Shatter the AI Memory Wall

Maximize return on CAPEX

GPUs, CPUs and embedded processors are leading the innovation charge for AI and other domain-specific workloads. However, they will only achieve their true processing capability when both local and off chip memory can keep pace. Current memory interface solutions and even some on the horizon are still falling short of keeping up with these processing capabilities and requirements.

CrossFire’s Bridgelet™ and Direct Chiplet Interface™ (DCI) solutions shatter the current processor to memory constrains with a 10X improvement in performance, energy savings and cost. How is this achieved? This patented approach makes off chip memory appear to the processor as if it is on chip. The only way to do that is to extend the internal memory fabric off chip.

Screen Shot 2019-02-03 at 12.08.32 PM.jpg

Bridgelet
 Direct Chiplet Interface (DCI)


Features
 Benefits
 Features
 Benefits

Disaggregates system I/O from memory functions
Allows chiplet reuse within or across product families
Ultra Short Reach (<500µ)
Lower power and area; Higher performance 

Die scale interconnect (sub-micron)
Higher density and performance; Lower power
Protocol Agnostic
One PHY for multiple protocols

Enables “Beyond the Reticle” constraint
Extend SoC designs
Die Scale Interconnect (sub-micron)
Higher density and performance; Lower power

Maximize Return on CAPEX

As the need for processing power has exploded, so has Capital Expenditures (CAPEX) on data center hardware. Some industry reports suggest these CAPEX expenditures will have surpassed $1 trillion by the end of 2029 with over half predicted to have been spent on AI and accelerated computing hardware. However, studies show that accelerated computing hardware is underutilized by 50% due to the memory bottleneck. Shattering the bottleneck will enable maximum revenue from CAPEX.

Lower Energy Use - Minimize OPEX

Lowering power use is not a new goal, but the pain points have grown. As the use of AI has increased exponentially so has the energy used for data transfer between processor to memory. Industry estimations calculate the energy used for this data transfer can range between 30 – 50% of total energy consumption. CrossFire’s solutions offer a 10X reduction in energy use for this growing operating expense.

Extending Memory Fabric

CrossFire’s patented Bridgelet disaggregates system I/O from memory functions. This enables the memory fabric to be extended off chip but appear to the processor as if on chip. This is achieved through a new approach called “Wire Abundance”. There are multiple benefits to this approach – improved performance, reduction in power and an unheard-of area savings.