Cores in Tesla c2050 card shows 112 cores instead of 448

[sanf]

Dear Users,

    The Tesla C2050 GPU has 448 cores as mentioned by NVIDIA. But here in the following output, its showing 112 cores:

#lib/gpu/nvc_get_devices

Device 0: “Tesla C2050”
Revision number: 2.0
Total amount of global memory: 2.62 GB
Number of multiprocessors: 14
Number of cores: 112
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 32768
Warp size: 32
Maximum number of threads per block: 1024
Maximum sizes of each dimension of a block: 1024 x 1024 x 64
Maximum sizes of each dimension of a grid: 65535 x 65535 x 1
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Clock rate: 1.15 GHz
Concurrent copy and execution: Yes

Is there any difference between CUDA core and a normal core (as highlighted above)?

Also, there are no sample example programs inside CUDA Toolkit. Does anybody have such programs, which can compare both GPU & CPU performance?

Thanks

[avidday]

That is just a problem with the deviceQuery example. The previous generations of cards had 8 cores per multiprocessor, whereas the GF100 cards have 32 cores per multiprocessor. Because CUDA API doesn’t (yet) report the core count, only the multiprocessor count, that code has a fixed constant of 8 cores per multiprocessor. So it reports 14 * 8 = 112 instead of 14 * 32 = 448 as it should. There is nothing wrong with your card and you have not misunderstood the specifications.

[avidday]

That is just a problem with the deviceQuery example. The previous generations of cards had 8 cores per multiprocessor, whereas the GF100 cards have 32 cores per multiprocessor. Because CUDA API doesn’t (yet) report the core count, only the multiprocessor count, that code has a fixed constant of 8 cores per multiprocessor. So it reports 14 * 8 = 112 instead of 14 * 32 = 448 as it should. There is nothing wrong with your card and you have not misunderstood the specifications.

[sanf]

That is just a problem with the deviceQuery example. The previous generations of cards had 8 cores per multiprocessor, whereas the GF100 cards have 32 cores per multiprocessor. Because CUDA API doesn’t (yet) report the core count, only the multiprocessor count, that code has a fixed constant of 8 cores per multiprocessor. So it reports 14 * 8 = 112 instead of 14 * 32 = 448 as it should. There is nothing wrong with your card and you have not misunderstood the specifications.

Thanks.

We’re facing one more problem with a CUDA ported application LAMMPS. It has been installed successfully, as per the instructions given in the pdf :

http://lammps.sandia.gov/workshops/Feb10/M…own/gpu_tut.pdf

But it failed with following error:

cat gpu_test_lj_out

lmp_openmpi: pair_gpu_cell.cu:486: void build_cell_list(double*, int*, cell_list&, int, int, int, int, int, int, int, int): Assertion `err == cudaSuccess’ failed.

LAMMPS (7 Apr 2010)

Lattice spacing in x,y,z = 1.6796 1.6796 1.6796

Created orthogonal box = (0 0 0) to (33.5919 33.5919 33.5919)

1 by 1 by 1 processor grid

Created 32000 atoms

---

• Using GPGPU acceleration for LJ-Cut:

---

GPU 0: Tesla C2050, 112 cores, 2.6 GB, 1.1 GHZ

---

Setting up run …

[nvidia:13239] *** Process received signal ***

[nvidia:13239] Signal: Aborted (6)

[nvidia:13239] Signal code: (-6)

[nvidia:13239] [ 0] /lib64/libpthread.so.0 [0x359be0e4c0]

[nvidia:13239] [ 1] /lib64/libc.so.6(gsignal+0x35) [0x359b230215]

[nvidia:13239] [ 2] /lib64/libc.so.6(abort+0x110) [0x359b231cc0]

[nvidia:13239] [ 3] /lib64/libc.so.6(__assert_fail+0xf6) [0x359b229696]

[nvidia:13239] [ 4] …/…/src/lmp_openmpi [0x7275cc]

[nvidia:13239] [ 5] …/…/src/lmp_openmpi(_Z12_lj_gpu_cellIffEdR13LJ_GPU_MemoryIT_T0_EPPdS

5_S6_PiiiibbPKdS9_+0x127) [0x7205b7]

[nvidia:13239] [ 6] …/…/src/lmp_openmpi(Z11lj_gpu_cellPPdS_S0_PiiiibbPKdS3+0x4f) [0x71bd6f]

[nvidia:13239] [ 7] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS12PairLJCutGPU7computeEii+0xb1) [0x6a0861]

[nvidia:13239] [ 8] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS6Verlet5setupEv+0x12a) [0x71771a]

[nvidia:13239] [ 9] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS3Run7commandEiPPc+0x449) [0x6f5919]

[nvidia:13239] [10] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS5Input15execute_commandEv+0xe0e) [0x5fd45e]

[nvidia:13239] [11] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS5Input4fileEv+0x34d) [0x5fe60d]

[nvidia:13239] [12] …/…/src/lmp_openmpi(main+0x4a) [0x604c7a]

[nvidia:13239] [13] /lib64/libc.so.6(__libc_start_main+0xf4) [0x359b21d974]

[nvidia:13239] [14] …/…/src/lmp_openmpi(__gxx_personality_v0+0x421) [0x4600a9]

[nvidia:13239] *** End of error message ***

The input file is:

cat in.lj

3d Lennard-Jones melt

newton off

variable x index 1

variable y index 1

variable z index 1

variable xx equal 20*$x

variable yy equal 20*$y

variable zz equal 20*$z

units lj

atom_style atomic

lattice fcc 0.8442

region box block 0 ${xx} 0 ${yy} 0 ${zz}

create_box 1 box

create_atoms 1 box

mass 1 1.0

velocity all create 1.44 87287 loop geom

pair_style lj/cut/gpu one/node 0 2.5

pair_coeff 1 1 1.0 1.0 2.5

neighbor 0.3 bin

neigh_modify delay 0 every 20 check no

fix 1 all nve

run 100

What’s the wrong here?

The OS is RHEL-5.3 64 bit. The cuda toolkit says RHEL-5.4 64bit (cudatoolkit_3.1_linux_64_rhel5.4.run). Is this compatibility issue causing the failure?

Is there anybody successful in running this application on GPU?

Thanks once again

[sanf]

That is just a problem with the deviceQuery example. The previous generations of cards had 8 cores per multiprocessor, whereas the GF100 cards have 32 cores per multiprocessor. Because CUDA API doesn’t (yet) report the core count, only the multiprocessor count, that code has a fixed constant of 8 cores per multiprocessor. So it reports 14 * 8 = 112 instead of 14 * 32 = 448 as it should. There is nothing wrong with your card and you have not misunderstood the specifications.

Thanks.

We’re facing one more problem with a CUDA ported application LAMMPS. It has been installed successfully, as per the instructions given in the pdf :

http://lammps.sandia.gov/workshops/Feb10/M…own/gpu_tut.pdf

But it failed with following error:

cat gpu_test_lj_out

lmp_openmpi: pair_gpu_cell.cu:486: void build_cell_list(double*, int*, cell_list&, int, int, int, int, int, int, int, int): Assertion `err == cudaSuccess’ failed.

LAMMPS (7 Apr 2010)

Lattice spacing in x,y,z = 1.6796 1.6796 1.6796

Created orthogonal box = (0 0 0) to (33.5919 33.5919 33.5919)

1 by 1 by 1 processor grid

Created 32000 atoms

---

• Using GPGPU acceleration for LJ-Cut:

---

GPU 0: Tesla C2050, 112 cores, 2.6 GB, 1.1 GHZ

---

Setting up run …

[nvidia:13239] *** Process received signal ***

[nvidia:13239] Signal: Aborted (6)

[nvidia:13239] Signal code: (-6)

[nvidia:13239] [ 0] /lib64/libpthread.so.0 [0x359be0e4c0]

[nvidia:13239] [ 1] /lib64/libc.so.6(gsignal+0x35) [0x359b230215]

[nvidia:13239] [ 2] /lib64/libc.so.6(abort+0x110) [0x359b231cc0]

[nvidia:13239] [ 3] /lib64/libc.so.6(__assert_fail+0xf6) [0x359b229696]

[nvidia:13239] [ 4] …/…/src/lmp_openmpi [0x7275cc]

[nvidia:13239] [ 5] …/…/src/lmp_openmpi(_Z12_lj_gpu_cellIffEdR13LJ_GPU_MemoryIT_T0_EPPdS

5_S6_PiiiibbPKdS9_+0x127) [0x7205b7]

[nvidia:13239] [ 6] …/…/src/lmp_openmpi(Z11lj_gpu_cellPPdS_S0_PiiiibbPKdS3+0x4f) [0x71bd6f]

[nvidia:13239] [ 7] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS12PairLJCutGPU7computeEii+0xb1) [0x6a0861]

[nvidia:13239] [ 8] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS6Verlet5setupEv+0x12a) [0x71771a]

[nvidia:13239] [ 9] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS3Run7commandEiPPc+0x449) [0x6f5919]

[nvidia:13239] [10] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS5Input15execute_commandEv+0xe0e) [0x5fd45e]

[nvidia:13239] [11] …/…/src/lmp_openmpi(_ZN9LAMMPS_NS5Input4fileEv+0x34d) [0x5fe60d]

[nvidia:13239] [12] …/…/src/lmp_openmpi(main+0x4a) [0x604c7a]

[nvidia:13239] [13] /lib64/libc.so.6(__libc_start_main+0xf4) [0x359b21d974]

[nvidia:13239] [14] …/…/src/lmp_openmpi(__gxx_personality_v0+0x421) [0x4600a9]

[nvidia:13239] *** End of error message ***

The input file is:

cat in.lj

3d Lennard-Jones melt

newton off

variable x index 1

variable y index 1

variable z index 1

variable xx equal 20*$x

variable yy equal 20*$y

variable zz equal 20*$z

units lj

atom_style atomic

lattice fcc 0.8442

region box block 0 ${xx} 0 ${yy} 0 ${zz}

create_box 1 box

create_atoms 1 box

mass 1 1.0

velocity all create 1.44 87287 loop geom

pair_style lj/cut/gpu one/node 0 2.5

pair_coeff 1 1 1.0 1.0 2.5

neighbor 0.3 bin

neigh_modify delay 0 every 20 check no

fix 1 all nve

run 100

What’s the wrong here?

The OS is RHEL-5.3 64 bit. The cuda toolkit says RHEL-5.4 64bit (cudatoolkit_3.1_linux_64_rhel5.4.run). Is this compatibility issue causing the failure?

Is there anybody successful in running this application on GPU?

Thanks once again

[dgwsoft]

Dear Users,

The Tesla C2050 GPU has 448 cores as mentioned by NVIDIA. But here in the following output, its showing 112 cores:

#lib/gpu/nvc_get_devices

Device 0: “Tesla C2050”

Revision number: 2.0

Total amount of global memory: 2.62 GB

Number of multiprocessors: 14

Number of cores: 112

Total amount of constant memory: 65536 bytes

Total amount of shared memory per block: 49152 bytes

Total number of registers available per block: 32768

Warp size: 32

Maximum number of threads per block: 1024

Maximum sizes of each dimension of a block: 1024 x 1024 x 64

Maximum sizes of each dimension of a grid: 65535 x 65535 x 1

Maximum memory pitch: 2147483647 bytes

Texture alignment: 512 bytes

Clock rate: 1.15 GHz

Concurrent copy and execution: Yes

Is there any difference between CUDA core and a normal core (as highlighted above)?

Also, there are no sample example programs inside CUDA Toolkit. Does anybody have such programs, which can compare both GPU & CPU performance?

Thanks

I had the same issue with a couple of GTX 465 cards (reporting 88 instead of 352 cores). When I upgraded from the 2.3 Toolkit and SDK to the 3.1 Toolkit and SDK and reran deviceQuery, it got it right.

I guess the latest toolkit knows how many cores per multiprocessor for this device.

[dgwsoft]

Dear Users,

The Tesla C2050 GPU has 448 cores as mentioned by NVIDIA. But here in the following output, its showing 112 cores:

#lib/gpu/nvc_get_devices

Device 0: “Tesla C2050”

Revision number: 2.0

Total amount of global memory: 2.62 GB

Number of multiprocessors: 14

Number of cores: 112

Total amount of constant memory: 65536 bytes

Total amount of shared memory per block: 49152 bytes

Total number of registers available per block: 32768

Warp size: 32

Maximum number of threads per block: 1024

Maximum sizes of each dimension of a block: 1024 x 1024 x 64

Maximum sizes of each dimension of a grid: 65535 x 65535 x 1

Maximum memory pitch: 2147483647 bytes

Texture alignment: 512 bytes

Clock rate: 1.15 GHz

Concurrent copy and execution: Yes

Is there any difference between CUDA core and a normal core (as highlighted above)?

Also, there are no sample example programs inside CUDA Toolkit. Does anybody have such programs, which can compare both GPU & CPU performance?

Thanks

I had the same issue with a couple of GTX 465 cards (reporting 88 instead of 352 cores). When I upgraded from the 2.3 Toolkit and SDK to the 3.1 Toolkit and SDK and reran deviceQuery, it got it right.

I guess the latest toolkit knows how many cores per multiprocessor for this device.