Llama · Jul 2024 · 8B
Llama 3.1 8B: what runs it and what it costs today
At Q8_0 with 8K of context it needs about 11 GB of memory. The cheapest machine that runs it today is the NVIDIA Tesla P100 16GB at $127.63.
By quantization
Memory needed includes the KV cache for 8K of context and runtime overhead. The cheapest machine is the lowest live price, item plus shipping, that fits with room for that context.
| Quant | Needs | Cheapest machine that fits today | Price |
|---|---|---|---|
| FP16 | 19 GB | NVIDIA Tesla P40 | $273 |
| Q8_0 | 11 GB | NVIDIA Tesla P100 16GB | $127.63 |
| Q6_K | 9 GB | NVIDIA Tesla P100 16GB | $127.63 |
| Q5_K_M | 8 GB | NVIDIA Tesla P100 16GB | $127.63 |
| Q4_K_M | 7 GB | NVIDIA Tesla P100 16GB | $127.63 |
| Q3_K_M | 6 GB | NVIDIA Tesla P100 16GB | $127.63 |
| IQ2_M | 5 GB | NVIDIA Tesla P100 16GB | $127.63 |
Every machine at Q8_0 34 fit
| Machine | Memory | Fit at 8K | Est. speed | Price | Per GB |
|---|---|---|---|---|---|
| NVIDIA Tesla P100 16GB card | 16 GB | fits 44K ctx | ~56 tok/s | $127.63 eBay | $7/GB |
| AMD Radeon Instinct MI50 16GB card | 16 GB | fits 44K ctx | ~60 tok/s | $228.78 eBay | $14/GB |
| NVIDIA Tesla P40 card | 24 GB | fits 104K ctx | ~27 tok/s | $273 eBay | $11/GB |
| NVIDIA GeForce RTX 3060 12GB card | 12 GB | fits 15K ctx | ~28 tok/s | $320 eBay | $26/GB |
| AMD Radeon Instinct MI50 32GB card | 32 GB | fits 128K+ ctx | ~60 tok/s | $600 eBay | $18/GB |
| NVIDIA GeForce RTX 5060 Ti 16GB card | 16 GB | fits 44K ctx | ~34 tok/s | $617.99 eBay | $38/GB |
| NVIDIA GeForce RTX 4060 Ti 16GB card | 16 GB | fits 44K ctx | ~22 tok/s | $656.27 eBay | $41/GB |
| AMD Radeon RX 7900 XTX card | 24 GB | fits 104K ctx | ~57 tok/s | $900 eBay | $37/GB |
| NVIDIA RTX A4000 card | 16 GB | fits 44K ctx | ~34 tok/s | $950 eBay | $59/GB |
| NVIDIA GeForce RTX 3090 card | 24 GB | fits 104K ctx | ~72 tok/s | $1,420 eBay | $59/GB |
| NVIDIA RTX 4000 SFF Ada card | 20 GB | fits 74K ctx | ~21 tok/s | $1,510.30 eBay | $75/GB |
| AMD Radeon AI PRO R9700 card | 32 GB | fits 128K+ ctx | ~38 tok/s | $1,820 eBay | $56/GB |
| NVIDIA GeForce RTX 4090 card | 24 GB | fits 104K ctx | ~77 tok/s | $2,083.48 eBay | $86/GB |
| NVIDIA RTX A5000 card | 24 GB | fits 104K ctx | ~59 tok/s | $2,325.65 eBay | $96/GB |
| NVIDIA RTX PRO 4000 Blackwell card | 24 GB | fits 104K ctx | ~52 tok/s | $2,827.16 eBay | $117/GB |
| NVIDIA RTX PRO 4000 Blackwell SFF card | 24 GB | fits 104K ctx | ~33 tok/s | $2,899 eBay | $120/GB |
| Framework Desktop, Ryzen AI Max+ 395, 128 GB desktop box | 128 GB | fits 128K+ ctx | ~15 tok/s | $3,464.65 eBay | $27/GB |
| GMKtec EVO-X2, Ryzen AI Max+ 395, 128 GB desktop box | 128 GB | fits 128K+ ctx | ~15 tok/s | $3,649.99 eBay | $28/GB |
| NVIDIA RTX PRO 4500 Blackwell card | 32 GB | fits 128K+ ctx | ~69 tok/s | $4,450 eBay | $139/GB |
| NVIDIA RTX A6000 card | 48 GB | fits 128K+ ctx | ~59 tok/s | $4,500 eBay | $93/GB |
| Apple Mac Studio M4 Max, 128 GB desktop box | 128 GB | fits 128K+ ctx | ~39 tok/s | $4,899 eBay | $38/GB |
| NVIDIA GeForce RTX 5090 card | 32 GB | fits 128K+ ctx | ~140 tok/s | $4,920.99 eBay | $153/GB |
| ASUS Ascent GX10 desktop box | 128 GB | fits 128K+ ctx | ~21 tok/s | $6,099.99 eBay | $47/GB |
| Acer Veriton GN100 desktop box | 128 GB | fits 128K+ ctx | ~21 tok/s | $6,367.19 eBay | $49/GB |
| HP ZGX Nano AI Station desktop box | 128 GB | fits 128K+ ctx | ~21 tok/s | $6,440.40 eBay | $50/GB |
| Gigabyte AI TOP ATOM desktop box | 128 GB | fits 128K+ ctx | ~21 tok/s | $6,999 eBay | $54/GB |
| NVIDIA RTX PRO 5000 Blackwell card | 48 GB | fits 128K+ ctx | ~100 tok/s | $7,099.99 eBay | $147/GB |
| NVIDIA RTX 6000 Ada card | 48 GB | fits 128K+ ctx | ~74 tok/s | $7,698.43 eBay | $160/GB |
| NVIDIA DGX Spark desktop box | 128 GB | fits 128K+ ctx | ~21 tok/s | $7,799.98 eBay | $60/GB |
| Apple Mac Studio M3 Ultra, 256 GB desktop box | 256 GB | fits 128K+ ctx | ~58 tok/s | $8,499 eBay | $33/GB |
| NVIDIA A100 80GB PCIe card | 80 GB | fits 128K+ ctx | ~150 tok/s | $15,299 eBay | $191/GB |
| NVIDIA RTX PRO 6000 Blackwell card | 96 GB | fits 128K+ ctx | ~140 tok/s | $16,499 eBay | $171/GB |
| Apple Mac Studio M3 Ultra, 512 GB desktop box | 512 GB | fits 128K+ ctx | ~58 tok/s | $17,499 eBay | $34/GB |
| NVIDIA H100 80GB PCIe card | 80 GB | fits 128K+ ctx | ~150 tok/s | $26,995.21 eBay | $337/GB |
Estimates assume the whole model in memory with llama.cpp-style inference; see how we estimate. Prices are the cheapest live eBay listing that passed our filters, item plus shipping.