aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorDavid Phillips <david@sighup.nz>2017-05-23 14:51:31 +1200
committerDavid Phillips <david@sighup.nz>2017-05-23 14:51:31 +1200
commit8cd351ddc25bc5e6c549b99dc5fd3e9607f5a580 (patch)
treefcb632e0c755767744ff3a1770730c3e05caf59e
parent5b2aa52f06dbf3d5aed7165b1b9fe5ca1cf38af4 (diff)
downloadsand-leek-8cd351ddc25bc5e6c549b99dc5fd3e9607f5a580.tar.xz
Update readme with new benchmarks
-rw-r--r--README.md13
1 files changed, 10 insertions, 3 deletions
diff --git a/README.md b/README.md
index cb18582..266cbaf 100644
--- a/README.md
+++ b/README.md
@@ -29,11 +29,18 @@ that keys are sane.
Preliminary benching shows sand leek to be faster than some of the other
similar tools out there when pushing work across cores.
+I have also written a "slightly parallel" base32 algorithm which uses
+SSSE3 if support is given from the compiler and target platform.
+Preliminary benchmarks seem to indicate that this gives a performance
+benefit of between roughly 3% and 30%. Although, these higher
+performance gains seem to be only when running at a reduced worker
+thread count.
+
| CPU(s) | Max throughput | -t |
|---------------------------------------------|---------------:|---:|
-| 2× Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz | 95 MH/s | 32 |
-| Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz | 38 MH/s | 8 |
-| AMD A4-1200 APU with Radeon(TM) HD Graphics | 2.3 MH/s | 2 |
+| 2× Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz | 103.3 MH/s | 32 |
+| Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz | 39.2 MH/s | 8 |
+| AMD A4-1200 APU with Radeon(TM) HD Graphics | 2.6 MH/s | 2 |
| ARMv6-compatible processor rev 7 (v6l) | 0.22 MH/s | 1 |
## Inspiration