View @vitalik.ca's post on Firefly

Details

Now, scaling.

There are two buckets here: short-term and long-term.

Short term scaling I've written about elsewhere. Basically:

* Block level access lists (coming in Glamsterdam) allow blocks to be verified in parallel.

* ePBS (coming in Glamsterdam) has many features, of which one is that it becomes safe to use a large fraction of each slot (instead of just a few hundred milliseconds) to verify a block

* Gas repricings ensure that gas costs of operations are aligned with the actual time it takes to execute them (plus other costs they impose). We're also taking early forays into multidimensional gas, which ensures that different resources are capped differently. Both allow us to take larger fractions of a slot to verify blocks, without fear of exceptional cases.

There is a multi-stage roadmap for multidimensional gas.

First, in Glamsterdam, we separate out "state creation" costs from "execution and calldata" costs. Today, an SSTORE that changes a slot from nonzero -> nonzero costs 5000 gas, an SSTORE that changes zero -> nonzero costs 20000. One of the Glamsterdam repricings greatly increases that extra amount (eg. to 60000); our goal doing this + gas limit increases is to scale execution capacity much more than we scale state size capacity, for reasons I've written before ( ethresear.ch/t/hyper-scaling-… ). So in Glamsterdam, that SSTORE will charge 5000 "regular" gas and (eg.) 55000 "state creation gas".

State creation gas will NOT count toward the ~16 million tx gas cap, so creating large contracts (larger than today) will be possible.

One challenge is: how does this work in the EVM? The EVM opcodes (GAS, CALL...) all assume one dimension. Here is our approach. We maintain two invariants:

* If you make a call with X gas, that call will have X gas that's usable for "regular" OR "state creation" OR other future dimensions

* If you call the GAS opcode, it tells you you have Y gas, then you make a call with X gas, you still have at least Y-X gas, usable for any function, _after_ the call to do any post-operations

What we do is, we create N+1 "dimensions" of gas, where by default N=1 (state creation), and the extra dimension we call "reservoir". EVM execution by default consumes the "specialized" dimensions if it can, and otherwise it consumes from reservoir. So eg. if you have (100000 state creation gas, 100000 reservoir), then if you use SSTORE to create new state three times, your remaining gas goes (100000, 100000) -> (45000, 95000) -> (0, 80000) -> (0, 20000). GAS returns reservoir. CALL passes along the specified gas amount from the reservoir, plus _all_ non-reservoir gas.

Later, we switch to multi-dimensional _pricing_, where different dimensions can have different floating gas prices. This gives us long-term economic sustainability and optimality (see vitalik.eth.limo/general/2024… ). The reservoir mechanism solves the sub-call problem at the end of that article.

Now, for long-term scaling, there are two parts: ZK-EVM, and blobs.

For blobs, the plan is to continue to iterate on PeerDAS, and get it to an eventual end-state where it can ideally handle ~8 MB/sec of data. Enough for Ethereum's needs, not attempting to be some kind of global data layer. Today, blobs are for L2s. In the future, the plan is for Ethereum block data to directly go into blobs. This is necessary to enable someone to validate a hyperscaled Ethereum chain without personally downloading and re-executing it: ZK-SNARKs remove the need to re-execute, and PeerDAS on blobs lets you verify availability without personally downloading.

For ZK-EVM, the goal is to step up our "comfort" relying on it in stages:

* Clients that let you participate as an attester with ZK-EVMs will exist in 2026. They will not be safe enough to allow the network to run on them, but eg. 5% of the network relying on them will be ok. (If the ZK-EVM breaks, you _will not_ be slashed, you'll just have a risk of building on an invalid block and losing revenue)

* In 2027, we'll start recommending for a larger minority of the network to run on ZK-EVMs, and at the same time full focus will be on formally verifying, maximizing their security, etc. Even 20% of the network running ZK-EVMs will let us greatly increase the gaslimit, because it allows gas limits to greatly increase while having a cheap path for solo stakers, who are under 20% anyway.

* When ready, we move to 3-of-5 mandatory proving. For a block to be valid, it would need to contain 3 of 5 types of proofs from different proof systems. By this point, we would expect that all nodes (except nodes that need to do indexing) will rely on ZK-EVM proofs.

* Keep improving the ZK-EVM, and make it as robust, formally verified, etc as possible. This will also start to involve any VM change efforts (eg. RISC-V)