Best case: code was written by aliens with efficient implementation of thinking algorithms, so there is no need to simulate details of brain chemestry and allow emergent behavior to work itself out.
Revisit this answer for details. Take the specs as:
If you already knew how the brain worked to produce intelligence, writing that program fairly directly would require 1015 FLOPS (Blue Gene/P circa 2007) and 100 Terabytes.
Suppose the smallest item that can be manipulated is the size of a dime. They can be laid out in trays or stacked compactly in rolls if you are careful.
If each dime stores a byte, you normally move them around in bulk and don’t examine the detailed value. But it has a 4×4 grid of dots on each side, or a written hex digit, on each side. Math work could be done on expanded registers (like an abacus?) But main storage is this size.
Each byte is 2½ grams and 500 cubic millimeters. Multiplying out, it totals 50 million cubic meters and 250 billion kilograms. The mass is possible physically, and the size would be 20 of the Great Pyramid of Giza. That’s just the size of the bytes, and warehouses to allow storage and access would be significantly larger.
Copying a block of memory would be the labor (time and logistics) of finding a palette of goods in a warehouse and moving it to another, in a large city. The storage would be over many miles.
I don't think floating-point is that important, and can be a refinement of the abacus technique. A skilled operator using abacus or mental work is limited by how fast he can accept directions. As a bound, let’s over-estimate and say he can do 10 operations per second on bytes, reading the opcodes as distinct pictograms and whizzing through the steps. I think finding the argument values to work with will be a bottleneck. One instruction would need to operate on values large enough to store addresses and reasonable signifacance of quantities, but to get an undisputed lower bound let’s suppose it’s possible with stone-tech and human effort, 10 instructions per second.
1015 instructions would take 3 million man-years.
A million people in a concentrated area is Ancient Rome, not stone age. Say 50,000 dedicated people (more like the pyramid builders) plus the urban infrastructure to feed them. Working 8-hour days, it would take 180 years to compute 1 second of AI program.
As noted, the logistics of bringing the right data and instructions to them would be worse. But you have caching levels, and multiple threads, and designs more suited to the architecture.
Before the AI could think, “um…” the same effort could invent a modern society and build semiconductors.