With up to 288 GB memory and 20 PFLOPS performance, AMD’s latest Instinct GPUs give Nvidia’s Blackwell a run for its money and big companies like Oracle are already on board.
Key Highlights –
- AMD launches MI350 and MI355X AI chips under the Instinct brand
- MI355X achieves up to 20 PFLOPS FP4 performance, supports 288 GB HBM3E memory
- Power-hungry: 1,400 W liquid cooled MI355X; 1,000 W MI350X with air cooling
- Built on CDNA 4 architecture supporting FP4, FP6, FP8, FP16
- Matches Nvidia’s Blackwell chips in GPU specs and dense workloads
- Supported by major OEM partners like Dell, HPE, Cisco, Oracle, Supermicro
- Oracle signed deal for 30,000 MI355X GPUs for AI clusters
- Part of AMD’s annual AI chip roadmap and expanding software ecosystem
Overview
AMD has introduced the MI350 and MI355X Instinct GPUs, its next generation AI chips aimed at enhancing the performance and memory capabilities of its data center lineup. The MI355X, in particular, delivers up to 20 PFLOPS of FP4 performance, thanks to its support for 288 GB of HBM3E memory a move designed to rival Nvidia’s high end Blackwell Ultra series.
AMD’s new CDNA 4 architecture, these chips also support newer precision formats like FP6 and FP4, which are growing in importance for AI inference workloads. Industry benchmarks show that MI355X competes strongly with Nvidia’s GB200 Superchip for large scale LLM inference tasks.
The chips consume significant power, with the MI355X drawing around 1,400 W and requiring liquid cooling, while the MI350X operates at 1,000 W with air cooling . Both feature an impressive 8 TB/s memory bandwidth and up to 40 PFLOPS FP4 performance in multi GPU setups.
Market Strategy & Ecosystem
AMD is taking its AI ambitions seriously by working with top tier OEM’s like Dell, Hewlett Packard Enterprise, Cisco, Oracle, and Supermicro to integrate MI350 series chips into AI servers and cloud platforms.
A standout endorsement comes from Oracle, which recently struck a multi-billion dollar deal to deploy 30,000 MI355X GPUs in its data center cluster, giving AMD a strong foothold against Nvidia’s dominance.
See Also – HP Launches Google Beam Powered 3D Video Calling Setup at $25,000
Why It Matters
- Closing the Nvidia gap: AMD’s MI350-series matches or surpasses Nvidia’s Blackwell GPUs on critical specs like memory bandwidth and FLOPS
- Advanced precision support: Adding FP6 and FP4 gives AMD an edge in inference tasks, potentially boosting energy and compute efficiency
- Partner momentum: Backing from major OEM’ and Oracle signals strong commercial adoption and trust
Summary at a Glance
| Feature | MI350X | MI355X |
|---|---|---|
| Memory | 288 GB HBM3E | 288 GB HBM3E |
| Memory Bandwidth | 8 TB/s | 8 TB/s |
| FP4 Performance | ~18 PFLOPS | ~20 PFLOPS |
| Power | 1,000 W (air) | 1,400 W (liquid) |
Read Also
- Apple’s First Foldable iPhone Might Launch in 2026 – Here’s What the Leaks Say
- 16 Billion Passwords Leaked & Cock.li Email Hack: What You Need to Know Now
- Donkey Kong Bananza: Nintendo’s Bananas-Powered Switch 2 Adventure
- 2026 Nissan Leaf Reinvented as SUV with 303-Mile Range
- Tesla to Pause Model Y and Cybertruck Production at Giga Texas for One Week