WEBVTT

00:00:00.450 --> 00:00:03.860
Seed State OS starts with a practical question.

00:00:03.870 --> 00:00:08.540
When only part of a system changes, why repeat every operation?

00:00:08.710 --> 00:00:12.560
Beyond Brute Force Systems develops retained-state reconstruction for

00:00:12.570 --> 00:00:15.050
compatible applications and devices.

00:00:15.170 --> 00:00:20.210
Keep useful state, identify the required change, reconstruct the affected

00:00:20.220 --> 00:00:23.830
result, and verify it before the application continues.

00:00:23.970 --> 00:00:28.190
Here are the measured results, with each comparison kept separate.

00:00:28.370 --> 00:00:33.283
In the retained CPU repeat study, compatible state validation used about two

00:00:33.293 --> 00:00:37.480
thousand, six hundred and sixty-eight times less CPU time than expanded

00:00:37.490 --> 00:00:38.490
validation.

00:00:38.690 --> 00:00:42.360
The original study recorded about two thousand, five hundred and forty-eight

00:00:42.370 --> 00:00:43.110
times.

00:00:43.370 --> 00:00:46.940
Each used ninety-six fixtures and two thousand, eight hundred and eighty

00:00:46.950 --> 00:00:48.160
checks per arm.

00:00:48.610 --> 00:00:52.040
An existing local atlas and receipts remained available.

00:00:52.050 --> 00:00:55.680
Initial atlas construction was outside this timer.

00:00:55.850 --> 00:01:00.290
In the GPU study, one of four thousand and ninety-six tiles changed.

00:01:00.530 --> 00:01:02.590
The other tiles remained applicable.

00:01:02.690 --> 00:01:06.910
Updating the changed tile required four thousand and ninety-six times less

00:01:06.920 --> 00:01:08.800
dispatched logical work.

00:01:08.930 --> 00:01:13.160
Measured GPU execution was approximately three hundred and fifty-three to

00:01:13.170 --> 00:01:15.650
three hundred and fifty-four times shorter.

00:01:15.910 --> 00:01:21.170
The wider operational GPU update comparison recorded approximately twenty-one

00:01:21.180 --> 00:01:25.290
to twenty-seven times shorter times, depending on the output contract and

00:01:25.300 --> 00:01:25.840
repeat.

00:01:25.970 --> 00:01:28.300
This used warm retained state.

00:01:28.470 --> 00:01:32.940
Initial construction and the independent full-output evaluator were outside

00:01:32.950 --> 00:01:34.470
the operational timer.

00:01:34.590 --> 00:01:38.970
Those costs must be included wherever the deployment requires them.

00:01:39.090 --> 00:01:44.600
On the same FPGA, a separate bounded task reduced median engine cycles from

00:01:44.610 --> 00:01:47.150
two thousand, five hundred and twenty-three to thirteen.

00:01:47.970 --> 00:01:52.130
That is ninety-nine point four eight percent fewer engine cycles.

00:01:52.450 --> 00:01:57.140
For this FPGA experiment, timing the task together with communication and

00:01:57.150 --> 00:02:01.360
verification reduced elapsed time by twenty point nine percent.

00:02:01.370 --> 00:02:04.560
Communication still occupied much of that measurement.

00:02:04.570 --> 00:02:08.570
This is one task comparison, not a summary of Seed State OS.

00:02:08.730 --> 00:02:12.842
Later Mac software receiver tests measured three point three six to sixteen

00:02:12.852 --> 00:02:17.910
point one two times shorter warm request and response times across four input

00:02:17.920 --> 00:02:21.290
conditions, with identical native output values.

00:02:21.450 --> 00:02:25.550
Those framing, parsing and delivery improvements were measured together.

00:02:25.650 --> 00:02:29.880
One-off process start and retained-data loading were measured separately.

00:02:29.970 --> 00:02:33.840
This was a software receiver comparison, rather than an FPGA or

00:02:33.850 --> 00:02:35.790
electrical-power test.

00:02:36.010 --> 00:02:38.840
Reuse can build across requests.

00:02:38.850 --> 00:02:43.080
More compatible bases in a qualified collection give later inputs more

00:02:43.090 --> 00:02:45.080
retained structure to use.

00:02:45.090 --> 00:02:48.750
Previously successful transforms can be nominated again.

00:02:48.970 --> 00:02:54.170
Retained transforms cut recurring producer cycles by thirty-one point five one

00:02:54.180 --> 00:02:54.720
percent.

00:02:54.730 --> 00:02:59.460
A sixteen-slot collection used seventeen point zero one percent fewer resident

00:02:59.470 --> 00:03:03.200
active cycles than an optimised conventional cache.

00:03:03.410 --> 00:03:07.910
In a separate growing-stream study, combining checked-history reuse with route

00:03:07.920 --> 00:03:13.170
nomination reduced conservative total time by fourteen point seven three

00:03:13.180 --> 00:03:17.070
percent against size-based selection with repeated history scans.

00:03:17.250 --> 00:03:19.960
Setup and diagnostics were included.

00:03:19.970 --> 00:03:24.680
All seven thousand, six hundred and eighty fresh reconstruction checks passed.

00:03:24.890 --> 00:03:29.020
Every new transition still required fresh permission and validation.

00:03:29.170 --> 00:03:33.440
The advantage depended on the workload; individual requests did not

00:03:33.450 --> 00:03:35.040
continually get faster.

00:03:35.050 --> 00:03:39.400
For a synthetic versioned-data workload, the archive and state component

00:03:39.410 --> 00:03:43.840
retained fifty-seven point three percent fewer logical bytes than trained

00:03:43.850 --> 00:03:48.140
Zstandard, and thirty-seven point nine percent fewer than compacted Borg.

00:03:48.150 --> 00:03:51.600
Zstandard completed that comparison faster.

00:03:51.610 --> 00:03:56.400
All nine thousand, two hundred and sixteen fresh reconstruction checks passed.

00:03:56.690 --> 00:04:00.880
Logical data size and allocated disk space are different measures.

00:04:00.970 --> 00:04:05.160
A later optimisation reduced complete pipeline time by twenty point two

00:04:05.170 --> 00:04:09.930
percent against the preceding Seed component, with unchanged retained data

00:04:09.940 --> 00:04:10.370
size.

00:04:10.570 --> 00:04:14.710
It retained fifty-seven point three percent fewer logical bytes than trained

00:04:14.720 --> 00:04:18.680
Zstandard, while taking forty-one point one percent more time.

00:04:18.890 --> 00:04:22.640
That gives customers a measured storage and time trade-off.

00:04:22.650 --> 00:04:25.470
Correctness stays alongside performance.

00:04:25.610 --> 00:04:30.770
The GPU study recorded one hundred and twenty exact trial results and refused

00:04:30.780 --> 00:04:32.610
twenty hostile controls.

00:04:32.770 --> 00:04:37.650
The FPGA comparison completed two hundred and fifty-six requests with zero

00:04:37.660 --> 00:04:38.750
wrong outputs.

00:04:38.930 --> 00:04:43.540
These studies do not combine into one multiplier or establish a whole-device

00:04:43.550 --> 00:04:44.870
electricity saving.

00:04:45.010 --> 00:04:49.370
The opportunity is to find repeated state work in application runtimes,

00:04:49.410 --> 00:04:53.390
graphics, digital twins, automation or devices.

00:04:53.570 --> 00:04:56.060
Begin with one valuable operation.

00:04:56.210 --> 00:05:00.710
Measure the complete workflow on the intended platform, including preparation,

00:05:00.730 --> 00:05:03.620
checking, reconstruction and fallback.

00:05:03.770 --> 00:05:07.730
The proposed licensing route supports application deployments, runtime

00:05:07.740 --> 00:05:10.760
integrations and device-volume agreements.

00:05:10.770 --> 00:05:15.640
Agree the scope, evaluate the customer workload, then set the commercial terms

00:05:15.650 --> 00:05:17.640
around the accepted deployment.

00:05:17.690 --> 00:05:22.460
A separate historical model estimated energy savings using assumed workload

00:05:22.470 --> 00:05:23.060
shares.

00:05:23.210 --> 00:05:27.902
Its three scenarios gave eighty-nine point seven two, ninety-four point seven

00:05:27.912 --> 00:05:28.330
one, and ninety-seven point seven one percent.

00:05:31.010 --> 00:05:35.790
These are conditional estimates from work, timing and byte-avoidance proxies.

00:05:35.970 --> 00:05:40.350
They include assumed unchanged overhead of ten, five and two percent

00:05:40.360 --> 00:05:41.570
respectively.

00:05:41.730 --> 00:05:45.530
They are not electrical measurements or promises for every device.

00:05:45.690 --> 00:05:47.420
Seed State OS.

00:05:47.620 --> 00:05:49.650
Retain useful structure.

00:05:49.690 --> 00:05:50.810
Focus the work.

00:05:50.970 --> 00:05:52.430
Verify the result.

00:05:52.650 --> 00:05:55.120
UK patent application pending.

00:05:55.130 --> 00:05:57.600
Beyond Brute Force Systems Limited.

00:05:57.610 --> 00:06:01.240
Bring us a state workload, and let us measure the opportunity.