AMD Instinct MI350P (CDNA 4, Air-Cooled PCIe)
Product Overview
AMD Instinct MI350P is an air-cooled PCIe add-in card AI accelerator announced by AMD at Advancing AI 2026 (July) — AMD's first PCIe GPU in 4 years, aimed at existing data center upgrades: generative / Agentic AI can be deployed within the power and thermal envelope of existing servers, without building new liquid-cooled facilities.
In essence, the MI350P is a halved MI350X: it belongs to the same CDNA 4 architecture with 4× XCD compute chiplets (TSMC 3nm) + 1× IOD (TSMC 6nm), 73 billion transistors, and 144GB HBM3E, and runs on air cooling alone (TDP 450-600W, dual-slot). AMD emphasizes its leading token output per dollar (token efficiency per unit cost up to 4.2× that of the NVIDIA RTX PRO 6000), making it the "enterprise upgrade" SKU of the MI400 series purchase scenarios.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | CDNA 4 |
| Process | XCD: TSMC 3nm; IOD: TSMC 6nm |
| Transistor Count | 73 billion |
| Compute Cores | 128 compute units / 512 matrix cores |
| Infinity Cache | 128 MB |
| Memory | 144 GB HBM3E |
| Memory Bandwidth | 4 TB/s |
| MXFP4 Compute | 4.6 PFLOPS |
| MXFP6 Compute | 4.6 PFLOPS |
| Model Support | Runs models up to 260B parameters at FP4 |
| TDP | 450-600 W |
| Form Factor | PCIe add-in card (dual-slot, air-cooled) |
| Availability | 2026 H2 (announced at Advancing AI 2026) |
📌 Official positioning: the MI350P targets enterprise customers who "do not build new data centers and upgrade within existing air-cooled racks"; AMD claims token output efficiency up to 5.1× that of the RTX PRO 6000 across different model scales, and token efficiency per unit cost up to 4.2× higher.
MI350P vs MI350X / MI355X
| Metric | MI350X | MI350P | MI355X |
|---|---|---|---|
| Architecture | CDNA 4 | CDNA 4 | CDNA 4 |
| Memory | 288 GB HBM3E | 144 GB HBM3E | 288 GB HBM3E |
| Memory Bandwidth | 8 TB/s | 4 TB/s | 8 TB/s |
| MXFP4 | ~9.4 PFLOPS (estimated) | 4.6 PFLOPS | ~4.6 PFLOPS (estimated) |
| Form Factor | OAM | PCIe (air-cooled) | OAM |
| TDP | ~1,000 W | 450-600 W | ~1,000 W |
| Positioning | Large-scale training/inference | Enterprise installed-base upgrade | Large-scale training/inference |
Enterprise Deployment Advantages
- No liquid cooling required: 450-600W air-cooled operation, compatible with existing server racks and power delivery
- Standard PCIe form factor: plug-and-play, no OAM baseboard needed
- Open software stack: ROCm + AMD Inference Microservices (AIMs), no licensing fees, Day-0 support for mainstream models
- Token efficiency per unit cost: up to 4.2× vs the RTX PRO 6000 (AMD official data)
Use Cases
- ✅ Enterprise generative / Agentic AI inference (existing data center upgrades)
- ✅ RAG / multimodal inference (144GB large memory)
- ✅ Budget-sensitive AI deployments (per-token cost first)
- ❌ Frontier large model training (insufficient compute; requires the MI455X)
- ❌ New AI factories (choose Helios for hyperscale deployments)
Vendor Information
| Item | Details |
|---|---|
| Vendor | AMD Corporation |
| Official Announcement | Advancing AI 2026 (2026-07) |
| Product Page | https://www.amd.com/en/products/accelerators/instinct/mi350.html |
| Availability | 2026 H2 |
Related Products
- AMD MI350 - Same-architecture flagship (CDNA 4)
- AMD MI355X - Same-architecture upgrade
- AMD MI455X - MI400 series flagship
- AMD MI430X - MI400 series HPC variant
- Full comparison table