How much VRAM do you need for a 70B model?
A 70B model at Q4_K_M is a ~42.5 GB file (bartowski GGUF on Hugging Face), so it needs roughly 44 GB+ of VRAM with headroom: two 24 GB cards or a 48 GB workstation card.
Llama-3.3-70B-Instruct needs about 44 GB of VRAM. The Q4_K_M GGUF is 42.52 GB (source), plus our 1.5 GB allowance for the runtime and a short-context KV cache.
No GPU or system with a dated price has that much memory. The 5 devices that fit it (NVIDIA RTX 6000 Ada Generation, NVIDIA RTX A6000, Apple Mac Studio M5 Max (40-core GPU, 64 GB), Apple Mac Studio M5 Ultra (64-core GPU, 96 GB), Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB) have no verified price in our data.
Cheapest fully priced machine that fits: the 48 GB: dual used RTX 3090 build at $5,443.94 for 48 GB, ~17 tok/s estimate. Cheapest multi-GPU route: 4× Intel Arc B580 = $1,320 for 48 GB, ~8 tok/s (cards only, no other parts).
Last updated Oct 16, 2026. Speeds are bandwidth-based estimates, not benchmarks. Prices are dated snapshots and move daily; "check price" means we have no verified price.
Every GPU and system that fits Llama-3.3-70B-Instruct
| GPU or system | VRAM | Bandwidth | Speed estimate | Low price | $ per tok/s |
|---|---|---|---|---|---|
| NVIDIA RTX 6000 Ada Generation | 48 GB | 960 GB/s | ~17 tok/s | check price | — |
| NVIDIA RTX A6000 | 48 GB | 768 GB/s | ~14 tok/s | check price | — |
| Apple Mac Studio M5 Max (40-core GPU, 64 GB) | 64 GB | 614 GB/s | ~11 tok/s | check price | — |
| Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) | 96 GB | 1200 GB/s | ~21 tok/s | check price | — |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | 192 GB | 273 GB/s | ~5 tok/s | check price | — |
Reaching 44 GB with identical cards
| Cards | Total VRAM | Cards only | Speed estimate | $ per tok/s (cards only) |
|---|---|---|---|---|
| 4× Intel Arc B580 | 48 GB | $1,320 | ~8 tok/s | $165 |
| 3× AMD Radeon RX 9070 | 48 GB | $1,710 | ~12 tok/s | $143 |
| 3× AMD Radeon RX 7800 XT | 48 GB | $1,797 | ~11 tok/s | $163 |
| 4× NVIDIA GeForce RTX 3060 12 GB | 48 GB | $1,820 | ~6 tok/s | $303 |
| 3× AMD Radeon RX 9060 XT 16 GB | 48 GB | $1,860 | ~6 tok/s | $310 |
| 3× NVIDIA GeForce RTX 5060 Ti 16 GB | 48 GB | $2,340 | ~8 tok/s | $293 |
| 3× AMD Radeon RX 9070 XT | 48 GB | $2,430 | ~12 tok/s | $203 |
| 2× AMD Radeon RX 7900 XTX | 48 GB | $2,700 | ~17 tok/s | $159 |
| 4× NVIDIA GeForce RTX 5070 | 48 GB | $3,200 | ~12 tok/s | $267 |
| 3× NVIDIA GeForce RTX 5070 Ti | 48 GB | $3,360 | ~16 tok/s | $210 |
| 2× NVIDIA GeForce RTX 3090 | 48 GB | $3,778 | ~17 tok/s | $222 |
| 4× NVIDIA GeForce RTX 4070 SUPER | 48 GB | $3,864 | ~9 tok/s | $429 |
| 3× NVIDIA GeForce RTX 5080 | 48 GB | $4,800 | ~17 tok/s | $282 |
| 2× NVIDIA GeForce RTX 4090 | 48 GB | $6,400 | ~18 tok/s | $356 |
| 2× NVIDIA GeForce RTX 5090 | 64 GB | $13,000 | ~32 tok/s | $406 |
Card prices only — no CPU, board, memory, PSU or case. Speed uses one card's bandwidth because a layer split reads one card at a time.
Fully priced builds that fit it
| Build | VRAM | Total | Fits Llama-3.3-70B-Instruct? | Speed estimate | $ per tok/s |
|---|---|---|---|---|---|
| 16 GB starter: Radeon RX 9070 | 16 GB | $1,744.44 | ❌ | — | — |
| 16 GB NVIDIA: RTX 5060 Ti 16 GB | 16 GB | $1,834.44 | ❌ | — | — |
| 24 GB: one used RTX 3090 | 24 GB | $2,933.45 | ❌ | — | — |
| 32 GB: two Radeon RX 9070 | 32 GB | $3,065.92 | ❌ | — | — |
| 48 GB: dual used RTX 3090 | 48 GB | $5,443.94 | ✅ | ~17 tok/s | $320 |
Speed calculator estimate
Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works
Questions people ask
How much VRAM do you need for Llama-3.3-70B-Instruct?
About 44 GB. The Q4_K_M GGUF is 42.52 GB and we add 1.5 GB for the runtime and a short-context KV cache. 5 of the 36 GPUs and systems we track have that much memory. Long contexts need more than our allowance.
What is the cheapest way to run Llama-3.3-70B-Instruct?
No GPU or system with a dated price has that much memory: the 5 devices that fit it (NVIDIA RTX 6000 Ada Generation, NVIDIA RTX A6000, Apple Mac Studio M5 Max (40-core GPU, 64 GB), Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) and Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB) have no verified price in our data. The cheapest fully priced machine that fits is our 48 GB: dual used RTX 3090 build at $5,443.94 for 48 GB, an estimated 17 tok/s. The cheapest way to reach the memory is to split the model across identical cards: 4 x Intel Arc B580 is $1,320 for 48 GB, at an estimated 8 tok/s — cards only, before the CPU, board, memory, PSU and case.
Can you run Llama-3.3-70B-Instruct on multiple GPUs?
Yes: VRAM adds up across cards, so 4 x Intel Arc B580 (48 GB, $1,320) covers the 44 GB this model needs. A layer split still reads from one card at a time, so the estimate stays at one card's bandwidth: about 8 tok/s.
How fast does Llama-3.3-70B-Instruct run?
Estimates across the 5 devices that fit it range from about 5 tok/s on the Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB (273 GB/s) to about 21 tok/s on the Apple Mac Studio M5 Ultra (64-core GPU, 96 GB) (1200 GB/s). These are bandwidth estimates, not benchmarks.
Related
- What the Intel Arc B580 can run
- What the AMD Radeon RX 9070 can run
- What the NVIDIA RTX 6000 Ada Generation can run
- What the NVIDIA RTX A6000 can run
- What GPU you need for Llama-3.3-70B-Instruct
- What GPU you need for Qwen3-32B
- Can the Arc B580 run Qwen3-14B?
- Can the Arc B580 run Mistral-Nemo-Instruct-2407?
- Can the Radeon RX 9070 run Mistral-Small-3.2-24B-Instruct-2506?
- Can the Radeon RX 9070 run gpt-oss-20b?
- Build: 48 GB: dual used RTX 3090 — $5,443.94, 48 GB
- Guide: Cheapest way to run gpt-oss-120b locally
- Guide: Best GPU for local LLMs (2026)
- Guide: Best PC for local LLMs under $2,000