AM AMD Enterprise Discrete GPU

AMD Instinct MI300X 192GB

Runs downloadable LLMs up to ~305B at Q4, ~745.3-910.9 tok/s on an 8B model.

AM
192GB · 5300 GB/s
192GB
VRAM
5300
GB/s bandwidth
305B
Max model (Q4)
$15,000
MSRP

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 556.5-680.2 t/s
≤2581k ctx · then ~4.2-5.1 (slow)
1113-1360.3 t/s
≤2627k ctx · then ~8.4-10.3 (slow)
1987.5-2429.2 t/s
≤2647k ctx · then ~15-18.3 (slow)
7-8B 208.7-255.1 t/s
≤1214k ctx · then ~1.6-1.9 (slow)
417.4-510.1 t/s
≤1275k ctx · then ~3.2-3.9 (slow)
745.3-910.9 t/s
≤1302k ctx · then ~5.6-6.9 (slow)
13B 128.4-157 t/s
≤910k ctx · then ~1-1.2 (slow)
256.8-313.9 t/s
≤990k ctx · then ~1.9-2.4 (slow)
458.7-560.6 t/s
≤1025k ctx · then ~3.5-4.2 (slow)
32B 52.2-63.8 t/s
≤424k ctx · then ~0.4-0.5 (slow)
104.3-127.5 t/s
≤546k ctx · then ~0.8-1 (slow)
186.3-227.7 t/s
≤600k ctx · then ~1.4-1.7 (slow)
70B 23.9-29.2 t/s
≤107k ctx · then ~0.2-0.2 (slow)
47.7-58.3 t/s
≤321k ctx · then ~0.4-0.4 (slow)
85.2-104.1 t/s
≤415k ctx · then ~0.6-0.8 (slow)
120B won't run 27.8-34 t/s
≤168k ctx · then ~0.2-0.3 (slow)
49.7-60.7 t/s
≤329k ctx · then ~0.4-0.5 (slow)
400B won't run won't run won't run

Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (AMD). Cyan = fits ~176.6 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 MI300X 192GB. Guidance — exact needs depend on your case, other parts and how much you offload to system RAM.

Power supply
≥ 1150W · 80+ Gold
Headroom over the 750W card plus the rest of the build.
System RAM
≥ 256 GB · 512 GB recommended
At least the 192 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 MI300X 192GB

Videos about the AMD Instinct MI300X 192GB

The AMD Instinct MI300X GPU can handle the Meta LLaMa 60B parameter model on a single GPU. #amd

AMD MI300X: LLM Setup Was SHOCKINGLY FAST!

More hardware like the AMD Instinct MI300X 192GB

About the AMD Instinct MI300X 192GB

AMD Instinct MI300X 192GB is a discrete GPU from AMD on CDNA 3, 2023. For local genAI the numbers that matter are its 192 GB (how big a model + context fits) and 5300 GB/s bandwidth (generation speed). Plan for a ~1150W power supply.

Frequently asked questions

What size LLM can the AMD Instinct MI300X 192GB run?

With ~176.6 GB usable it runs up to ~305B at Q4 (~86B at FP16) at 8k context. Bigger spills into system RAM (CPU/PCIe offload) and slows sharply.

How many tokens/sec does the AMD Instinct MI300X 192GB generate?

Generation is bandwidth-bound. At 5300 GB/s an 8B model at Q4 runs ~745.3-910.9 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 176.6 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 MI300X 192GB?

Besides the card: a ≥1150W 80+ Gold power supply, at least 192GB of system RAM (2× the 192GB VRAM is ideal for CPU offload), a free PCIe x16 slot, any modern multi-core CPU, and a ≥1TB NVMe SSD for model weights.