Dmitriy Sorkin

Making chains not suck at hardware @VelariumOne

[Dmitriy Sorkin @AiKiXD]

This one was hard to publish, actually. I checked and re-checked it multiple times, cause the numbers were so good. Anyway, it's pubic now.

[Velarium @VelariumOne]

We’ve built the fastest RISC-V engine we could find. So we made the benchmark public. Don’t believe us? Run it: https://t.co/QBk3lQiJGG

[Dmitriy Sorkin @AiKiXD]

@ProjectAglaya @emollick It's the job of your dev instructions to dictate the kind of ops you need. Add a ritual at the start of the turn to make a read on the situation and act accordingly.

[Dmitriy Sorkin @AiKiXD]

@a16zcrypto @cdixon Naturally, but the entire Eth ecosystem capacity is what? 30-40k TPS? The dude there implying 100k tps as the floor; realistically it's in the ballpark of 5 MILLION TPS. This just can't land. Luckily, I'm building something that can make this happen.

[Dmitriy Sorkin @AiKiXD]

@gakonst My headless harness was a cron job with tmux virtual typing, ever since codex came out. I guess it's time to ditch that, lol

[Dmitriy Sorkin @AiKiXD]

With latest optimizations, my RISC-V emulator can now spew dense traces at 8-10GHz. Pretty sure that's the fastest one on the planet.
I'll put out a public bench harness in a couple of weeks, once I round up some corners.

[Dmitriy Sorkin @AiKiXD]

@iammadab It's nice, but it's a half way there approach. I'm working at a solution where the RISC-V program is restored, lifted to SSA and recompiled with an optimizing compiler, reaching speeds ~4x than of an AOT transpiler. We'll be releasing public demo in the following weeks.

[Dmitriy Sorkin @AiKiXD]

@ClutchMarkets 0 chance, lol. The signal becomes indistinguishable from the background noise at about 20 lys. And no one is that close.

[Dmitriy Sorkin @AiKiXD]

@openvm_org Impressive. What's the expected ceiling? Do you see a path to multiples of that, say 4x, on a same sized cluster?

[Dmitriy Sorkin @AiKiXD]

@jackoholleran This will also not happen without proper infra. The current one just doesn't fit the bill.

[Dmitriy Sorkin @AiKiXD]

@ssafavi @hosseeb At @VelariumOne we're tackling one of the limitation of infrastructure: execution. We're very early but we've got a concept proof: a micro-benchmark shows a 230x increase in performance of a toy EVM contract execution against revm.

[Dmitriy Sorkin @AiKiXD]

When doing theory and research, Opus isn't much better than Sonnet. Both miss points and make stupid mistakes. But Opus does sound much smarter while at it.

[Dmitriy Sorkin @AiKiXD]

@LayerOnEth @protocollabs That's neat, pushing the heavy load off-chain. At @VelariumOne we're taking the opposite path: make contracts fast enough that more can stay on chain.

[Dmitriy Sorkin @AiKiXD]

@LineaBuild @alexand_belling @eth_proofs Nice move. But why keep a VM in the hot path? Lower to RISC-V for structure, then lower again to native for execution. That’s what we’re building at @VelariumOne.

[Dmitriy Sorkin @AiKiXD]

With LLMs reducing the cost of code to one of coffee, the important question is now not what to write, but what not to.

[Dmitriy Sorkin @AiKiXD]

@a16z Most attempts at this will fail in two ways:
- vibecode their way to death
- middle management will kill of the engineering
The rest will compete on products that can be prompted by a single janitor with a 10 layer chain of command.

[Velarium @VelariumOne]

Today we’re revealing Velarium.

A parallel execution layer for Ethereum chains - built to make existing stacks faster, without changing how they work.

Here's what we're building and why it matters 🧵👇

[Dmitriy Sorkin @AiKiXD]

@Alchemy ERC-4337 unlocked the wallet side, and now agents are free to blast without thinking about keys or gas. Infrastructure should be ready to handle high volumes of traffic: bursts of micro-ops and highly contended shared contracts. Exactly the thing I'm building.

[Dmitriy Sorkin @AiKiXD]

@EliBenSasson Perps are a contention stress test: they hammer hot markets/accounts with RMW, then add liquidation cascades on top. If you lane that into RMW phases and treat the commutative parts (fees, funding) as accumulators, you get more headroom than a “generic parallel EVM”.

[Dmitriy Sorkin @AiKiXD]

@VittoStack @iamjohnellison L1 still has a tight throughput cap. A lot of use cases require much higher volumes than what L1 can accommodate, like intra-day trading. You still have to batch and/or do an L2/app chain. And in some setups, off-chain matching isn’t really acceptable.