AM AMD Enterprise Discrete GPU

AMD Instinct MI325X

Runs downloadable LLMs up to ~410B at Q4, ~843.8-1031.3 tok/s on an 8B model.

AM
256GB · 6000 GB/s
256GB
VRAM
6000
GB/s bandwidth
410B
Max model (Q4)
1000W
TDP · PSU 1400W

What you can run

At 8k context, FP16 KV. Full-speed tok/s holds up to "max ctx"; beyond it the context spills to system RAM (CPU/PCIe offload) and drops to the degraded figure.

Model sizeFP16Q8Q4
3B 630-770 t/s
≤3479k ctx · then ~4.2-5.1 (slow)
1260-1540 t/s
≤3525k ctx · then ~8.4-10.3 (slow)
2250-2750 t/s
≤3545k ctx · then ~15-18.3 (slow)
7-8B 236.3-288.8 t/s
≤1663k ctx · then ~1.6-1.9 (slow)
472.5-577.5 t/s
≤1724k ctx · then ~3.2-3.9 (slow)
843.8-1031.3 t/s
≤1751k ctx · then ~5.6-6.9 (slow)
13B 145.4-177.7 t/s
≤1270k ctx · then ~1-1.2 (slow)
290.8-355.4 t/s
≤1349k ctx · then ~1.9-2.4 (slow)
519.2-634.6 t/s
≤1384k ctx · then ~3.5-4.2 (slow)
32B 59.1-72.2 t/s
≤649k ctx · then ~0.4-0.5 (slow)
118.1-144.4 t/s
≤771k ctx · then ~0.8-1 (slow)
210.9-257.8 t/s
≤824k ctx · then ~1.4-1.7 (slow)
70B 27-33 t/s
≤287k ctx · then ~0.2-0.2 (slow)
54-66 t/s
≤501k ctx · then ~0.4-0.4 (slow)
96.4-117.9 t/s
≤595k ctx · then ~0.6-0.8 (slow)
120B won't run 31.5-38.5 t/s
≤348k ctx · then ~0.2-0.3 (slow)
56.3-68.8 t/s
≤509k ctx · then ~0.4-0.5 (slow)
400B won't run won't run 16.7-20.4 t/s
≤14k ctx · then ~0.1-0.1 (slow)

Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (AMD). Cyan = fits ~235.5 GB usable. Beyond "ctx", KV spills to system RAM (CPU/PCIe offload) (rough degraded speed shown).

What else you need for a full build

Minimum supporting components for a desktop built around the AMD Instinct MI325X. Guidance — exact needs depend on your case, other parts and how much you offload to system RAM.

Power supply
≥ 1400W · 80+ Gold
Headroom over the 1000W card plus the rest of the build.
System RAM
≥ 256 GB · 512 GB recommended
At least the 256 GB of VRAM; 2× lets you CPU-offload models larger than VRAM.
Motherboard
1× free PCIe x16 slot (4.0/5.0)
A full-length x16 electrical slot; PCIe 4.0 is ample, 5.0 future-proofs.
CPU
Any modern multi-core
The GPU runs inference; more cores and memory channels only matter when offloading to RAM.
Storage
≥ 1 TB NVMe SSD
Weights are large — a 70B model at Q4 is ~40 GB; budget for a few.

Run popular models on the AMD Instinct MI325X

Videos about the AMD Instinct MI325X

Is the Nvidia Monopoly Over? Fine-tuning our Liquid AI LLMs on AMD MI325X

Massive AMD Instinct MI325X 2TB HBM3e Server by Gigabyte

More hardware like the AMD Instinct MI325X

About the AMD Instinct MI325X

AMD Instinct MI325X is a discrete GPU from AMD, 2024. For local genAI the numbers that matter are its 256 GB (how big a model + context fits) and 6000 GB/s bandwidth (generation speed). Plan for a ~1400W power supply.

Frequently asked questions

What size LLM can the AMD Instinct MI325X run?

With ~235.5 GB usable it runs up to ~410B at Q4 (~115.5B at FP16) at 8k context. Bigger spills into system RAM (CPU/PCIe offload) and slows sharply.

How many tokens/sec does the AMD Instinct MI325X generate?

Generation is bandwidth-bound. At 6000 GB/s an 8B model at Q4 runs ~843.8-1031.3 tok/s (estimate).

Why does long context slow it down?

The KV cache for the context must also fit fast memory; once weights+KV exceed 235.5 GB it spills to system RAM (CPU/PCIe offload) and throughput collapses. Each row shows the max usable context.

What else do I need to run the AMD Instinct MI325X?

Besides the card: a ≥1400W 80+ Gold power supply, at least 256GB of system RAM (2× the 256GB VRAM is ideal for CPU offload), a free PCIe x16 slot, any modern multi-core CPU, and a ≥1TB NVMe SSD for model weights.