But none of the current gpu apps are using the full capabilities of the cards like our old special sauce app did.

I disagree. I believe the GPUGRID app uses the full extent of the GPU. i get even more power draw from their app than I did with the SETI app. with GPU use pegged to 98-100%. but they way their apps are designed, GPU memory use is quite low at less than 1GB, yet PCIe bus use is sky high compared to other projects (which can make it a bottleneck if you dont account for that when building your system)

it just depends what metric you want to use to define "using the full capabilities" I guess.

SETI - pulls up to max power limit, high heat produced, 99+% GPU core load, medium VRAM use (~2GB), marginal PCIe load (PCIe 3.0 x1 is fine)
GPUGRID - pulls up to max power limit, very high heat produced, 99+% core load, minimal VRAM use (<1GB), significant PCIe load (PCIe 3.0 x8 or better)
Einstein GR - pulls about 75% TDP, medium heat produced, 99+% core load, minimal VRAM use (<1GB), insignificant PCIe load (even down to a PCIe 1.0 x1)
Einstein GW - these are more CPU bound, the best performance will come from a strong CPU driving the system to help the GPU perform the tasks faster
but you're looking at about the same 75% TDP, medium heat produces, GPU load depending on CPU performance, high VRAM requirements at times
>3GB, insignificant PCIe load (even down to a PCIe 1.0 x1)

personally I think the GPUGRID app is on par with the SETI Special CUDA app based on GPU load and power draw, they're just different types of calculations with different requirements. a fully unpacked and decompressed GPUGRID WU probably has no chance to fit into a normal GPU's available memory, so they dont even try, leading to the huge PCIe bus load sending data back and forth constantly.

We'll have to see if there are any issues with next gen hardware on GPUGRID when they are released. I think I remember you saying that GPUGRID had this issue last time with Turing, but I also think they have changed their apps since then and it might no longer be an issue. we'll just have to see.

---

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

But first I would have to purchase one of the new cards. Wonder where this generation prices out compared to Turing. Higher - Lower - the same . . .? ? ?

If previous new families is any guide they will start a bit higher, just after the remaining stock of the old families has had a price cut to clear the shelves.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

It depends which end of the line they start from - and it looks as if they're starting at the top.

Note that the RTX 2070 Super, 2080 Super, and 2080 Ti were all overkill (price/performance wise) for distributed computing: they contained lots of extra silicon for new functions that we didn't need here.

Hi Keith,
I would much rather stay away from ANYTHING Google. I don't like their politics, or their spying, or their data stealing, etc., etc. I use DuckDuckGo for my search engine. I haven't used Google search in over a year. I won't use it, no matter what. I don't use any computer outside my house. Yeah, I know, my cell phone is Google, but I hate Apple even more. ;)
Have a great day! :)
Siran

. . Ahhh, a man after my own heart .... dislike Google, hate Apple ... :)

Stephen

LOL

+2 :)

---

CAPT Siran d'Vel'nahr - L L & P _\\//
Winders 11 OS? "What a piece of junk!" - L. Skywalker
"Logic is the cement of our civilization with which we ascend from chaos using reason as our guide." - T'Plana-hath

We'll have to see if there are any issues with next gen hardware on GPUGRID when they are released. I think I remember you saying that GPUGRID had this issue last time with Turing, but I also think they have changed their apps since then and it might no longer be an issue. we'll just have to see.

The main issue was with Einstein and the new Turing cards with the new SM75 gpuarch naming when compiling with nvcc.

The new 75 SM definition for Turing cards wasn't defined so all the tasks immediately failed. But the new acemd3 application which is BOINC wrapper based should be able to handle any GPUGrid task AS LONG as the Nvidia drivers are updated to understand the new Ampere silicon cards.

So the new arch definition for Ampere silicon should be handled without issues I believe.

---

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

A proud member of the OFA (Old Farts Association)

It depends which end of the line they start from - and it looks as if they're starting at the top.

Note that the RTX 2070 Super, 2080 Super, and 2080 Ti were all overkill (price/performance wise) for distributed computing: they contained lots of extra silicon for new functions that we didn't need here.

Even if we had no use for the RTX or Tensor functions in the high end Turing cards, they DID give us more SM's or CUDA cores which DID improve the compute capabilities from the sheer improvement in parallelization.

I think the same will apply for Ampere. Based on the rumored CUDA core count gains from Turing to Ampere. Supposed to be some IPC gains in there too.

https://www.techpowerup.com/gpu-specs/geforce-rtx-3080.c3621

https://www.techpowerup.com/gpu-specs/geforce-rtx-3080-ti.c3581

https://www.techpowerup.com/gpu-specs/geforce-rtx-3090.c3622

---

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

A proud member of the OFA (Old Farts Association)

Hi Keith,
What a difference RTX3090 to my old EVGA Titans from 2013.
http://=https://www.techpowerup.com/gpu-specs/geforce-gtx-titan.c1996
I have 3 of them in my old V8-Xeon machine and they are working till today without any issue. The machine is complete aircooled. On the Intel Skulltrail Board working two Xeons E5405 with 80W TDP with 100%CPU usage over all these years.
It is time to build a new machine:
in late autumn I plan a new Server with 2x Xeon E5-2698v4 20Core 40line and RTX3080Ti

greetings :-)[url][/url]

---

D5400XS V8-Xeon

in late autumn I plan a new Server with 2x Xeon E5-2698v4 20Core 40line and RTX3080Ti

Even though it has less cores & threads, a Ryzen 9 3900XT would out perform one of those E5-2698v4s and use less power- much higher IPC & base & boost clock speeds.
Two Ryzen 9 3900XT systems would still use less power (while outperforming) that older dual socket system.

---

Grant
Darwin NT

hmmmm,,,,,
can´t imagine that Ryzen 9 3900XT can beat a Xeon 20 Core 40 threads (we must always see the summary output of wu`s/ KWh)
New Brand 22 core Intel Xeon E5-2699 v4 server processor cpu
available from US$100.00 - US$500.00 / Piece | 1.0 Piece/Pieces (Min. Order)
(2 Pieces max 1000,00 $) 44Cores 96 Threads
TDP 145 W summary TDP 290 W .

Well, I´m Impressed of the AMD Ryzen Threadripper PRO 3995WX with 64Cores 128 Threads (TDP: 280 W) $4,170 on newegg

must rethink my Server Project, but all is depending on budget

---

D5400XS V8-Xeon

Depends on your projects. If sticking to gpu apps that only need FP32, a 3090 will outperform the Titan by 4X.

If you are running MW OTOH, with the need for FP64, the Titan will be 2.5X faster than a 3090.

---

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

A proud member of the OFA (Old Farts Association)

can´t imagine that Ryzen 9 3900XT can beat a Xeon 20 Core 40 threads

Depends on the project and what kind of cpu application is used.

Prime benefit is the 3900XT will run a lot faster on all-cores compared to a Xeon. By at least a Ghz. That will shorten the crunch time considerably.

To give you an idea of just the difference in cpu architecture can bring between the Ryzen Zen+ architecture and the Ryzen Zen 2 architecture, compare my two hosts at Universe.

The Zen+ cpu is a Threadripper 2920X and the Zen 2 cpu is a 3950X. Both hosts run the same 4.2Ghz clock speeds with the same memory timings basically.

Yet, the 3950X cruises through Universe tasks in 30-50 minutes while the 2920X takes 70-120 minutes. Mainly because the internal floating point registers are twice as wide as the Threadripper.

---

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

A proud member of the OFA (Old Farts Association)

I'm impressed about your 3950X, yes it has AVX and AVX2 that's the clue.
Some of the new Xeon Gold series have AVX and AVX2, some models have 2 units AVX2 like Xeon Gold 6210U (Purley)
http://=https://www.alternate.de/Intel/Xeon-Gold-6210U-Prozessor/html/product/1541985?
But the price is € 1.699,- Euro in Germany. and it has 150W TDP, 20core, 40 threads, clock speed 2500 MHz - max 3900 MHz, DDR4-2933
I believe it can not beat your 3950X price € 729,-Euro, 16core 32 threads 105 W TDP, clock speed 3500 MHz -max 4700 MHz, DDR4-3200
so far a 3950X is a very well alternative

I will rethink my server project
thanks

---

D5400XS V8-Xeon

Annandtech have just posted their recommendations for Best CPUs for Workstations for August 2020.
Intel only makes the list in 1 category, for the rest it's AMD all the way based on performance alone. Based on price or availability, Intel just aren't in the running.

---

Grant
Darwin NT

I'm impressed about your 3950X, yes it has AVX and AVX2 that's the clue.
Some of the new Xeon Gold series have AVX and AVX2, some models have 2 units AVX2 like Xeon Gold 6210U (Purley)
http://=https://www.alternate.de/Intel/Xeon-Gold-6210U-Prozessor/html/product/1541985?
But the price is € 1.699,- Euro in Germany. and it has 150W TDP, 20core, 40 threads, clock speed 2500 MHz - max 3900 MHz, DDR4-2933
I believe it can not beat your 3950X price € 729,-Euro, 16core 32 threads 105 W TDP, clock speed 3500 MHz -max 4700 MHz, DDR4-3200
so far a 3950X is a very well alternative

I will rethink my server project
thanks

And the actuality of the Zen 2 processors like the 3950X is that it can easily run much faster memory than the specification shows. You should not have any issues running at least 3600Mhz or even 3800Mhz on the memory with some simple adjustments in the BIOS.

And memory clocks on Ryzen are directly tied to the speed of the cpu cores and the internal communication between cores and I/O.

The faster your memory, the faster the cpu gets.

I run 3600Mhz with 4 X 8 GB sticks of 3200Mhz XMP memory at CL14 latency. If I was only running two sticks, I probably could run 3800Mhz or even higher.

So that in itself makes the Ryzen choice a no-brainer over the maximum memory clocks that an equivalent core count Xeon every could attempt.

---

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

A proud member of the OFA (Old Farts Association)

The only situation an Intel workstation makes sense is if you have to have AVX-512 capability for some very uncommon applications. And very few Intel processors are actually AVX-512 capable in the first place.

For every other workstation use, the AMD processors win over Intel for almost all programs.

---

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

A proud member of the OFA (Old Farts Association)

New fibre transmission data rate set.

Working together, the team was able to achieve a data transmission rate of 178 terabits a second, or 178,000,000 megabits a second. According to UCL’s press release, that’s fast enough to download Netflix’s entire catalog in less than a second.

And the good thing about it was the record was set using current optic fibre- just the hardware at each end allowed the higher data rate. So by replacing the transmitter & receiver & any inline amplifiers, existing links will be able to support a huge increase in bandwidth over their current capabilities when the hardware becomes commercially available.

---

Grant
Darwin NT

I'd give an arm and a leg just to have standard 1Gb fiber link to the premises instead of my crappy DSL. Even 100Mb would be acceptable.

---

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

A proud member of the OFA (Old Farts Association)

I'd give an arm and a leg just to have standard 1Gb fiber link to the premises instead of my crappy DSL. Even 100Mb would be acceptable.

Here 100Mb/s fibre to the premises is considered as good as it gets.
There are some 250MB/s plans around, but given the 100Mb/s plans are $100/month, i haven't even bothered to checkout how much the faster ones are. Apparently there are some 1Gb/s plans for businesses, at appropriate pricing levels of course. And that's only if you're lucky enough to have Fibre to the premises, a few have either Wifi or Fibre to the Node with Coax to the premises, but the majority are stuck with Fibre to the Node, and 50+ year old copper to the home.

My parents had ADSL2+ for years now, and often the best sync speed they could usually get was 2MB/s (very occasionally 4MB/s), but often it was 1MB/s or less with frequent & often extended dropouts (their ISP ended up giving them a modem with mobile phone support- when the phone went down it would switch over to the mobile network, and even though it was only 3G it was still faster than their ADSL 2+ connection). Late last year they finally got Fibre to the node, with Coax to the premises. They've got a 25Mb/s connection, and have to reboot the modem every few weeks to keep their phone & net connection working.
Which i consider to be absolute crap (i'm generally good for 95MB/s download & 25MB/s upload and the only time the modem reboots is if the power goes out long enough for the UPS battery to go flat) , but compared to what they had- it's paradise.

---

Grant
Darwin NT

For the first time in years i'm getting some money back from my Tax return, so i had a look at what i could get (if i put in a bit of (well, a lot of) extra cash).

AMD Ryzen 9 3900X with Wraith Prism                               $  819
Gigabyte X570 UD Motherboard                                      $  279
Samsung 970 EVO Plus NVMe SSD 1TB                                 $  299
Corsair Vengeance LPX 64GB (2x32GB) 3200MHz CL16 DDR4             $  479
Corsair RM650 Gold 650W Power Supply                              $  199
Thermaltake Core P5 Tempered Glass V2 Edition Wall-Mount Chassis  $  269
ASUS ROG Strix LC 240 AIO Liquid CPU Cooler                       $  209
Microsoft Windows 10 Pro 64bit OEM DVD                            $  229
ASUS DRW-24D5MT 24x DVD Writer                                    $   22
                                                      Sub-total:  $2,804

Plus another $200 or so for freight. And i might swap the CPU cooler with a Kraken X73 (extra $100 or so). Just saw a review and it's very impressive- while it's considered noisy (for a water cooler), but it has excellent cooling performance. And the higher the load, the better it performs. With the weather here, excellent cooling isn't just desirable- it's necessary.

The case i want (an Obsidian 750D airflow edition) is no longer in production, and the pricing of the Threadrippers is just way out of my reach (plus all the extra RAM i'd need to support all those extra cores & threads and memory channels (it would be a bit silly to get something like that, then hamstring it by not using all of it's memory channels)).

3960X   $2339
3970X   $3329
3990X   $6579

Noises are the pandemic has caused a big increase in RAM inventories, so prices are expected to drop 10-20% over the next few months. And the next generation of Ryzens aren't far off, and the lack of hype, along with the very promising numbers that have leaked, make them way more attractive than the current series of CPUs.
So i'll probably hold off till March/April of next year. Should be able to get a much more powerful system, for around the same price, with no more really big performance gains in future CPU releases for the next 12months or more from then.

---

Grant
Darwin NT

I'm waiting for the shoe to drop, i.e. the new Ryzen 4000 Zen 3 processors and if there is another 15% IPC improvement over Zen 2. Leaks say there is another big jump in RAM speeds too.

I don't know what projects you are interested in, but I and everyone else running the BHspin tasks at Universe is seeing a 2X reduction in crunching times just by moving from Ubuntu 18.04.5 to Ubuntu 20.04.1 even though the kernel is the same. You might consider moving from Mint Tina to the latest.

---

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

A proud member of the OFA (Old Farts Association)