessed 2026-08-27).