● Covered by 1 source · 1 reportMedium impact
Meta to implement AMD’s custom Instinct MI450 accelerators with 144GB HBM4 for specific tasks
Meta will utilize custom AMD Instinct MI450 accelerators with 144GB of HBM4, reducing costs for recommendation workloads but limiting versatility in AI training. This shift may lead to significant material savings while relying on Nvidia for training larger models.
Key points
- Meta's custom accelerators will have 144GB HBM4 memory.
- Use of custom models may save tens of millions in costs.
- Reduced compute power limits training capability for LLMs.
Introduction to Meta's Custom Accelerators
Meta is reportedly planning to use AMD's custom Instinct MI450-based AI accelerators for certain workloads.
These accelerators are optimized specifically for recommendation systems used by Facebook and other social platforms.
Specifications of the Custom Hardware
The Instinct MI450 for Meta will feature 144GB of HBM4 memory compared to the full MI455X's 432GB.
This design will enable a better bandwidth-per-dollar ratio for recommendation workloads, albeit at a reduced performance level.
Cost Implications
The lowered memory capacity and compute power may significantly decrease Meta's bill of materials due to the high cost of HBM4 memory.
By adopting these custom models, Meta could potentially save tens of millions in expenses.
Trade-offs of Customization
While the custom accelerators will consume less power and perform well for recommendation tasks, their limitations make them unsuitable for training large language models (LLMs).
The reduced capacity also poses a challenge for interchangeability as these models can't be repurposed for other intensive tasks.
Conclusion
Meta's shift towards custom AMD hardware marks a notable strategy to cut operational costs, but with potential drawbacks in flexibility for broader AI workloads. This approach could impact how Meta engages with AI model training and inference in the future.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Reporting from
Meta will utilize custom AMD Instinct MI450 accelerators with 144GB of HBM4, reducing costs for recommendation workloads but limiting versatility in AI training. This shift may lead to significant material savings while relying on Nvidia for training larger models.