This one’s a bit out there, but I promise it’s relevant to blockchain tech. Stick with me.
For the longest time, CPUs & GPUs were just one slab of silicon called a silicon die. In the industry, this is called a monolithic die or a monolithic chip. Something like this:
Without going into details about silicon fabrication — basically, silicon always has small defects. The larger the die, the defect rates increase exponentially, on top of being more expensive to produce in the first place. At some point, the die becomes too large and defect rates too high that it becomes infeasible to produce. So, these inefficiencies mean the cost of producing dies increases superlinearly with size.
In 2019, Intel’s fastest CPU was Xeon Platinum, a 28 core CPU costing a staggering $10,000. AMD released their competitor — a 64 core CPU, costing $4,000 and completely obliterated Intel’s very best. Double the performance for half the price. That’s right — a 4x improvement overnight for an industry where you’d expect 2x in 18 months at best. Here’s what AnandTech had to say:
So has AMD done the unthinkable? Beaten Intel by such a large margin that there is no contest? For now, based on our preliminary testing, that is the case. The launch of AMD’s second generation EPYC processors is nothing short of historic, beating the competition by a large margin in almost every metric: performance, performance per watt and performance per dollar.
So how did AMD pull off this miraculous feat? They modularized the monolithic die:
Instead of one giant slab, you now have 8 tiny modules called “chiplets”, and one medium-sized die in the middle coordinating everything containing common elements shared across the chiplets like memory controllers and I/O. All these small dies combined can be larger than a single monolithic die. Also, given the superlinear costs of larger dies, it’d be far cheaper to produce as well. This is how AMD brought twice the performance at half the cost and beat Moore’s Law manifold.
Of course, Intel’s response was to FUD this revolutionary approach, derogatorily calling it “glued together”. Needless to say, they were scrambling to design their own modular architectures. A couple of years later, Intel is now all-in on the modular architecture. Their latest supercomputer chip Ponte Vecchio contains a staggering 47 smaller dies to make one whole.
Obviously, they are not going to call them “chiplets” or “modules” — they have come up with their own terminology — tiles. This is an order of magnitude greater performance than would be possible with a single monolithic die. NVIDIA, AMD, Intel and others — everyone building high-performance processors is going to transition to modular processors or be technologically obsolete. Note that mobile processors can get away with monolithic designs as they are tiny anyway with low-performance, but even mobile SoCs will eventually make the pivot.
So, what does this have to do with blockchains? Those who have been reading my blog know that I have been calling L1s “monolithic blockchains” for several months now. I took my inspiration directly from the silicon industry — now you know!
Just like AMD split up the CPU to unlock impossible efficiency — blockchains are undergoing the same revolution. We’re splitting up chains between execution, security and data availability layers. A monolithic blockchain scales to a few thousand or tens of thousands of TPS at most despite crippling centralization and security compromises. A modular architecture will scale to millions of TPS, while still remaining highly secure and decentralized. It’s orders of magnitude better on scalability, security, and decentralization.
Scroll back up to the earliest AMD Rome die. See those 8 tiny chiplets surrounding the central I/O die? That central chip is the security & DA chain — those 8 chiplets are rollups. This was just the beginning, as Intel’s Ponte Vecchio shows above, you can have 47 dies each with different specific characteristics. We’ll see more efficient solutions like volitions where one execution layer settles on one security chain, but spreads DA across multiple layers. The skies are the limits.
The modular revolution is here for both processors and blockchains. Monolithic processors and monolithic blockchains are technologically obsolete. Sure, the monolithic blockchain incumbents will FUD and lie through their teeth (just like Intel did)— but make no mistake —modular architectures are the only way the blockchain industry will attain global scale.
