167 terraflops/sec 2.place in world ranking list of civil use now in Research-Center Juelich Germany
regards ;-)
heinz ~seti_britta

---

D5400XS V8-Xeon

167 terraflops/sec 2.place in world ranking list of civil use now in Research-Center Juelich Germany
regards ;-)
heinz ~seti_britta

If used on S@H, how long before it hit the No 1 spot?

---

167 terraflops/sec 2.place in world ranking list of civil use now in Research-Center Juelich Germany
regards ;-)
heinz ~seti_britta

If used on S@H, how long before it hit the No 1 spot?

Well, let's see, the top host has 18,000,000 credits (roughly), and can do 1560*2 MFlops/sec. It seems to be producing much less lately, so let's look at an active host to get the RAC/FLOP conversion. Msattler's fastest box does 3761*4 MFlops, and has an RAC of 4941. So, the Machine that does 167 TFlops/Sec would have an RAC of about 167*1000000*4941/15044=54.8 million.

At 54.8 million credits per day, it would take approximately 8 hours for it to pass the #1 host in total credits.

---

167 terraflops/sec 2.place in world ranking list of civil use now in Research-Center Juelich Germany
regards ;-)
heinz ~seti_britta

If used on S@H, how long before it hit the No 1 spot?

Well, let's see, the top host has 18,000,000 credits (roughly), and can do 1560*2 MFlops/sec. It seems to be producing much less lately, so let's look at an active host to get the RAC/FLOP conversion. Msattler's fastest box does 3761*4 MFlops, and has an RAC of 4941. So, the Machine that does 167 TFlops/Sec would have an RAC of about 167*1000000*4941/15044=54.8 million.

At 54.8 million credits per day, it would take approximately 8 hours for it to pass the #1 host in total credits.

Um... probably somewhat less. Not all flop measures get the same answer for the same hardware. The Top500 list, I think, is calibrated by LinPack speed, and I don't think the flop count for that matches the BOINC benchmark flops especially well.

Case in point, here in New Mexico the local paper just announced that the state is buying a late-model SGI supercomputer for installation by next summer, at which time it is expected to be about #5 on the Top500 list, at a 172 Teraflop rating (very close to the machine of this thread on that measure).

That SGI machine as installed in NM will 3584 quad-core processors of the same type and generation as msattler's quads. They are the Xeon flavor, which is irrelevant for this point, and will certainly not be so viciously overclocked as Mr. Sattler does, so would produce rather less than 3000 times his output, not the well over 10000 times predicted by the method proposed here.

Almost all the machines on the Top500 list these days use large numbers of individual processors, each of which is actually slower than the processors at the top of our list (because older, and operated for short-term accuracy and long-term reliability). The ones that are best for large problems have highly capable interconnect between the processors, but those are losing market share to ones which just use Ethernet (and are much cheaper, but have inferior latency and pose a different programming problem). For the SETI task, which has utterly negligible interprocessor communication, pretty much nothing that makes supercomputers "super" matters, so it would be a big waste of money and power to run SETI on them.

The days when a Cray had vastly higher uniprocessor performance than consumer desktop machines are long gone, and not likely to return. The handwriting was on the wall by the time of the 80486, and the convergence has accelerated since then.

---

> idea * Dr. Eric Korpela - write a Proposal to them and request *use* for SETI@home / BOINC ?

---

BOINC Wiki . . .

Science Status Page . . .

The top500 list shall be released on November 13, that is tomorrow, at least in Italy.
Tullio

---

TOP500 List - November 2007

Greetings from Germany NRW
Ulli

TOP500 List - November 2007

Ah, they have updated it. I'm puzzled to see the New Mexico state machine at number 3--I thought it was not to finish being installed for some months, and I thought this list was supposed to be measured performance.

Anyway, if you look on the right side of the first page for the graph generator and click the "development" tab before setting up a query, you can choose "processor family" to see a year by year graph of what has been powering these beasts. Over the last very few years, the sum of AMD's x86_64 and Intel's EMT64 has risen from under 1% of the systems in November 2003 to what looks like about 80% of the systems in November 2007. The fraction of total flops is smaller, but still over half today (Power is the other consistent player, mostly because of IBM's dominant offerings in this arena).

Ethernet grows from zero systems in 2001 to over half of the current list.

---

Folding@home is claiming 1556 teraflops. No comment.
Tullio

---

Another way of counting the beans would be the way we count. Ie, steps necessary to perform quad-dimensional path-integrals:-)
Currently we are performing 43,614,208,000 steps per second... Ontop of that we have all the routing and checking operations necessary to run a a four-dimensional computer topology.
Still a bit left to blue gene, but we're getting there.

---

It is SEXY to DONATE!
Skype = demiurg2