docs / benchmarks / llm

Benchmarks index
  1. Overview
  1. Exemptions
  2. Methodology

aot

  1. Overview
  2. Components

codec

  1. Overview
  2. Base32
  3. Base64
  4. Deflate (dev.cajeta.codec)

codec / csv

  1. CSV

codec / json

  1. JSON

collection

  1. Overview
  2. ArrayList
  3. BPlusTree
  4. Cache
  5. Collectors
  6. HashMap
  7. HashSet
  8. Heap
  9. ImmutableList
  10. ImmutableMap
  11. ImmutableSet
  12. LinkedList
  13. RedBlackTree
  14. Sort

collection / ltm

  1. LtmBPlusTree

concurrent

  1. Overview
  2. AtomicInt32
  3. AtomicInt64
  4. Channel
  5. Fiber
  6. FiberLocal
  7. Lock
  8. Mutex
  9. RwLock
  10. Semaphore
  11. Tasks

gfx

  1. Overview
  2. Texture2D

hash

  1. Overview
  2. DefaultHasher
  3. Hash
  4. SipHash
  5. XXHash3

http

  1. Overview
  2. Client
  3. Server
  4. Server-sent events
  5. Start-up and footprint
  6. TLS
  7. WebSocket

io

  1. Overview
  2. Buffer

io / file

  1. Directory walk
  2. File
  3. FileInfo
  4. FileReader
  5. FileWriter
  6. Path
  7. Watcher

io / net

  1. BufferPool
  2. ByteBuffer
  3. Server
  4. TcpListener
  5. TcpStream
  6. UdpSocket

io / net / uri

  1. Uri
  2. UriBuilder

lang

  1. Overview
  2. Allocation and ownership
  3. Calls and dispatch
  4. Exceptions
  5. Guid
  6. Loops
  7. Math and SIMD
  8. Optional
  9. Pair
  10. Slice
  11. String
  12. StringBuilder
  13. Whole programs: the Benchmarks Game

lang / stream

  1. ArrayStream
  2. Stream

llm

  1. Overview
  2. Decode
  3. Kernel parity
  4. Load and first token
  5. Long context
  6. Memory
  7. Prefill

math

  1. Overview
  2. Ray
  3. Rotation
  4. Tensor
  5. Transform

math / fft

  1. Fft

math / linalg

  1. LinAlg

math / npio

  1. Npy

math / poly

  1. Poly

math / random

  1. Generator

math / stats

  1. Stats

ml

  1. Overview
  2. Classical models
  3. Neural models
  4. XGBoost

nucleo

  1. Overview
  2. Column
  3. Fuse
  4. Table
  5. Tape
  6. Transforms

primavera

  1. Overview
  2. Build time
  3. Dependency injection at scale
  4. REST endpoint
  5. Start-up and footprint
  6. Templated page

process

  1. Overview
  2. Command
  3. Process

reflect

  1. Overview
  2. Class

search

  1. Overview

search / distance

  1. Distance

search / fuzzy

  1. Matcher

search / ngram

  1. Index

time

  1. Overview
  2. Clock
  3. DateTimeFormatter
  4. Duration
  5. Instant
  6. LocalDate
  7. LocalDateTime
  8. LocalTime
  9. ZonedDateTime

toolchain

  1. Overview
  2. Binary size
  3. Build tool
  4. Compile time
  5. Debugger
  6. Notebook kernel start-up
  7. Start-up

toolchain / kernel-startup

  1. Overview
  2. Where cold-start time goes

wire

  1. Overview
  2. CompressStream
  3. Encoder
  4. SchemaEncoder

xpu

  1. Overview
  2. Device
  3. KernelBuffer
  4. KernelStream
  5. KernelTimer

xpu / mesh

  1. MeshSimplifier
Machine
no runs yet

dev.cajeta.llm against llama.cpp

Prefill and decode tokens per second, keyed by operating system, driver, model and platform. The matrix below is the page to read; the others are its cuts. Method: measuring.

Page
DecodeDecode tokens per second at one and four sequences
PrefillPrompt-processing tokens per second at 512 and 4096 tokens, keyed like Decode
Load and first tokenTime to load a model from safetensors and from GGUF, and time to the first generated token
MemoryResident memory at a 4096-token context
Long contextDecode at 8k and 32k tokens of context, where attention dominates
Kernel parityThe parity legs: one matrix-multiply shape per row, device-timed on each side’s own stream, joined under the five rules the parity harness enforces

The matrix

One block per model. Rows are operating system × driver × backend; the two column groups are prefill and decode tokens per second; each cell holds cajeta-llm beside llama.cpp with the ratio printed. A row is published only when both platforms ran in the same interleaved session on the same model file and the greedy-agreement gate passed.

prefill · decode · tokens/s

Not yet measured on Proton.

Legend. OS: Ubuntu 26.04 (Proton, native), Ubuntu 24.04 / WSL2 (Phoenix), Windows 11 (Phoenix, native, when both engines’ Windows builds are green). Driver: the NVIDIA driver with its CUDA version, the ROCm version, or the Vulkan driver (RADV, or the vendor ICD) as reported at run time. Model: name, parameters, quantization and file checksum. Platform: cajeta-llm <version> <backend> or llama.cpp <commit> <backend>.

Standing

Written after the first published run.

Source: bench/llm/README.md · 1 min read · 275 words