@GergelyOrosz: More engineers should do this kind of napkin math and ask: “what is the *theoretical* limit of throughput for this syst…
Summary
The tweet highlights a fireside chat at AIE Worlds Fair where Simon Eskildsen emphasizes the value of napkin math for estimating theoretical system throughput limits and understanding why actual performance differs.
View Cached Full Text
Cached at: 07/01/26, 08:04 AM
More engineers should do this kind of napkin math and ask: “what is the theoretical limit of throughput for this system? And why is it lower, in practice?”
Smashing fireside with @Sirupsen at AIE Worlds Fair
Steve Derico (@stevederico): Simon Eskildsen @Sirupsen w/ @GergelyOrosz at @aiDotEngineer World’s Fair
“I hate benchmarks…this should take 10ms in napkin math”
Napkin Math = raw material costs > bandwidth to DRAM / NVME / EBS volume > roundtrip to s3 time + cost > 1GB of memory, S3, disk > spot vs
Similar Articles
@GergelyOrosz: That @Sirupsen measured, published then memorized about 3x these numbers on networking+latency "basics" is only surpris…
Gergely Orosz praises Sirupsen for his deep understanding of networking and latency basics, referencing an interview where Sirupsen used napkin math to build a 10x cheaper or faster product.
@nrehiew_: This was great! A fun followup is given some arbitrary topology and architecture, compute what is the expected compute/…
A tweet reply to a blog post about collective communication in TPU/GPU clusters, suggesting a follow-up to compute expected compute/communication duration for a 1T MoE on Blackwell NVL72 and estimate theoretical MFU and bottlenecks.
@charles_irl: typical engineering velocity with agents
A tweet from @charles_irl discusses typical engineering velocity when using AI agents, likely sharing insights on the topic.
The Final Bottleneck
A reflective blog post on how AI acceleration in code generation overwhelms review processes, creating a new bottleneck in software engineering. Draws parallels to historical industrial bottlenecks and suggests throttling input as a necessary response.
@bcherny: Something I have been thinking about: in the past, the best engineers I knew spent a lot of time automating their work …
A thought-provoking thread arguing that automating engineering tasks, especially with AI agents like Claude, is more important than ever. It advocates encoding domain knowledge as infrastructure to accelerate development and enable contributions from non-engineers.