DEPLOY

AI chip · AMD

AMD Instinct MI300X

CDNA 3 data-center accelerator with 192 GB HBM3, positioned as the primary alternative to Nvidia H100/H200 for LLM inference workloads. Fabricated by TSMC.

Part of 1 rack design.

Market position

MI300X is the chip that gave AMD real inference share against H100/H200: 192 GB HBM3 (2.4x H100 80GB) at $10-15K street. Meta, Microsoft (Copilot inference), Oracle and Vultr became public buyers. First real dual-source cloud GPU since Volta.

Efficiency and power

How much work you get per watt and per dollar, and how much power a full rack draws.

Perf per watt
3.49 FP8 TFLOPS/W
2,614 TFLOPS ÷ 750 W = 3.49
List $ per FP8 TFLOP
$5.74
$15,000 ÷ 2,614 TFLOPS = $5.74
Rack power
6 kW per AMD Instinct MI300X Platform (8-GPU OAM baseboard) (~750 W per chip)
6 kW ÷ 8 chips (whole-system, includes CPU, memory, NICs, cooling)
Cooling
air
Depends on the rack design

What fits in 192 GB

Which open-source LLMs run on one AMD Instinct MI300X, by precision. Weights only: add roughly 15% headroom for real serving. If a model does not fit at FP16, try INT8 or INT4 (smaller quality trade-off than most people expect).

ModelParamsFP16INT8INT4
Llama 3.1 8B
dense
8 B16 GB ✓8 GB ✓4 GB ✓
Llama 3.1 70B
dense
70 B140 GB ✓70 GB ✓35 GB ✓
Llama 3.1 405B
dense
405 B810 GB ✗405 GB ✗203 GB ✗
Llama 3.3 70B
dense
70 B140 GB ✓70 GB ✓35 GB ✓
DeepSeek V3
MoE (671B total, 37B active per token)
671 B1342 GB ✗671 GB ✗336 GB ✗
DeepSeek R1
MoE (671B total, 37B active per token)
671 B1342 GB ✗671 GB ✗336 GB ✗
Qwen 2.5 7B
dense
7 B14 GB ✓7 GB ✓4 GB ✓
Qwen 2.5 72B
dense
72 B144 GB ✓72 GB ✓36 GB ✓
Mixtral 8x7B
MoE (46.7B total, 12.9B active per token)
46.7 B93 GB ✓47 GB ✓23 GB ✓
Mixtral 8x22B
MoE (141B total, 39B active per token)
141 B282 GB ✗141 GB ✓71 GB ✓
Gemma 2 27B
dense
27 B54 GB ✓27 GB ✓14 GB ✓
Command R+
dense
104 B208 GB ✗104 GB ✓52 GB ✓
Kimi K2
MoE (1T total, 32B active per token)
1000 B2000 GB ✗1000 GB ✗500 GB ✗

Math: FP16 = params × 2 bytes; INT8 = params × 1 byte; INT4 = params × 0.5 bytes. MoE models sum every expert (full weights on disk), not the per-token active subset.

Common questions

How much does AMD Instinct MI300X cost?

Launch list price was $15,000 when it shipped in 2024-01-01. Street prices swing with supply and how old the generation is. Cloud rental rates vary widely by provider. See list price per TFLOP for the full ranking.

How much memory does AMD Instinct MI300X have?

192 GB of HBM3, running at 5,300 GB/s. Straight from the vendor datasheet. See chips with the most memory for context.

How much power does one AMD Instinct MI300X draw?

750 W at the chip. A full server draws more once you add CPU, memory, networking and cooling: see the rack power number above. Compare to other chips on perf-per-watt.

Which open-source LLMs fit on one AMD Instinct MI300X?

Of 13 open-source LLMs we track, 7 fit at FP16, 9 at INT8, and 9 at INT4 (for example Llama 3.1 8B, Llama 3.1 70B and Llama 3.3 70B). Full table above with each model's memory need.

See every answer we publish →

Key facts

Class
Compute SoC (AI accelerator)
Designer
AMD
Safety-critical?
No (data-center inference / training)
Record as of
2026-09-27
Most recent source
2023-12-01 (across 3 sources on this page)
Specifications (14 fields, click to expand)

Straight from the vendor datasheet. Dense throughput shown first; sparse (2:4) numbers in parentheses where the vendor publishes them. Full datasheet linked below.

Process node
TSMC N5 + N6 (chiplet)
Transistors
153 B
CUDA cores
19,456
TDP
750 W
Memory
192 GB HBM3
Memory bandwidth
5,300 GB/s
PCIe
Gen 5 x16 (896 GB/s)
FP16 (dense)
1,307 TFLOPS
FP8 (dense)
2,614 TFLOPS
INT8 (dense)
2,614 TOPS
Form factor
OAM
Announced
2023-12-06
Released
2024-01-01
Launch price
$15,000 (list)

Source: vendor datasheet

Generation

Foundry & process

Fabbed at
TSMC (all chips TSMC makes →)
Process node
TSMC N5 + N6 (chiplet)
Transistors
153 B

Rack designs that use it

Whole-rack systems buyers order in bulk (NVL72, HGX, TPU pods) that ship with AMD Instinct MI300X inside.

See all rack designs →

Compare with

See every chip comparison →

Benchmarks

Published performance numbers by workload. Vendor datasheet figures where noted; independent measurements otherwise.

WorkloadValueUnitSourceAs of
memory capacity192GB HBM3vendor2023-12-01
peak tflops fp16 dense1307TFLOPSvendor2023-12-01
peak tflops fp8 dense2614TFLOPSvendor2023-12-01

Sources

Adoption rows appear as we confirm each chip-in-robot pairing from a public source. See every chip we track for the full catalog or the AMD page.