Hardware to run Llama 3.1 405B

Dense 405 B — heavy bandwidth need. Now eclipsed by DeepSeek V4-Pro, Kimi K2.6 and GLM-5.1 on every coding/agent benchmark; kept for chat-only setups with 256+ GB. Q2 fits in 2× DGX Spark cluster or M3 Ultra 256.

Quantization Q2 Q2_K Q4 Q4_K_M Q5 Q5_K_M Q8 Q8_0

Availability ⚡ Only plug & play

Cheapest

4× Strix Halo cluster (512 GB unified)

AMD · rack of 4 mini-PCs, 10 GbE fabric

$11,500

tokens / secQ4

120B-MoE 55 t/s

235B-MoE 24 t/s

671B-MoE 5.0 t/s

Memory512 GB · 384 usable

Bandwidth256 GB/s

Idle / Active32 W / 480 W

Sticker$11,500

Why: Lowest sticker that still fits Llama 3.1 405B ($12k USD).

Amazon (Strix Halo Mini PC)GMKtec EVO-X2Bosgame M5Framework

📺 Reviews on YouTube

▶ Kimi-K2 (1T) / GLM 4.7 (355B) on a 4-node Strix Halo cluster — 512 GB unified memory ▶ AMD Strix Halo / Ryzen AI Max+ 395 — honest review (single-node baseline)

Fastest

12× RTX Pro 6000 Blackwell rack (1152 GB)

NVIDIA · 8U server rack (multi-node, 1-2 chassis)

$118,000

tokens / secQ4

120B-MoE —

235B-MoE 260 t/s

671B-MoE 95 t/s

Memory1152 GB · 1116 usable

Bandwidth1792 GB/s

Idle / Active340 W / 7400 W

Sticker$118,000

Why: Highest measured tg/s — 78 t/s on Llama 3.1 405B-class models at Q4.

AmazonCoreWeaveSuperMicroSHI enterprise

All-rounder

Single AMD Instinct MI355X 288 GB workstation

AMD · 4U server (OAM, liquid-cooled)

$28,000

tokens / secQ4

120B-MoE 320 t/s

235B-MoE 130 t/s

671B-MoE 50 t/s

Memory288 GB · 282 usable

Bandwidth8000 GB/s

Idle / Active140 W / 1400 W

Sticker$28,000

Why: Top quartile across speed, value, memory headroom, and efficiency — the "buy this if unsure" pick.

Dell XE9712AMD MI350 seriesSupermicro UBB

📺 Reviews on YouTube

▶ AMD's MI350/355X Advancing AI Event Recap ▶ AMD AI Event w/ Craft Computing — MI350/355X Launch, ROCm 7 ▶ AMD Visit and Tour! Featuring ROCm 7 and AMD Instinct

Best value

Quad RTX Pro 6000 Blackwell build (384 GB)

NVIDIA · workstation / 4U pedestal

$38,000

tokens / secQ4

120B-MoE —

235B-MoE 145 t/s

671B-MoE 40 t/s

Memory384 GB · 372 usable

Bandwidth1792 GB/s

Idle / Active100 W / 2200 W

Sticker$38,000

Why: Best $/tg-per-second — ~$874 per t/s.

AmazonNeweggB&H PhotoPNY Pro

Best CUDA

8× RTX Pro 6000 Blackwell server (768 GB)

NVIDIA · 4U server (e.g. SuperMicro AS-4125GS-TNRT)

$78,000

tokens / secQ4

120B-MoE —

235B-MoE 220 t/s

671B-MoE 75 t/s

Memory768 GB · 744 usable

Bandwidth1792 GB/s

Idle / Active220 W / 4800 W

Sticker$78,000

Why: Strongest CUDA-only software stack among fitting builds.

AmazonPNY ProCoreWeaveSuperMicro

Most VRAM

DGX B200 — 8× B200 server (1.44 TB HBM3e)

NVIDIA · 10U DGX server

$475,000

tokens / secQ4

120B-MoE 330 t/s

235B-MoE 165 t/s

671B-MoE 105 t/s

Memory1440 GB · 1404 usable

Bandwidth8000 GB/s

Idle / Active900 W / 10200 W

Sticker$475,000

Why: 1404 GB usable — most headroom for batching and longer contexts.

NVIDIA DGX B200SuperMicro HGX B200Lambda Labs

📺 Reviews on YouTube

▶ Inside a 1.44 TB HBM3e NVIDIA HGX B200 AI Server from ASRock Rack

Efficient

Mac Studio M3 Ultra 512 GB

Apple · small desktop

$14,199

tokens / secQ4

120B-MoE 55 t/s

235B-MoE 22 t/s

671B-MoE 8.0 t/s

Memory512 GB · 480 usable

Bandwidth819 GB/s

Idle / Active12 W / 220 W

Sticker$14,199

Why: 220 W active — lowest power draw of the fitting builds.

Apple RefurbB&H Photo (used)eBay (used)

📺 Reviews on YouTube

▶ M3 Ultra Mac Studio Review ▶ M3 Ultra vs RTX 5090 — The Final Battle

Cheapest

Mac Studio M3 Ultra 512 GB

Apple · small desktop

$14,199

tokens / secQ4

120B-MoE 55 t/s

235B-MoE 22 t/s

671B-MoE 8.0 t/s

Memory512 GB · 480 usable

Bandwidth819 GB/s

Idle / Active12 W / 220 W

Sticker$14,199

Why: Lowest sticker that still fits Llama 3.1 405B ($14k USD).

Apple RefurbB&H Photo (used)eBay (used)

📺 Reviews on YouTube

▶ M3 Ultra Mac Studio Review ▶ M3 Ultra vs RTX 5090 — The Final Battle

Fastest

Single AMD Instinct MI355X 288 GB workstation

AMD · 4U server (OAM, liquid-cooled)

$28,000

tokens / secQ4

120B-MoE 320 t/s

235B-MoE 130 t/s

671B-MoE 50 t/s

Memory288 GB · 282 usable

Bandwidth8000 GB/s

Idle / Active140 W / 1400 W

Sticker$28,000

Why: Highest measured tg/s — 39 t/s on Llama 3.1 405B-class models at Q4.

Dell XE9712AMD MI350 seriesSupermicro UBB

📺 Reviews on YouTube

▶ AMD's MI350/355X Advancing AI Event Recap ▶ AMD AI Event w/ Craft Computing — MI350/355X Launch, ROCm 7 ▶ AMD Visit and Tour! Featuring ROCm 7 and AMD Instinct

Every other build that runs Llama 3.1 405B

6 additional builds fit Llama 3.1 405B at Q4_K_M (256 GB usable minimum), sorted by sticker price.

Build	Price	Memory	Bandwidth	tg/s (Q4)	Active W	5-yr power
4× DGX Spark cluster (512 GB unified, CUDA)NVIDIA · rack of 4 desktops	$20k	512 / 488 GB	273 GB/s	9.6 t/s	920 W	$3.4k
8× Strix Halo cluster (1024 GB unified)AMD · rack of 8 mini-PCs, 10/25 GbE fabric	$23k	1024 / 768 GB	256 GB/s	9.6 t/s	960 W	$3.4k
2× Mac Studio M3 Ultra 512 GB cluster (TB5 / MLX)Apple · two desktops, Thunderbolt 5 RDMA	$28k	1024 / 960 GB	819 GB/s	8.7 t/s	440 W	$1.5k
8× DGX Spark cluster (1024 GB unified, CUDA)NVIDIA · rack of 8 desktops, 200 GbE fabric	$44k	1024 / 976 GB	273 GB/s	13 t/s	1840 W	$7k
8× H100 80 GB serverNVIDIA · server rack	$280k	640 / 620 GB	3350 GB/s	33 t/s	5600 W	$20k
DGX H200 — 8× H200 server (1.13 TB HBM3e)NVIDIA · 8U DGX / HGX server rack	$380k	1128 / 1100 GB	4800 GB/s	47 t/s	6500 W	$24k

Last updated 2026-07-11