AMD's 256-core EPYC hits 91% of RTX 5090 bandwidth — LLMs no longer need GPUs

AMD released a 256-core EPYC processor this week, hitting 91% of the consumer flagship RTX 5090's memory bandwidth. On the surface, this looks like a routine CPU upgrade. In reality, we see it as a fork in the hardware roadmap for local LLM inference (running models on local servers rather than the cloud). The old default was "running large models = stacking NVIDIA GPUs." Now AMD is telling the market with a single CPU: there is more than one path.

What this is

EPYC is AMD's server CPU lineup. This new variant packs 256 cores paired with 16-channel DDR5-12800 memory. Memory bandwidth — how much data the CPU can move per second — is the critical bottleneck for LLM inference. Models routinely weigh tens to hundreds of GB, and they do not run fast if the data cannot move. The RTX 5090 is one of today's most powerful consumer GPUs, capable of running large models primarily thanks to its 1.8 TB/s memory bandwidth. AMD's new CPU reaches 91% of that figure, meaning a CPU has come close to consumer flagship GPU bandwidth for the first time.

Industry view

The hardware community broadly reads this as a positive signal for on-premises deployment. For enterprises that do not want to be locked into NVIDIA or worry about data leaving their perimeter, this is a new alternative path.

The pushback is just as clear. Reddit threads are full of self-deprecating jokes: a motherboard to run this CPU plus a 2TB DDR5 kit costs "two kidneys plus a small country's GDP." The more serious critique: even with comparable bandwidth, CPUs still lag dedicated GPUs badly on energy efficiency (compute per watt) and long-context handling (processing long documents in a single pass). Meanwhile, NVIDIA's H200 and B200 roadmap keeps pushing bandwidth higher — how long AMD's edge lasts is an open question.

Impact on regular people

For enterprise IT: Private AI deployment now has an additional route, but it is not a cost-saving play in the short term — think of it as a strategic "don't put all eggs in one basket."

For working professionals: No immediate impact on day-to-day work. But if your company eventually goes the local AI route, your data stays inside the corporate network rather than passing through the cloud — the boundary changes.

For the consumer market: Consumer GPUs (5090/4090 class) remain the workhorse for local AI. This CPU is still far out of reach for ordinary enthusiasts.