VRAM
192 GB
TDP
750 W
MSRP
$18k
Perf/W
0.10 tok/s/W
Cost/1K tok
$0.0026/k
Tested
2025-07-11
AMD Instinct MI300X 192GB has a source-attributed result of 72.0 tok/s on Llama 3 70B Q4 (batch 1, 8192-token context, Q4_K_M; runtime not documented). This is a reference report, not an independently verified lab result. A 70B Q4_K_M artifact needs roughly 40GB for weights alone; smaller memory configurations require explicit offload or model splitting and do not establish full-GPU residency.
AMD Instinct MI300X 192GB — CDNA3-architecture datacenter APU. 192GB HBM3 at 5.3 TB/s, 750W TDP. AMD's answer to the H100/H200 for AI workloads. ROCm software stack maturing through 2025-2026.
192GB VRAM at competitive bandwidth makes this a credible H200 alternative for inference. vLLM ROCm support is production-grade. Llama 3.1 70B Q4 runs at ~120+ tok/s. The software tax (ROCm vs CUDA) is the real consideration — fewer engines support it, and community tooling is thinner. Best for operators who already have AMD infrastructure or who are specifically chasing $/VRAM ratios.
Best for organizations with existing AMD infrastructure or those specifically chasing $/VRAM ratios at datacenter scale. The software maturity gap vs NVIDIA is real but narrowing.
70B FP16 or smaller quantized artifacts with measured context headroom. Standard 405B Q4 exceeds 192GB before runtime overhead.
ROCm driver compatibility with specific kernel versions — pin your ROCm + kernel combination once stable. vLLM ROCm branch occasionally lags upstream.
Ubuntu 22.04 LTS (reference), RHEL 9, other Linux (ROCm-supported), Windows (unsupported), macOS (unsupported)
750W TDP. Datacenter power and cooling required. Not suitable for residential. ROCm power management less mature than NVIDIA's — expect higher idle draw.
AMD Instinct MI300X 192GB with 192GB VRAM at 750W TDP, scored across 12 workloads with 20 benchmark records.
Reference workload
DeepSeek-R1 7B
215 tok/s
Quantization
Q4_K_M
8K context · batch 1
| Workload | Score | Quant | Context | σ | Status | Tested |
|---|---|---|---|---|---|---|
| Llama 3 70B Q4 llm | 72.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2025-07-11 |
| Llama 3 70B Q8 llm | 48.0tok/s | Q8_0 | 8K | , | Curated Aggregate | 2025-07-15 |
| Llama 3 8B FP16 llm | 260.0tok/s | FP16 | 8K | , | Curated Aggregate | 2025-07-30 |
| SDXL image gen image | 22.0img/min | FP16 | , | , | Curated Aggregate | 2024-12-22 |
| Embedding throughput embedding | 9600.0emb/s | FP16 | 1K | , | Curated Aggregate | 2026-02-12 |
| Qwen 2.5 14B llm | 142.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2024-12-04 |
| DeepSeek-R1 7B llm | 215.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2026-03-19 |
| Gemma 2 9B llm | 188.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2025-12-08 |
| Mistral 7B llm | 268.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2025-05-19 |
| Llama 3 8B Q4 llm | 122.0tok/s | Q4_K_M | 128K | , | Curated Aggregate | 2025-04-08 |
| Whisper transcription audio | 95.0x RT | FP16 | , | , | Curated Aggregate | 2025-01-12 |
| Asure 12B llm | 130.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2025-05-01 |
There is insufficient comparable evidence to score AMD Instinct MI300X 192GB. A score needs results for multiple devices with matching workload, runtime version, quantization, context and batch size.
Planning estimates for 192GB: weights plus at least 2GB or 10% overhead. Actual KV cache depends on model, context and cache format; confirm the artifact before buying.
Trust score
6/10
MyAI rating
Not scored
Runs
Not documented
Freshness
Aging
Source-linked row with explicit verification status.
Record date: 2026-03-19; 174 days old.
Enterprise buying note
Retailer we'd check first
Amazon search plus reseller quotes
For datacenter and accelerator parts, Amazon is useful for spotting live listings, accessories, or used pulls, but serious procurement usually happens through integrators, brokers, or cloud partners.
Enterprise procurement, not retail
This silicon is typically acquired through an authorized OEM partner, system integrator, or hyperscaler reseller. For on-demand access, compare hourly rates at RunPod, Vast.ai, Lambda Labs, or your existing cloud provider before committing to capital expenditure.
There is insufficient comparable evidence for a purchase recommendation. Check model fit, software support and a current seller quote.
Affiliate note: we do not have a device-level ASIN yet, so this opens tagged Amazon search results for the exact product name.