The first glimmer of a possible Ampere consumer card?

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

The first glimmer of a possible Ampere consumer card?

Well, as they say- with that memory it would be a Quadro, but they use the Quadro silicon with less RAM & lower Double Precision performance for the x080 /x080Ti & Titan type cards.
So it gives us an idea of the type of performance we can look forward to when the consumer Ampere cards come out, and it looks like the performance potential over the current hardware is going to be not too far short (if any) of incredible.

Notice how in the past (pretty much for AMD, Intel and Nvidia) the more noise and hype before a new release the more underwhelming it is when it comes out? And the less you hear about it before hand the better it is? And Nvidia have kept this as quiet as anything in recent history. And now with AMDs RNDAx series finally putting some pressure on Nvidia, there's a chance Nvidia won't up the price on this lot, in order to put the boot in to AMD.
So, fingers crossed.

---

Grant
Darwin NT

Yes, as mentioned in the article, too much memory for consumer cards. But the underlying silicon could have a similar variation for the consumer cards.

I think likely as per history that the performance scaling between generations has been typically 15% from generation to generation. I see no reason to doubt they will achieve same. Could be much more owing the fact they are changing process nodes from 12nm to 7nm. I wonder how much we will actually get in compute improvements. Rumor is that the RT cores are getting the benefit of the die shrink which we have no use for.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I think likely as per history that the performance scaling between generations has been typically 15% from generation to generation.

If the numbers in the database are accurate, these cards with only slightly over a 1GHz clock are almost 40% faster than current equivalent hardware.

---

Grant
Darwin NT

I think likely as per history that the performance scaling between generations has been typically 15% from generation to generation.

If the numbers in the database are accurate, these cards with only slightly over a 1GHz clock are almost 40% faster than current equivalent hardware.

and with 73% more cores than the 2080ti (what they are comparing it to).

---

Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

Just wondering: imagine the nightmare to keep them well feeded here.

They could crunch a WU in less than 20 secs. Or about 100 WU each 5 min on a 4 GPU host. LOL

---

I think likely as per history that the performance scaling between generations has been typically 15% from generation to generation.

If the numbers in the database are accurate, these cards with only slightly over a 1GHz clock are almost 40% faster than current equivalent hardware.

and with 73% more cores than the 2080ti (what they are comparing it to).

So with the higher clock speeds of production cards, architectural improvements, and mostly of all the 7nm process benefits, double the performance (or slightly more) of the current hardware isn't an unrealistic expectation.
Wow.

---

Grant
Darwin NT

That does sound like an amazing advance! I hope the price is not too scary.

---

That does sound like an amazing advance! I hope the price is not too scary.

Depends, did you scary with the 3K tag of the Titan V? Expect a "panic movie" price tag.

---

Keep in mind only two of the OpenCL Geekbench 5 tests test math performance, the Particle Physics and SFFT tests. The rest of the tests test the graphics performance and produce most of the score weighting.

It would be more interesting to see the CUDA performance tests which are not the default for the test and have to be manually set in the parameter invocation.

The older Geekbench 4 tests would have been more indicative of the compute performance.

Of course, we can only hope the compute performance is 40% better than Turing.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Just wondering: imagine the nightmare to keep them well feeded here.

They could crunch a WU in less than 20 secs. Or about 100 WU each 5 min on a 4 GPU host. LOL

Probably true, these new cards have 118 or 108 SM's or CU's. Closest comparison would be to Petri's Titan V with 80 SM's.

And yes he can do non-overflow tasks in 20 seconds if they are easy ones with high angle ranges.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Thanks Rob and Jim for your response to my question. I will continue to crunch.

I have a query in regards to the amount of work awaiting validation this is over 13 million but there is only around 6.1 million results in circulation. I am thinking this means there are lots of tasks waiting to expire and then be re-sent. Would I be somewhere close or am I completely wrong?

---

Thanks Rob and Jim for your response to my question. I will continue to crunch.

I have a query in regards to the amount of work awaiting validation this is over 13 million but there is only around 6.1 million results in circulation. I am thinking this means there are lots of tasks waiting to expire and then be re-sent. Would I be somewhere close or am I completely wrong?

Partly right.
The main issue is that they have to resend every WU that is a noise bomb, or finishes very, very early due to the problems with the RX 5000 series drivers trashing WUs, in order to keep erroneous results from being put in to the Science database.

---

Grant
Darwin NT

Thanks Grant for your response. One further question. Are they just allowing these tasks to timeout before resending them to what I assume would be not a card with the RX 5000 series driver?

---

Are they just allowing these tasks to timeout before resending them to what I assume would be not a card with the RX 5000 series driver?

There is no other mechanism for WUs that are sent out & aren't returned within a few days. The deadline is the mechanism for re-issuing such WUs.

Normally a WU is only sent out twice- that's what happens when it is first issued (_0 & _1 WUs) if both are returned & the WU Validated then that's it. If it's not Validated, then it's sent out again (_2), if it comes back & still not validated it's sent out again (_3) and so on till up to the 10th.
If everyone returned work in a day or 2 then even if it was sent out 9 more times than the initial release, it would be cleared in less than 3 weeks. But if it's not returned, it has to reach the deadline date before it will be re-issued. So it could take more than 6 months for some of these to clear (the vast majority are cleared within 2 months, but there those that take over 4 months).

---

Grant
Darwin NT

Looks like there are 11 ThreadRipers in the top 100.

7 Ryzen 5/7's

5 Ryzen 9's

4 Intel i9's

36 i7's

---

A proud member of the OFA (Old Farts Association).

Thanks Grant for filling me in. It will be interesting to see how long it takes for the number to drop and stay under the 13 million mark. I can't even remember what it used to be

---

Thanks Grant for filling me in. It will be interesting to see how long it takes for the number to drop and stay under the 13 million mark. I can't even remember what it used to be

Around about the 4 million mark, up to 5/5.5 million occasionally.

---

Grant
Darwin NT

Thanks Grant for filling me in. It will be interesting to see how long it takes for the number to drop and stay under the 13 million mark. I can't even remember what it used to be

https://web.archive.org/web/20191115164019/https://setiathome.berkeley.edu/show_server_status.php
thank Keith for this one.

Thank you for the graph and your answer Grant. I still struggle to believe how low the results waiting to be purged are I haven't seen them over 500,000 in probably almost a month unless it's been after an outage

---