How is GIMPS' sustained throughput defined?

[NBtarheel_33]

So it looks as though the new prime has brought in a considerable amount of new computing power (thanks and welcome to the new users!). The 30-day sustained GIMPS throughput has swollen to a record 380 TFLOP/s, with the theoretical 30-day maximum throughput exploding to 1.25 PFLOP/s. This is the first time that the GIMPS distributed supercomputer has (theoretically) breached the petaflop barrier.

For historical perspective, in November 2005, GIMPS was knocking on the door of 20 TFLOP/s. We now produce nineteen times that: what constituted GIMPS' typical 24-hour throughput a decade ago, is now completed every 76 minutes.

It is worth noting that the actual 30-day throughput has been typically 35-40% of the theoretical maximum. Therefore, as the new users continue to complete assignments, I expect that we should sustain 450-500+ TFLOP/s in the next few weeks.

[Madpoo]

Originally Posted by NBtarheel_33

So it looks as though the new prime has brought in a considerable amount of new computing power (thanks and welcome to the new users!). The 30-day sustained GIMPS throughput has swollen to a record 380 TFLOP/s, with the theoretical 30-day maximum throughput exploding to 1.25 PFLOP/s. This is the first time that the GIMPS distributed supercomputer has (theoretically) breached the petaflop barrier.

For historical perspective, in November 2005, GIMPS was knocking on the door of 20 TFLOP/s. We now produce nineteen times that: what constituted GIMPS' typical 24-hour throughput a decade ago, is now completed every 76 minutes.

It is worth noting that the actual 30-day throughput has been typically 35-40% of the theoretical maximum. Therefore, as the new users continue to complete assignments, I expect that we should sustain 450-500+ TFLOP/s in the next few weeks.

In case you haven't seen this... a "beta" set of graphs:
http://www.mersenne.org/primenet/graphs.php

[NBtarheel_33]

GIMPS has set yet another new record for 30-day sustained computing throughput: 405+ TFLOP/s, beating the old record of 400+ TFLOP/s set on 2016-02-19 in the wake of the discovery of the 49th Mersenne prime and the associated influx of new users and computing power.

The maximum theoretical 30-day sustained throughput record remains 1,201+ TFLOP/s, also set on 2016-02-19. The present value of this metric is 1,037+ TFLOP/s, which still places GIMPS in the impressive realm of petaflop-level computing power, while showing a 17.4% increase in computing efficiency from 33.3% to 39.1% overall, in just shy of one year.

[Madpoo]

Originally Posted by NBtarheel_33

GIMPS has set yet another new record for 30-day sustained computing throughput: 405+ TFLOP/s, beating the old record of 400+ TFLOP/s set on 2016-02-19 in the wake of the discovery of the 49th Mersenne prime and the associated influx of new users and computing power.

The maximum theoretical 30-day sustained throughput record remains 1,201+ TFLOP/s, also set on 2016-02-19. The present value of this metric is 1,037+ TFLOP/s, which still places GIMPS in the impressive realm of petaflop-level computing power, while showing a 17.4% increase in computing efficiency from 33.3% to 39.1% overall, in just shy of one year.

Feb 9th (the most recent spike on the graph) had the anomalous single result with 527 GHz-days of credit, so I wouldn't put too much stock in that one.

There has been discussion on whether/how much to cap GHz-day credit to prevent crazy things like that.

[Gordon]

Originally Posted by Madpoo

Feb 9th (the most recent spike on the graph) had the anomalous single result with 527 GHz-days of credit, so I wouldn't put too much stock in that one.

There has been discussion on whether/how much to cap GHz-day credit to prevent crazy things like that.

I see your 527 GHz days and raise you 4000.

ECM on very tiny exponents with very large B2 bounds...

[GP2]

Originally Posted by Gordon

I see your 527 GHz days and raise you 4000.

I see your 4000 GHz days and raise you 527511 (the actual amount of credit, MadPoo's figure was off by three orders of magnitude).

[science_man_88]

Originally Posted by GP2

I see your 4000 GHz days and raise you 527511 (the actual amount of credit, MadPoo's figure was off by three orders of magnitude).

in other words he used the wrong prefix and therefore should have used THz days instead.

[NBtarheel_33]

Originally Posted by NBtarheel_33

GIMPS has set yet another new record for 30-day sustained computing throughput: 405+ TFLOP/s, beating the old record of 400+ TFLOP/s set on 2016-02-19 in the wake of the discovery of the 49th Mersenne prime and the associated influx of new users and computing power.

The maximum theoretical 30-day sustained throughput record remains 1,201+ TFLOP/s, also set on 2016-02-19. The present value of this metric is 1,037+ TFLOP/s, which still places GIMPS in the impressive realm of petaflop-level computing power, while showing a 17.4% increase in computing efficiency from 33.3% to 39.1% overall, in just shy of one year.

New record: 421.679 TFLOP/s. Efficiency is now at 40% (1054.193 TFLOP/s maximum).

[NBtarheel_33]

This is getting as bad (as good?) as the US stock market.

New record: 431.005 TFLOP/s.

[NBtarheel_33]

Originally Posted by NBtarheel_33

This is getting as bad (as good?) as the US stock market.

New record: 431.005 TFLOP/s.

New record for 30-day sustained throughput: 436.648 TFLOP/s. The post-M50 crowd seems to be hanging around a little longer than those joining on after previous prime discoveries.