Just some fun playing with a Tesla P100 plus a question...

[JonRussell]

Hi,

I had an NVidia Tesla P100 to play with this week, we were installing it in a big Hadoop cluster for a client. But I had it for the day to play with before we handed it over. I ran the mfaktc code for a few minutes.

I achieved 617 GHz d/days on the P100 (3584 cores, 4.7 TeraFLOPS DP)

Date    Time | class   Pct |   time     ETA | GHz-d/day    Sieve     Wait
Apr 26 14:57 | 4604  99.8% |  0.021    n.a. |    616.78    82485    n.a.%
Apr 26 14:57 | 4605  99.9% |  0.021    n.a. |    616.78    82485    n.a.%
Apr 26 14:57 | 4617 100.0% |  0.021    n.a. |    616.78    82485    n.a.%
found 2 factors for M3321932839 from 2^50 to 2^71 [mfaktc 0.21 75bit_mul32_gs]
tf(): total time spent: 20.802s

For comparison, I ran the same mfaktc code on my NVidia GeForce GTX 750 Ti graphics card and achieved 100 GHz d/days

Date    Time | class   Pct |   time     ETA | GHz-d/day    Sieve     Wait
Apr 26 14:50 | 4604  99.8% |  0.129    n.a. |    100.41    82485    n.a.%
Apr 26 14:50 | 4605  99.9% |  0.129    n.a. |    100.41    82485    n.a.%
Apr 26 14:50 | 4617 100.0% |  0.130    n.a. |     99.63    82485    n.a.%
found 2 factors for M3321932839 from 2^50 to 2^71 [mfaktc 0.21 75bit_mul32_gs]
tf(): time spent since restart:    1m 44.186s
      estimated total time spent:  2m  5.023s

I was surprised a £100 graphics card was only 1/6th slower that a £5,000 P100 ?

Which begs the question ... what is the difference between the Telsa card architecture and a normal NVidia graphics card, to warrant the huge price difference ?

The P100 has 3584 cores, 12GB RAM, 1328 MHz clock.
The GeForce GTX 1080 Ti is almost exactly the same spec and has 3584 cores, 11GB RAM, 1480 MHz clock.
They are both the same NVidia Pascal Architecture.

However, the P100 is about £5,000 ! and the GTX 1080 Ti is about £700 ?

Whats the difference ?

Regards,

Jon.

[VictordeHolland]

Tesla
pure compute card (no video output)
option to turn on ECC (you lose 1/8 of the available memory)
1/2 the DoublePrecision of their SinglePrecision performance
HBM2 memory give it 540-720GB/s of memory bandwidth (depending on whether it is a 12GB or 16GB model)

GTX1080 Ti
a consumer card (with video output)
no ECC
1/32 the DoublePrecision of their SinglePrecision performance
GDDR5X memory (484GB/s memory bandwidth)

ECC + DP performance + Memory Bandwidth is what makes Tesla cards desirable for compute. If a company is doing seismic imaging/simulations/waterflow calculations etc, with engineers costing 1000s of dollars each month, you might also give them the proper tools of a couple of $1000s.

Nvidia uses product differentiation to distinguish between different markets (and get higher profit margins):
Tesla = Server/compute
Quadro = Workstation
Geforce = Consumer/gaming

[LaurV]

As mentioned, it was not the best test you could do. Mfaktc is mostly integer math and little SP float, no DP float, neither huge memory access. You can find cards with even lower price, giving a better Mfaktc score. See here.

This card is strong at DP float (1/2 SP float performance, compared with 1/32, or 1/8, or 1/3 for top cards from the former Tesla/Fermi series, etc). You should have tried cudaLucas or other DP-heavy stuff. In this table, P100 may stay somewhere about 150 or so (unfortunately nobody reported it yet, but Oliver and others around did some tests and published the results).

[kladner]

Something does not seem right here. I am running a GTX 1060. With the warm weather we are having, the settings I have run it at ~2000MHz. It runs a little higher when ambient is lower. It is currently turning out 615 GHz-d/d**. It has 1280 cores.
**running mfaktc

[JonRussell]

Shall I run the cudaLucas code and post the results somewhere to add to the database ?

Regards,

Jon.

[Mark Rose]

Originally Posted by kladner

Something does not seem right here. I am running a GTX 1060. With the warm weather we are having, the settings I have run it at ~2000MHz. It runs a little higher when ambient is lower. It is currently turning out 615 GHz-d/d**. It has 1280 cores.
**running mfaktc

That's what I get with mine, which has a 1.61 GHz base clock. Its fan is running at 91% and the temperatures are in the low 80's, so I'm thinking of returning it and getting a model with more than one fan.

[kladner]

Originally Posted by Mark Rose

That's what I get with mine, which has a 1.61 GHz base clock. Its fan is running at 91% and the temperatures are in the low 80's, so I'm thinking of returning it and getting a model with more than one fan.

I love Gigabyte coolers. Across several brands of cards, they are the quietest, with cooling as good or better than the others.
The 1060 is running at 78 C, ATM. The twin fans are at 94%, 2900 RPM.

[Mark Rose]

I have the EVGA GeForce GTX 1060 SC GAMING, ACX 2.0 (Single Fan) I have is easily the loudest card I've ever owned. It makes the GTX 580 silent by comparison.

The two fan EVGA GeForce GTX 1070 SC GAMING ACX 3.0 produces little noise and runs cool.

Both cards are drawing almost the same amount of electricity, but the 1070 is producing a lot more TF.

[Xyzzy]

FWIW, we run our 1060 at 50% TDP. It is silent and barely breaks 60° C. We get ~430 GHz-d/day.

[LaurV]

Originally Posted by JonRussell

Shall I run the cudaLucas code and post the results somewhere to add to the database ?

Yes please, follow the steps described here in the title. You can PM James on this forum, for details.

Edit: Optionally, you can post on this forum too, to make us envy you...