AMD has introduced the Instinct MI350P, a new PCIe AI accelerator aimed at enterprise servers — and the big deal here is simple: this is AMD’s first PCIe Instinct card in around four years.
Instead of forcing companies into expensive new rack-scale systems, the MI350P is designed as a dual-slot, drop-in card for standard air-cooled servers. For businesses in Malaysia and SEA that are trying to run more AI workloads locally — think chatbots, internal copilots, recommendation systems, analytics, or private inference — that matters. Not every company has the budget or infrastructure to jump straight into massive AI clusters.
The MI350P is based on AMD’s CDNA 4 GPU architecture and uses TSMC’s 3nm process for its compute dies. It comes in a 4 XCD configuration, which is basically half of what AMD uses on the higher-end MI350X. There is also a single I/O die built on TSMC’s 6nm process.
On paper, the card packs 128 compute units, 8,192 stream processors, 512 matrix cores, and a peak clock of 2,200MHz. AMD says the chip contains 73 billion transistors. For memory, it carries 144GB of HBM3E on a 4096-bit bus, delivering up to 4TB/s of bandwidth. There is also 128MB of Infinity Cache.
Performance-wise, AMD is pushing this hard for enterprise AI formats. The MI350P supports lower-precision MXFP6 and MXFP4, plus sparsity acceleration for common 8-bit and 16-bit formats. AMD claims around 2,299 TFLOPS, going up to 4,600 peak TFLOPS at MXFP4. That is the kind of number aimed at serious inference workloads, not gaming PCs or normal creator rigs.
Power draw is still very server-class. The MI350P has a 600W total board power rating, uses a 16-pin connector, and can reportedly be configured down to 450W. The card is 10.5 inches long, or 267mm, and uses passive cooling, so this is meant for proper server airflow — not your average tempered-glass desktop build, bro.
The most interesting part for SEA businesses is the PCIe form factor. A lot of regional companies want AI capability, but building new high-density infrastructure is painful and expensive. A PCIe accelerator that can slot into existing compatible servers gives IT teams a more practical upgrade path. For Malaysian firms already running on-prem data rooms, this could be more realistic than buying into huge proprietary AI systems from day one.
AMD is positioning the MI350P against NVIDIA’s H200 NVL, which is also a PCIe accelerator and comes with 141GB of HBM3E. Wccftech notes that H200 NVL cards cost around US$30,000 to US$40,000, which is roughly RM140,000 to RM190,000 before tax, shipping, support, and local enterprise markup. So yeah, none of this is remotely consumer-level money.
AMD says the MI350P is now available through various partners. It also supports AMD’s ROCm platform and an enterprise-ready AI software stack, which is important because hardware alone is not enough in AI deployments. The ecosystem, developer tools, and long-term software support will decide whether companies actually adopt it.
For gamers, this will not boost your FPS directly. But for the wider tech scene, it matters because AI infrastructure is becoming part of everything — cloud gaming services, content tools, customer support, esports analytics, game testing, and even moderation systems. If AMD can make AI acceleration easier to deploy in normal servers, more regional companies may get access to serious AI compute without needing hyperscaler-level budgets.
Source: Wccftech Gaming