GPU Owl 3080ti Benchmarks - Page 3

Magellan3s · 2022-04-26, 14:25

Quote:

Originally Posted by henryzz

According to https://www.anandtech.com/show/15978...idias-rtx-3090 GDDR6X would actually be better named GQDR6X. Maybe the quad data rate is the extra factor of 2.

GreenWithEnvy shows Mem Clock Max as 10551 and GPU Clock Max as 2100!

https://i.ibb.co/zm8j9HG/Screenshot-...6-09-22-58.png

kriesel · 2022-04-26, 14:58

"Memory Transfer Rate, what Nvidia Settings reports and changes, is different from the effective Memory Clock, what is actually being displayed by GWE" https://gitlab.com/leinardi/gwe/-/bl...ease/README.md (emphasis on effective etc mine). Effective memory clock in this context means the clock rate that would occur, if memory data bits were being sent one at a time per physical signal trace, unencoded, without PAM4 encoding etc., which they clearly per GDDR6X specifications, are not.

Magellan3s · 2022-04-26, 22:57

Quote:

Originally Posted by kriesel

"Memory Transfer Rate, what Nvidia Settings reports and changes, is different from the effective Memory Clock, what is actually being displayed by GWE" https://gitlab.com/leinardi/gwe/-/bl...ease/README.md (emphasis on effective etc mine). Effective memory clock in this context means the clock rate that would occur, if memory data bits were being sent one at a time per physical signal trace, unencoded, without PAM4 encoding etc., which they clearly per GDDR6X specifications, are not.

I don't know what to tell you, that's just what the program is telling me! That's why I posted a screenshot!

https://i.ibb.co/zm8j9HG/Screenshot-...6-09-22-58.png

kriesel · 2022-04-27, 00:48

"memory is running at 1188 MHz (19 Gbps effective)." https://www.techpowerup.com/gpu-spec...ti-20-gb.c3831
1188 MHz there is a clock frequency. Hz is a unit of frequency. Gbps is gigabits per second, a unit of data rate. Two totally different things, different names, different units, different definitions.
Note also the word "effective", which can be used as marketing speak for "not really, but we're going to claim this very high performance spec anyway." (I can type at xxx,xxx bits per hour. But that's 8 bits per keypress. The keypress frequency is 1/8 the bit rate.)

An apple is fruit. An orange slice is fruit. An orange slice is not an apple. No matter how much someone would wish it to be the same, or believe it to be, or how many others they may get to agree or imply that an orange slice is an apple.

Tell me you're learning the difference, or trying?

VBCurtis · 2022-04-27, 05:52

Ken, you're the one who should be learning in this thread.

The least ambiguous number to cite for speed is the one used and displayed by the drivers. You can complain all you want about how it's not a truthful speed, but those complaints should be aimed at the manufacturer, not at another forumite who is simply adopting the language used by the maker of his card. Why turn to a borderline insult like "tell me you're learning the difference, or trying?" when the OP doesn't care about your personal views?

chris2be8 · 2022-04-27, 15:27

Quote:

Originally Posted by kriesel

"memory is running at 1188 MHz (19 Gbps effective)."

That implies each memory cycle transfers 16 bits. Which is a plausible value. Although not directly useful when running programs on it.

Magellan3s · 2022-04-27, 18:45

Quote:

Originally Posted by VBCurtis

Ken, you're the one who should be learning in this thread.

The least ambiguous number to cite for speed is the one used and displayed by the drivers. You can complain all you want about how it's not a truthful speed, but those complaints should be aimed at the manufacturer, not at another forumite who is simply adopting the language used by the maker of his card. Why turn to a borderline insult like "tell me you're learning the difference, or trying?" when the OP doesn't care about your personal views?

Eh, Curtis... I guess I should have wrote "effective memory clock" instead of "memory clock"

kriesel · 2022-04-30, 18:33

It gets even worse.
NVIDIA's own utility for querying the card does not return the "effective" number given in their spec sheets, OR the clock frequency given in common third party utilities.
As in https://www.seimaxim.com/kb/gpu/nvidia-smi-cheat-sheet
(Run as admin in Windows, root on Linux, suggested)

Code:

nvidia-smi -q -d SUPPORTED_CLOCKS

returns, for GTX1080 and RTX2080 respectively:

Code:

==============NVSMI LOG==============

Timestamp                                 : Tue Apr 26 19:18:03 2022
Driver Version                            : 456.71
CUDA Version                              : 11.1

Attached GPUs                             : 2
GPU 00000000:01:00.0
    Clocks
        Graphics                          : 1809 MHz
        SM                                : 1809 MHz
        Memory                            : 4513 MHz
        Video                             : 1620 MHz
    Applications Clocks
        Graphics                          : N/A
        Memory                            : N/A
    Default Applications Clocks
        Graphics                          : N/A
        Memory                            : N/A
    Max Clocks
        Graphics                          : 1961 MHz
        SM                                : 1961 MHz
        Memory                            : 5005 MHz
        Video                             : 1708 MHz
    Max Customer Boost Clocks
        Graphics                          : N/A
    SM Clock Samples
        Duration                          : 11.27 sec
        Number of Samples                 : 100
        Max                               : 1822 MHz
        Min                               : 1771 MHz
        Avg                               : 1808 MHz
    Memory Clock Samples
        Duration                          : 11.27 sec
        Number of Samples                 : 100
        Max                               : 4513 MHz
        Min                               : 4513 MHz
        Avg                               : 4513 MHz

...

GPU 00000000:03:00.0
    Clocks
        Graphics                          : 1680 MHz
        SM                                : 1680 MHz
        Memory                            : 6800 MHz
        Video                             : 1560 MHz
    Applications Clocks
        Graphics                          : N/A
        Memory                            : N/A
    Default Applications Clocks
        Graphics                          : N/A
        Memory                            : N/A
    Max Clocks
        Graphics                          : 2100 MHz
        SM                                : 2100 MHz
        Memory                            : 7000 MHz
        Video                             : 1950 MHz
    Max Customer Boost Clocks
        Graphics                          : N/A
    SM Clock Samples
        Duration                          : Not Found
        Number of Samples                 : Not Found
        Max                               : Not Found
        Min                               : Not Found
        Avg                               : Not Found
    Memory Clock Samples
        Duration                          : Not Found
        Number of Samples                 : Not Found
        Max                               : Not Found
        Min                               : Not Found
        Avg                               : Not Found

Drilling down in https://www.nvidia.com/en-sg/geforce/10-series/ gives for GTX1080
"Memory Specs: 10 Gbps Memory Speed
8 GB GDDR5X Standard Memory Config
256-bit Memory Interface Width
320 Memory Bandwidth (GB/sec)"
I suppose the vague term "memory speed" could mean peak effective bit rate per signal line.
Although elsewhere historically it is used for DIMM data rate: https://www.crucial.com/support/memo...-compatability

And in https://www.nvidia.com/en-me/geforce...ards/rtx-2080/ NVIDIA gives
"Memory Specs:
14 Gbps 14 Gbps Memory Speed
8 GB GDDR6 8 GB GDDR6 Standard Memory Config
256-bit 256-bit Memory Interface Width
448 GB/s 448 GB/s Memory Bandwidth (GB/sec)"

I think an issue with the GWE utility could be graphical layout. Putting a clear description of what was being displayed would use more characters than the app appears to have allowed for. Or its author may not have spent the time to burrow into the terminology carefully. It is time consuming.

It would not have surprised me if there was a design something like:
Software settable memory clock generator A -> distribution at frequency A to the various memory chips -> on-chip doubler (to b) -> doubler (to c) -> RAM clocking->PAM4 output at 2-bit voltage levels at rate c (computed bit rate d = 2c = 4b = 8A). Or maybe A -> distribution -> on-chip 4x PLL to c -> PAM4 output with effective bit rate d. Or it's QDR, clock A & 90 degree phase shifted A', transfers on both rising and falling edges (edge count c=4 A, PAM4 bit rate d= 2 c = 8 A.

GPU-Z, MSI Afterburner etc report A; NVIDIA-SMI reports c; various sources (spec sheets, GWE) report d.

I had looked for and not found a reasonably recent reference for GPU circuit designs. References I had found are higher level, architectural, and old enough to precede recent GPU memory types.

This one is too vague too.

Drilling down further, I do find PLLs, in the memory package documentation:
The GDDR6 Micron 16gigibit MT61K512M32 spec sheet: Table 2 at bottom of page 7 uses units of GHz on clock signals CK_t, CK_c, WCK_t, WCK_c; 1.5 or 3 GHz respectively; 3, 6 or 12 Gbit/s/"pin" for data signals, on the fine-pitch ball grid array package. Fig 4 on page 8 shows frequency and data rate ratios relative to the clock fundamental, and clock/data alignment in time. Fig 5 shows example clock and interface circuitry.
That is perhaps relevant to my RTX2080 and similar generation GPUs. The internal clock indicated as used in the memory chip ("Internal WCK") is 3. GHz. f, 2f & 4f clock signals; 2f, 4f, 8f data etc rates.

The GDDR6X Micron 8gigibit MT61K256M32 spec sheet Table 1 page 5, Figure 3 page 6, and Figure 4 page 7 provide analogous data relevant to RTX30xx GPUs (PAM4 mode); 2.5 or 5 GHz clock signals, ~20Gbit/sec PAM4 (voltage-level 2-bit encoded) effective data rate = 10 GBaud/pin. The internal clock indicated as used in the memory chip ("Internal WCK") is 2.5 GHz.
Also https://media-www.micron.com/-/media...fe1f9a5ff0231e

So in summary, it looks like at every stage other than the memory chip manufacturer, there's loose handling of clock and data signal rate terminology. (I haven't checked, but I'd expect other GDDR6 chip manufacturers to be similarly clear about their products. A quick online search did not yield any other GDDR6X manufacturers. NVIDIA and every stage downstream from there could do better in using precise terminology to limit confusion, in my opinion.)

LaurV · 2022-05-05, 15:51

Something may still be fishy with that A100 system, crunching 10xx us/iter for a 1xxM exponent.
Does the system do other things in the same time?
I am getting 12xx us/iter for a 332M exponent currently crunching, also linux.
For the same exponent (332M) the V100 gets 18xx us/iter.
Just saying...
Not a linux guy.

tdulcet · 2022-05-06, 12:13

Quote:

Originally Posted by LaurV

Something may still be fishy with that A100 system, crunching 10xx us/iter for a 1xxM exponent.
Does the system do other things in the same time?
I am getting 12xx us/iter for a 332M exponent currently crunching, also linux.
For the same exponent (332M) the V100 gets 18xx us/iter.

Thanks to @Magellan3s generously providing me access to a V100 GPU last week, I was able to confirm this severe performance regression, as I got 1235 us/iter on a wavefront first time exponent with the GpuOwl master branch (v7.2-93-ga5402c5). For comparison, I got 641 us/iter with the GpuOwl v6 branch (v6.11-382-g98ff9c7). After doing more testing, I found that it was mostly caused by this commit, which @preda then reverted last weekend (see here).

This can be seen by looking at the timings on the V100 GPU for the last 14 commits for this same wavefront exponent (values are in us/iter):

Code:

v7.2-93-ga5402c5,1236
v7.2-92-g5fb55ca,1235
v7.2-91-g9c22195,1235
v7.2-90-g75d4a1d,1235
v7.2-89-g885b8af,1236
v7.2-88-g599b2b2,1235
v7.2-87-gc1d9e26,1235
v7.2-86-gddf3314,1235
v7.2-85-g6122a0e,1236
v7.2-84-gce4fe12,1236
v7.2-83-g1aff945,1235
v7.2-82-g7bea16f,653
v7.2-81-g5f17913,656
v7.2-80-g9a975f9,656
v7.2-79-g3b4b060,645

Further testing showed that this commit actually caused a more serious regression, as many exponents produce an error:

Code:

20220430 10:13:23 Tesla V100-SXM2-16GB-0 Exception gpu_error: INVALID_KERNEL_ARGS fftMiddleIn at clwrap.cpp:324 run

This can be seen by looking at the timings for those same last 14 commits, but with 43 exponents selected by @kriesel (see here):

Code:

,57000991,63000083,67000177,73004279,76000207,84000017,95000011,103246861,113000033,125939521,131000021,144202441,150000029,169000061,187101781,205000013,223000051,247001701,260001727,283000171,295000007,331000037,367000099,403000007,438000131,487001743,509000099,559001657,580001651,650004253,720000049,791000053,861000113,960009689,999999929,1100000017,1138000001,1250000029,1410000023,1690000133,1891000019,1960000019,2147483563
v7.2-93-ga5402c5,646,,742,,841,939,1037,1137,1236,1335,1434,1536,1679,1896,,,,,,,,,,,,,,,7662,8667,,,,,,,,,,,,,
v7.2-92-g5fb55ca,645,,741,,841,939,1037,1136,1236,1334,1434,1536,1680,1897,,,,,,,,,,,,,,,7662,8666,,,,,,,,,,,,,
v7.2-91-g9c22195,645,,741,,841,939,1037,1136,1235,1335,1434,1536,1680,1897,,,,,,,,,,,,,,,7662,8666,,,,,,,,,,,,,
v7.2-90-g75d4a1d,646,,741,,841,939,1037,1136,1236,1335,1434,1536,1679,1897,,,,,,,,,,,,,,,7662,8666,,,,,,,,,,,,,
v7.2-89-g885b8af,647,,741,,841,939,1037,1136,1235,1335,1434,1536,1680,1897,,,,,,,,,,,,,,,7663,8666,,,,,,,,,,,,,
v7.2-88-g599b2b2,647,,741,,841,939,1037,1136,1235,1335,1434,1536,1679,1896,,,,,,,,,,,,,,,7662,8666,,,,,,,,,,,,,
v7.2-87-gc1d9e26,646,,741,,841,939,1037,1136,1235,1335,1434,1536,1679,1896,,,,,,,,,,,,,,,7662,8665,,,,,,,,,,,,,
v7.2-86-gddf3314,647,,742,,841,939,1037,1136,1236,1335,1434,1536,1679,1896,,,,,,,,,,,,,,,7662,8666,,,,,,,,,,,,,
v7.2-85-g6122a0e,647,,741,,841,939,1037,1136,1235,1335,1434,1536,1679,1896,,,,,,,,,,,,,,,7662,8666,,,,,,,,,,,,,
v7.2-84-gce4fe12,645,,741,,841,939,1037,1137,1235,1335,1434,1536,1679,1896,,,,,,,,,,,,,,,7662,8666,,,,,,,,,,,,,
v7.2-83-g1aff945,647,,741,,841,939,1037,1136,1235,1335,1434,1536,1679,1896,,,,,,,,,,,,,,,7663,8666,,,,,,,,,,,,,
v7.2-82-g7bea16f,347,408,401,470,451,501,546,610,653,706,759,807,852,962,1054,1184,1262,1376,1480,1548,1689,1918,2112,2367,2550,2778,2978,3134,3591,4688,5159,5763,6140,6750,7324,7591,7244,9559,10522,12523,13808,14983,15507
v7.2-81-g5f17913,350,400,405,460,454,505,549,613,656,711,763,810,856,966,1059,1190,1266,1380,1490,1550,1703,1918,2106,2372,2547,2778,2984,3129,3573,4621,5093,5710,6072,6655,7253,7539,7204,9448,10375,12386,13634,14802,15341
v7.2-80-g9a975f9,350,386,405,438,454,505,549,613,656,711,764,809,841,949,1043,1172,1243,1355,1468,1527,1704,1917,2107,2373,2545,2778,2986,3128,3500,3934,4361,4836,5237,5641,6147,6452,7149,8127,9022,10794,11749,12718,13339
v7.2-79-g3b4b060,342,386,395,439,443,495,540,602,645,700,752,798,841,950,1044,1173,1243,1355,1468,1526,1697,1921,2106,2367,2546,2778,2992,3136,3500,3933,4361,4834,5237,5641,6145,6452,7154,8169,9047,10831,11784,12747,13377

The empty cells are where it failed to run due to this error. If anyone is interested, I attached the full CSV data for these 43 exponents, but for the last 71 commits.

Magellan3s · 2022-05-07, 18:15

This test is an ASUS RTX 3090 STRIX

+100 MHX +900 Mhz Memory

GPU Owl V6

Code:

jesus@Magallan:~/gpuowl-6$ ./gpuowl -prp 113613007 -iters 30000
2022-05-07 13:12:03 gpuowl 
2022-05-07 13:12:03 config: -user Magallanes -cpu Magellan -block 1000 -maxAlloc 23500M
2022-05-07 13:12:03 config: -prp 113613007 -iters 30000 
2022-05-07 13:12:03 device 0, unique id ''
2022-05-07 13:12:03 Magellan 113613007 FFT: 6M 1K:12:256 (18.06 bpw)
2022-05-07 13:12:03 Magellan Expected maximum carry32: 4CFA0000
2022-05-07 13:12:03 Magellan OpenCL args "-DEXP=113613007u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=12u -DPM1=0 -DMM2_CHAIN=1u -DMAX_ACCURACY=1 -DWEIGHT_STEP_MINUS_1=0x1.d7719ff404155p-1 -DIWEIGHT_STEP_MINUS_1=-0x1.eae2bbc5c8218p-2  -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only "
2022-05-07 13:12:03 Magellan 

2022-05-07 13:12:03 Magellan OpenCL compilation in 0.00 s
2022-05-07 13:12:06 Magellan 113613007 OK        0 loaded: blockSize 1000, 0000000000000003
2022-05-07 13:12:06 Magellan validating proof residues for power 8
2022-05-07 13:12:06 Magellan Proof using power 8
2022-05-07 13:12:13 Magellan 113613007 OK     2000   0.00%; 2145 us/it; ETA 2d 19:42; 0f1a44508c206809 (check 2.18s)
2022-05-07 13:13:12 Magellan Stopping, please wait..
2022-05-07 13:13:14 Magellan 113613007 OK    30000   0.03%; 2124 us/it; ETA 2d 19:01; 32d4895e2a4b9a36 (check 2.18s)
2022-05-07 13:13:14 Magellan Exiting because "stop requested"
2022-05-07 13:13:14 Magellan Bye
jesus@Magallan:~/gpuowl-6$

GPU Owl Newest Version

Code:

jesus@Magallan:~/gpuowl-master$ ./gpuowl -prp 113613007 -iters 30000
20220507 13:08:25 GpuOwl VERSION 
20220507 13:08:25 GpuOwl VERSION 
20220507 13:08:25 config: -user Magallanes -cpu Magellan -block 1000 -maxAlloc 23500M
20220507 13:08:25 config: -prp 113613007 -iters 30000 
20220507 13:08:25 device 0, unique id ''
20220507 13:08:25 Magellan 113613007 FFT: 6M 1K:12:256 (18.06 bpw)
20220507 13:08:25 Magellan 113613007 OpenCL args "-DEXP=113613007u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=12u -DMM2_CHAIN=1u -DMAX_ACCURACY=1 -DWEIGHT_STEP=0.92078876355848627 -DIWEIGHT_STEP=-0.47938054461158819 -DIWEIGHTS={0,-0.45791076534214703,-0.41227852333612641,-0.36280502904659512,-0.30916693173607174,-0.25101366149694165,-0.18796513798337899,-0.11960928626782931,} -DFWEIGHTS={0,0.84471473710626932,0.70148623064852611,0.56937836233036643,0.44752769654326469,0.33513783709142603,0.23147422207537149,0.13585932291445821,}  -cl-std=CL2.0 -cl-finite-math-only "
20220507 13:08:26 Magellan 113613007 

20220507 13:08:26 Magellan 113613007 OpenCL compilation in 0.67 s
20220507 13:08:26 Magellan 113613007 maxAlloc: 22.9 GB
20220507 13:08:26 Magellan 113613007 P1(0) 0 bits
20220507 13:08:26 Magellan 113613007 PRP starting from beginning
20220507 13:08:28 Magellan 113613007 OK         0 on-load: blockSize 1000, 0000000000000003
20220507 13:08:28 Magellan 113613007 validating proof residues for power 8
20220507 13:08:28 Magellan 113613007 Proof using power 8
20220507 13:08:35 Magellan 113613007 OK      2000   0.00% 0f1a44508c206809 2130 us/it + check 2.16s + save 0.11s; ETA 2d 19:12
20220507 13:08:52 Magellan 113613007     10000 28f5eefd6236e274 2174
20220507 13:09:14 Magellan 113613007     20000 d556e5c56bf104e0 2164
20220507 13:09:36 Magellan 113613007 Stopping, please wait..
20220507 13:09:38 Magellan 113613007 OK     30000   0.03% 32d4895e2a4b9a36 2175 us/it + check 2.20s + save 0.11s; ETA 2d 20:38
20220507 13:09:38 Magellan Exiting because "stop requested"
20220507 13:09:38 Magellan Bye
jesus@Magallan:~/gpuowl-master$

Quote:

Originally Posted by Magellan3s

GPU is an EVGA 3080ti FTW3
OS is Linux 20.04.4

Slight GPU Overclock @
+200 MHZ +1000 Mhz Memory

For Wavefront Exponent

Code:

jesus@Magellan:~/gpuowl-6$ ./gpuowl -prp 113613007 -iters 30000
2022-04-19 10:55:04 gpuowl 
2022-04-19 10:55:04 config: -user Magallan3s -cpu Magellan -maxAlloc 10500M -yield
2022-04-19 10:55:04 config: -prp 113613007 -iters 30000 
2022-04-19 10:55:04 device 0, unique id ''
2022-04-19 10:55:04 Magellan 113613007 FFT: 6M 1K:12:256 (18.06 bpw)
2022-04-19 10:55:04 Magellan Expected maximum carry32: 4CFA0000
2022-04-19 10:55:05 Magellan OpenCL args "-DEXP=113613007u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=12u -DPM1=0 -DMM2_CHAIN=1u -DMAX_ACCURACY=1 -DWEIGHT_STEP_MINUS_1=0x1.d7719ff404155p-1 -DIWEIGHT_STEP_MINUS_1=-0x1.eae2bbc5c8218p-2  -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only "
2022-04-19 10:55:05 Magellan 

2022-04-19 10:55:05 Magellan OpenCL compilation in 0.83 s
2022-04-19 10:55:06 Magellan 113613007 OK        0 loaded: blockSize 400, 0000000000000003
2022-04-19 10:55:06 Magellan validating proof residues for power 8
2022-04-19 10:55:06 Magellan Proof using power 8
2022-04-19 10:55:09 Magellan 113613007 OK      800   0.00%; 2132 us/it; ETA 2d 19:18; 420cf6918603e7e1 (check 0.91s)
2022-04-19 10:56:12 Magellan Stopping, please wait..
2022-04-19 10:56:13 Magellan 113613007 OK    30000   0.03%; 2158 us/it; ETA 2d 20:06; 32d4895e2a4b9a36 (check 0.91s)
2022-04-19 10:56:13 Magellan Exiting because "stop requested"
2022-04-19 10:56:13 Magellan Bye

For anyone reading this post or interested:

3080TI vs 3090 performance for PRP = about the same

2022-04-26, 14:58	#24
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 7·13·89 Posts	"Memory Transfer Rate, what Nvidia Settings reports and changes, is different from the effective Memory Clock, what is actually being displayed by GWE" https://gitlab.com/leinardi/gwe/-/bl...ease/README.md (emphasis on effective etc mine). Effective memory clock in this context means the clock rate that would occur, if memory data bits were being sent one at a time per physical signal trace, unencoded, without PAM4 encoding etc., which they clearly per GDDR6X specifications, are not. Last fiddled with by kriesel on 2022-04-26 at 15:00

2022-04-27, 00:48	#26
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 7×13×89 Posts	"memory is running at 1188 MHz (19 Gbps effective)." https://www.techpowerup.com/gpu-spec...ti-20-gb.c3831 1188 MHz there is a clock frequency. Hz is a unit of frequency. Gbps is gigabits per second, a unit of data rate. Two totally different things, different names, different units, different definitions. Note also the word "effective", which can be used as marketing speak for "not really, but we're going to claim this very high performance spec anyway." (I can type at xxx,xxx bits per hour. But that's 8 bits per keypress. The keypress frequency is 1/8 the bit rate.) An apple is fruit. An orange slice is fruit. An orange slice is not an apple. No matter how much someone would wish it to be the same, or believe it to be, or how many others they may get to agree or imply that an orange slice is an apple. Tell me you're learning the difference, or trying? Last fiddled with by kriesel on 2022-04-27 at 01:05

2022-04-30, 18:33	#30
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 7×13×89 Posts	It gets even worse. NVIDIA's own utility for querying the card does not return the "effective" number given in their spec sheets, OR the clock frequency given in common third party utilities. As in https://www.seimaxim.com/kb/gpu/nvidia-smi-cheat-sheet (Run as admin in Windows, root on Linux, suggested) Code: nvidia-smi -q -d SUPPORTED_CLOCKS returns, for GTX1080 and RTX2080 respectively: Code: ==============NVSMI LOG============== Timestamp : Tue Apr 26 19:18:03 2022 Driver Version : 456.71 CUDA Version : 11.1 Attached GPUs : 2 GPU 00000000:01:00.0 Clocks Graphics : 1809 MHz SM : 1809 MHz Memory : 4513 MHz Video : 1620 MHz Applications Clocks Graphics : N/A Memory : N/A Default Applications Clocks Graphics : N/A Memory : N/A Max Clocks Graphics : 1961 MHz SM : 1961 MHz Memory : 5005 MHz Video : 1708 MHz Max Customer Boost Clocks Graphics : N/A SM Clock Samples Duration : 11.27 sec Number of Samples : 100 Max : 1822 MHz Min : 1771 MHz Avg : 1808 MHz Memory Clock Samples Duration : 11.27 sec Number of Samples : 100 Max : 4513 MHz Min : 4513 MHz Avg : 4513 MHz ... GPU 00000000:03:00.0 Clocks Graphics : 1680 MHz SM : 1680 MHz Memory : 6800 MHz Video : 1560 MHz Applications Clocks Graphics : N/A Memory : N/A Default Applications Clocks Graphics : N/A Memory : N/A Max Clocks Graphics : 2100 MHz SM : 2100 MHz Memory : 7000 MHz Video : 1950 MHz Max Customer Boost Clocks Graphics : N/A SM Clock Samples Duration : Not Found Number of Samples : Not Found Max : Not Found Min : Not Found Avg : Not Found Memory Clock Samples Duration : Not Found Number of Samples : Not Found Max : Not Found Min : Not Found Avg : Not Found Drilling down in https://www.nvidia.com/en-sg/geforce/10-series/ gives for GTX1080 "Memory Specs: 10 Gbps Memory Speed 8 GB GDDR5X Standard Memory Config 256-bit Memory Interface Width 320 Memory Bandwidth (GB/sec)" I suppose the vague term "memory speed" could mean peak effective bit rate per signal line. Although elsewhere historically it is used for DIMM data rate: https://www.crucial.com/support/memo...-compatability And in https://www.nvidia.com/en-me/geforce...ards/rtx-2080/ NVIDIA gives "Memory Specs: 14 Gbps 14 Gbps Memory Speed 8 GB GDDR6 8 GB GDDR6 Standard Memory Config 256-bit 256-bit Memory Interface Width 448 GB/s 448 GB/s Memory Bandwidth (GB/sec)" I think an issue with the GWE utility could be graphical layout. Putting a clear description of what was being displayed would use more characters than the app appears to have allowed for. Or its author may not have spent the time to burrow into the terminology carefully. It is time consuming. It would not have surprised me if there was a design something like: Software settable memory clock generator A -> distribution at frequency A to the various memory chips -> on-chip doubler (to b) -> doubler (to c) -> RAM clocking->PAM4 output at 2-bit voltage levels at rate c (computed bit rate d = 2c = 4b = 8A). Or maybe A -> distribution -> on-chip 4x PLL to c -> PAM4 output with effective bit rate d. Or it's QDR, clock A & 90 degree phase shifted A', transfers on both rising and falling edges (edge count c=4 A, PAM4 bit rate d= 2 c = 8 A. GPU-Z, MSI Afterburner etc report A; NVIDIA-SMI reports c; various sources (spec sheets, GWE) report d. I had looked for and not found a reasonably recent reference for GPU circuit designs. References I had found are higher level, architectural, and old enough to precede recent GPU memory types. This one is too vague too. Drilling down further, I do find PLLs, in the memory package documentation: The GDDR6 Micron 16gigibit MT61K512M32 spec sheet: Table 2 at bottom of page 7 uses units of GHz on clock signals CK_t, CK_c, WCK_t, WCK_c; 1.5 or 3 GHz respectively; 3, 6 or 12 Gbit/s/"pin" for data signals, on the fine-pitch ball grid array package. Fig 4 on page 8 shows frequency and data rate ratios relative to the clock fundamental, and clock/data alignment in time. Fig 5 shows example clock and interface circuitry. That is perhaps relevant to my RTX2080 and similar generation GPUs. The internal clock indicated as used in the memory chip ("Internal WCK") is 3. GHz. f, 2f & 4f clock signals; 2f, 4f, 8f data etc rates. The GDDR6X Micron 8gigibit MT61K256M32 spec sheet Table 1 page 5, Figure 3 page 6, and Figure 4 page 7 provide analogous data relevant to RTX30xx GPUs (PAM4 mode); 2.5 or 5 GHz clock signals, ~20Gbit/sec PAM4 (voltage-level 2-bit encoded) effective data rate = 10 GBaud/pin. The internal clock indicated as used in the memory chip ("Internal WCK") is 2.5 GHz. Also https://media-www.micron.com/-/media...fe1f9a5ff0231e So in summary, it looks like at every stage other than the memory chip manufacturer, there's loose handling of clock and data signal rate terminology. (I haven't checked, but I'd expect other GDDR6 chip manufacturers to be similarly clear about their products. A quick online search did not yield any other GDDR6X manufacturers. NVIDIA and every stage downstream from there could do better in using precise terminology to limit confusion, in my opinion.)

2022-05-05, 15:51	#31
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 10358₁₀ Posts	Something may still be fishy with that A100 system, crunching 10xx us/iter for a 1xxM exponent. Does the system do other things in the same time? I am getting 12xx us/iter for a 332M exponent currently crunching, also linux. For the same exponent (332M) the V100 gets 18xx us/iter. Just saying... Not a linux guy. Last fiddled with by LaurV on 2022-05-05 at 15:52

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Benchmarks	Pjetrode	Information & Answers	3	2018-01-07 23:23
RPS benchmarks	pinhodecarlos	Riesel Prime Search	29	2014-12-07 07:13
GPU Benchmarks	houding	Hardware	7	2014-07-09 10:48
LLR benchmarks	Retep	Riesel Prime Search	4	2008-11-06 22:15
Benchmarks	Vandy	Hardware	6	2002-10-28 13:45

2022-04-27, 05:52	#27
VBCurtis "Curtis" Feb 2005 Riverside, CA 1011101110100₂ Posts	Ken, you're the one who should be learning in this thread. The least ambiguous number to cite for speed is the one used and displayed by the drivers. You can complain all you want about how it's not a truthful speed, but those complaints should be aimed at the manufacturer, not at another forumite who is simply adopting the language used by the maker of his card. Why turn to a borderline insult like "tell me you're learning the difference, or trying?" when the OP doesn't care about your personal views?