ctLame
14.1×
MP3 · bit-identical to LAME 4.0
45 hours of CD audio in 83 s instead of 19½ minutes – almost 2,000 times real time.
Product
ThinkMeta ConcurrentTasks
ThinkMeta ConcurrentTasks is a C++ runtime that makes fine-grained parallelism worthwhile: it drives many small tasks on a few threads with almost no per-switch cost, so an algorithm can use every core without being rewritten around a new programming model. The science behind it is described under Research.
Four encoders as proof
What fine-grained parallelism gains is shown by four encoders built on ThinkMeta ConcurrentTasks. Each is a C++ port of a well-known reference encoder, writes exactly the same output byte for byte, and spreads the work on a single file over all cores. That is hard because these encoders are sequential by design: every frame and every block depends on the state the one before it left behind.
ctJpeg
6.9×
JPEG · bit-identical to jpegli
A 39-megapixel photo in 0.11 s – with jpegli’s smaller files, 3.4 to 4.7 times as fast as libjpeg-turbo on large images.
ctDeflate
4.9×
gzip · bit-identical to libdeflate 1.26
One 1 GB file at 303 MB/s at level 9, streaming with asynchronous I/O – a 5.2 GB file needs less than 150 MB of memory.
ctOpus
4.4×
Opus · bit-identical to opusenc
At maximum complexity twice as fast as opusenc at minimum complexity. Live: 1,024 streams in real time, one task each, without a missed deadline.
Same bits. A fraction of the time.
Looking for first adopters
We are starting a first-adopter programme for teams who want to run ThinkMeta ConcurrentTasks on a real, compute- or latency-critical workload – game engines, medical imaging and signal processing, simulation, real-time media – and help shape the first release. Get in touch.
Status
ThinkMeta ConcurrentTasks is in active development; the Windows build is the current focus.
Roadmap
- Windows version – end of 2027
- Embedded systems – end of 2028
- Xbox and PlayStation – end of 2029
Target dates.