Strategic Progress towards 120 TOPS/W for AI Inference at the Edge and in Datacenters
Mythic announces the selection of Silicon Storage Technology’s (SST) memBrain technology – a subsidiary of Microchip Technology – to power its next generation of Analog Processing Units (APUs).
By leveraging SST’s onboard non-volatile SuperFlash memory (eNVM), Mythic aims for a record energy efficiency of 120 TOPS/W through analog computation in memory (aCIM), up to 100 times more efficient than a traditional GPU.
These APUs will bridge the gap between edge systems and central infrastructures, providing a compact, low-power AI accelerator particularly suited for inference workloads.
SuperFlash + memBrain: A Proven Foundation for aCIM
With 150 billion SuperFlash units deployed, SST has become a global eNVM memory reference:
- utilized in 10 major foundries,
- adopted in automotive, industrial, consumer electronics industries,
- known for its reliability, endurance, and long-term retention.
For Mythic, the choice of memBrain is based on:
- 8 bits per cell storage (8bpc),
- reading current < 10nA,
- 10 years of retention under nominal conditions,
- 100,000 endurance cycles,
- MAC operations in a single cycle for analog computation.
Result: a highly efficient solution, ideal for neuromorphic analog chips.
Goal: Faster, Greener, More Scalable AI
The combination of Mythic and memBrain paves the way for:
- a 50 – 100 fold reduction in consumption depending on workloads,
- a significant decrease in the external DRAM requirement,
- local execution of demanding AI models with minimal thermal budget.
These advantages apply to:
- intelligent industrial sensors,
- ADAS/AV automotive platforms,
- low-power datacenter AI accelerators.
Microchip – Mark Reiten, VP Edge AI: “memBrain provides Mythic with a highly efficient foundation to advance AI inference at the edge and in datacenters.”
Mythic – Dr. Taner Ozcelik, CEO: “After analyzing existing eNVM technologies, memBrain was the only one capable of meeting our performance and consumption requirements.”
Roadmap: 40nm – 28nm – 22nm
The memBrain technology:
- is currently produced in 40nm and 28nm,
- will transition to 22nm to improve density, speed, and efficiency,
- leveraging standard SuperFlash modules for easy integration.
This information aligns with a clear dynamic: replacing power-hungry digital AI architectures with highly efficient analog accelerators.
With memBrain, Mythic positions itself to:
- offer high-performance AI on battery,
- reduce hardware costs,
- push the limits of TOPS/W ratio,
- provide a credible alternative to GPUs for inference.
Microchip Technology Inc. | Silicon Storage Technology (SST) | Mythic
