Broadcom releases VMware VCF 9.1 with intelligent memory tiering and AMD MI350 support, enabling high-performance on-premises AI clusters.
Why Broadcom Is Steering VCF Toward AI Workloads
VMware Cloud Foundation has always been positioned as the integrated stack for running enterprise workloads on-premises: compute, storage, and networking managed as one platform. The 9.1 release under Broadcom's ownership extends that idea to a workload class that used to live almost exclusively in the public cloud — large-scale AI training and inference. The pitch is that organizations with data residency requirements, existing hardware investments, or predictable long-running jobs can now keep those clusters inside their own data centers without giving up the operational model VCF already provides.
The practical draw here is control. Teams that fine-tune models or run steady inference pipelines often find that renting accelerators by the hour becomes the largest line item in their budget. Owning the infrastructure changes that math, and VCF 9.1 is aimed at the point where that shift becomes operationally realistic rather than a science project.
Intelligent Memory Tiering
Memory is frequently the real constraint in AI infrastructure, not raw compute. Models and their working sets can exceed the capacity of the fastest available memory, forcing awkward tradeoffs between performance and how much you can fit on a node. Intelligent memory tiering addresses this by treating memory as a layered resource: hot, frequently accessed data stays in the fastest tier, while colder data moves to larger, slower capacity automatically.
For platform teams, the value is that this placement is managed rather than hand-tuned. When evaluating tiering in your own environment, it helps to reason about a few things:
- The access pattern of your workloads — training, batch inference, and interactive serving each stress memory differently.
- How much working-set data genuinely needs to sit in the fastest tier versus what can tolerate a slower path.
- Whether the tiering behavior is observable, so you can confirm the automatic placement matches what your jobs actually need.
AMD MI350 Support and Hardware Choice
Adding support for AMD's MI350 accelerators matters as much for procurement as for performance. A validated path to a second accelerator vendor gives buyers leverage and reduces the risk of being locked into a single supply chain — a real concern when high-end accelerators are hard to source. It also means the platform team can standardize their operational tooling while leaving the underlying hardware decision open.
If you are planning around this, treat accelerator support as one input among several. Power draw, cooling, rack density, and driver maturity all shape what a cluster actually costs to run. Validated support inside VCF reduces the integration burden, but the physical realities of hosting dense AI hardware still need to be designed for.
What to Weigh Before Committing
An on-premises AI platform is a long-term commitment, so the decision should rest on workload steadiness and existing constraints rather than the appeal of new features. Predictable, sustained demand and firm data-governance requirements point toward owning the stack; spiky or exploratory workloads often still favor rented capacity.
A sensible approach is to start with a bounded pilot: run a representative workload, measure how the memory tiering behaves under real load, and confirm the accelerator path fits your operational tooling before scaling out. That keeps the evaluation grounded in your own numbers instead of general expectations.