HYLE

Θεῶν δημιούργημα

Divine Craft

Built for AI inference, resource streaming and low-latency services. Measured wins against SOTA performance libraries across 18 equivalent workloads.¹

Explore the performance Explore the engineering

BENCHMARKED AGAINST SOTA

The best libraries. A higher bar.

Highway, Boost.Beast, Abseil, Folly, oneTBB: SOTA libraries built around performance. HYLE goes head to head with them under equivalent functionality and semantics, with 18 wins across key workloads.

18external benchmark wins

SIMD

Peak advantage

7.463×

SIMD sorting

VQSort partial / full sorting

vs. Highway VQSort

1.147–7.463×

io-http

Peak advantage

3.065×

HTTP streaming

Fragmented HTTP/1 streaming lifecycle

vs. Boost.Beast

3.030–3.065×

io-websocket

Peak advantage

3.354×

WebSocket masking

WebSocket masking, 124 B–1 MiB

vs. Boost.Beast

2.005–3.354×

sched

Peak advantage

1.663×

Request scheduling

Request fan-out + collect on four P-cores

vs. Marl

1.646–1.663×

Each headline is the upper end of the range for its stated workload and reference implementation.

These implementations are already highly optimized. Even a few percentage points of further improvement matter. From incremental gains to multiples, every result below keeps its workload, competitor and full reported range in view.

Full performance comparison · 18

Scroll horizontally to see each workload, reference and result.

HYLE external benchmark wins under equivalent functionality and semantics
Module	Winning workload	Reference implementation	HYLE advantage
container-intrusive	Non-owning list: stable sort + merge	Boost.Intrusive	1.088–1.119×
util-core	Bound-member FunctionRef traversal calls	Abseil FunctionRef	1.245–1.264×
error	Owning failure, first-error, Join failure, Result<T&> borrow	Abseil Status/StatusOr	1.21–4.78×
timer-queue	Complete lifecycle of 64 timers	Seastar timer queue	1.129–1.190×
mem-reclaim	Immutable snapshot reclamation read	Folly reclamation	1.365–1.381×
mem	Logical range allocator: alignment, split, coalesce, ownership	OffsetAllocator, equivalent-contract adapter	1.031–1.043×
mem-vm	Model-loader virtual range lazy release	llama.cpp vector rebuild route	1.031–1.047×
SIMD	VQSort partial / full sorting	Highway VQSort	1.147–7.463×
container	Model-cache LRU: lookup, insert, weight eviction	Abseil adapter	1.222–1.273×
container-queue	Unbounded typed MPMC, 4 producers / 4 consumers	oneTBB queue	1.261–1.289×
container-text	Base64 encoding of small 20-byte records	simdutf	1.033–1.045×
sched	Request fan-out + collect on four P-cores	Marl	1.646–1.663×
sched	Completion wake	Marl	1.0266–1.0285×
io-http	Fragmented HTTP/1 streaming lifecycle	Boost.Beast	3.030–3.065×
io-tls	Verified asynchronous TLS handshake	Asio, same OpenSSL	1.099–1.102×
io-http2	64-stream small-response multiplexing + flow control	nghttp2	1.466–1.478×
io-transfer	64 MiB HTTP Range resume + verification	libcurl	1.206–1.247×
io-websocket	WebSocket masking, 124 B–1 MiB	Boost.Beast	2.005–3.354×

ENGINEERING BEHIND THE PERFORMANCE

Performance with the whole system in mind.

Memory, containers, SIMD, scheduling and I/O share allocator, error and lifecycle conventions. Build on a coherent foundation and spend less time reconciling separate subsystems.

Cancellation has a complete lifecycle.

Cancellation, timeouts, late completions and resource release have explicit ownership rules. Async work has a defined path through completion, draining and shutdown.

Keep resources under your control.

Inject allocators, set queue capacities and apply memory quotas. Backpressure and allocation failures have explicit outcomes that the application can handle.

Test the paths where things go wrong.

Validation covers malformed input, allocation failures, concurrency races and complete flows across public APIs, using fuzzing, fault injection and memory and thread sanitizers.

Bring only what you need.

Choose components through a single dependency graph, with locked third-party versions and offline builds. Independent consumer projects check public interfaces and integration boundaries.

TECHNOLOGY PREVIEW

More to come.

HYLE is under active development. Details of its future release and availability will be announced later.