Devotd Github
NewInference & serving

nicejunjie/strata-rdna3 — a Strata build tuned for RX 7900 XTX

Covers the AMD RDNA3 tuning in nicejunjie/strata-rdna3 and the throughput measurements reported in the repository.

Repositorynicejunjie/strata-rdna3 ↗

Covers the AMD RDNA3 tuning in nicejunjie/strata-rdna3 and the throughput measurements reported in the repository.

GitHub description: Private Strata customizations for 9950x3d2+dual 7900xtx; upstream Niko1221/Strata

stars
★ 55
Forks
12
Language
C++
License
MIT
Read original source ↗View on GitHub ↗

nicejunjie/strata-rdna3 is a user version of the upstream project Niko1221/Strata, tuned for AMD RDNA3 graphics cards, and in particular the RX 7900 XTX. The repository description states that the tuning targets a 9950x3d2 CPU with a dual 7900xtx setup. It is written in C++ and uses the MIT license. There is no latest release registered.

The repository presents the results of testing multiple quantization formats and context lengths on single- and dual-card setups in its README. According to the repository post, prefill throughput is over 3500 per second on a single card and over 7000 per second on dual cards. English prose decoding is given as 81 tokens per second on a single card and 127 tokens per second on dual cards, and coding decoding is described as 20~25 percent faster than prose. The README also explains the measurement procedure for checking input length and retrieval results, and covers tensor parallelism across cards and processing paths that depend on prompt length.

These figures are the values presented in the repository post and the README. Although the README contains tests of several formats and context lengths along with the measurement procedure, FACTS holds no information that asserts the results apply unchanged to other configurations. It is therefore more accurate to treat this fork as a record of measurements tuned for the RX 7900 XTX configuration.