20170216 - Sediment Layers


Parallella Blog : Reaching 1M Cores
Interesting articles on the Parallella Blog including A million core RISC CPU. The Does it Work? Yes! slide shows a 128x128 Matmul reaching 12 GFlop/s on the initial 16 core board with a 32 Glop/s capacity (algorithm doing 37.5% math, 62.5% overhead). Amazing that Adapteva managed a 1024 core chip on 16ff+ on a $1 million budget. This specific post however is about how to actually build a million core chip in a near term process. The Minimize Tile Size slide gets to the point: 512 bytes of memory/core, write-only single mesh network, 16-bit integer core. Interesting in that SRAM is 53.3% of die area, they drop memory to 1/128 per node, and ultimately only get a 10:1 scaling with all other changes. Reaching past the point of diminishing returns in ALU/memory ratio. Then 16x scaling via 16 separate chips connected by an interposer (similar to HBM memory).

Programming a real beast with 4x4 interposed chips would be quite interesting. One of the chips would have a 256x256 grid of nodes, so minimum on-chip communication latency horizontally would be at least 256 cycles. With only 512 bytes of memory, the "program" effectively is factored out into the network and the data. Believe it will take generations of humans to really grasp that kind of software programming methodology. I'm certainly struggling with it.

Sediment Layers
At one point I thought about grabbing and developing on the $99 4x4 core parallel board. Ultimately what caused me to say no: ARM A9, "Runs Linux". For practical reasons the 16-core Epiphany RISC chip is effectively only a parallel co-processor. And I'm really looking for a parallel machine that is more like a C64, devoid of layers.

The larger problem I face daily is that the computer industry buries hardware under layers of software faster than Moore's law. Drowning in technical dept and complexity, strangled by people who use APIs as a tool to force an agenda, ultimately wishing I could return to the era where the dystopia was a novel and not a roadmap.

The larger battle is really important, as ultimately the line between computers and the mind will blur. Even now your 'data' is an extension of your memory. Fast forward to a future where the mind is fully augmented by a machine. Then apply current market trends. Do you really want the company that manufactured part of your mind to have the ability to read your thoughts and distribute them to paying 3rd parties? Do you really want the government to be able to read your thoughts? Do you really want your mind to be updated at the will of that company? Do you really want your mind to be conditionally accessable depending on if you pay for it? Etc. The current world is already badly in need of a "bill of digital rights" to ensure that the rule of how markets optimize does not fully erode our basic freedoms.