Hitting the Memory Wall: Implications of the Obvious
As noted above, the right solution to the problem of the memory wall is probably something that we haven’t thought of — but we would like to see the discussion engaged. It would appear that we do not have a great deal of time.
Tear down the
memory wall.
HOLONOMIX
The Technology behind the Revolution.
Defining Constraint
Processor performance accelerates faster than memory can keep up. In 1995, Wulf and McKee gave it a name:
The Memory Wall
Brute force became the answer.
Bottleneck-effect
Bandwidth
Move more data to close the gap.
Power
Drive more current to move more data.
Thermal
Dissipate more heat from more power.
Infrastructure
Build more facilities to house more cooling.
Capital
Spend more to build what the last generation required.
Each generation solves last year's constraint and finances next year's crisis.
One structural representation. Eight services eliminated.
The wall does not fall to brute force.
You cannot overpower the constraints of modern computing. But when you change the basis to structure, you render them irrelevant. Changing the geometry of the data is the direct path to democratization.
A Paradigm Shift?...
The bandwidth bottleneck vanishes.
The thermal burden evaporates.
The hardware monopoly shatters.
Scale changes sides.
...No. A Revolution.
HOLONOMIX
When you change the basis, everything after it is arithmetic.
The first implementation is HX-SDP — the Structural Data Platform. One representation replaces eight infrastructure services. The cost drops from $149K to $20K per month. The 200 TB of redundant copies disappear.