Overclock.net › Forums › Graphics Cards › AMD/ATI › Polaris Bios Editing ( RX5xx / RX4xx )
New Posts  All Forums:Forum Nav:

Polaris Bios Editing ( RX5xx / RX4xx ) - Page 205

post #2041 of 3328
Bandwidth is a combination of clock bus width and timing. That's why you get more gb/s in oclmembench/aida.

Timings can and do change the bandwidth, in the same way on a broadband connection, if you have latency you can't max out your line speed. That's why they had those tcp tweaks.

So we have to say bandwidth "number" vs useable bandwidth. On my ddr3, I can't go below 9-9-9 on 1600, so I gain by doing 10-10-10 on 1866. Some people can run 8 8 8 1600 which ends up being better on even memory benches. Less overhead means more useable bandwidth.

That's why I wondered what's better, best timings w minor oc Or stock/mid timings with a higher oc. I guess it depends on the ram and what the imc can handle??


Edit: see the confusion about what "bandwidth" means even on the first post:
Quote:
Uber-Mix Extreme v2.1 (Click to hide)
555000000000000022CC1C00AD615B41C0570E152DCB74090068C7000B031420FA8900A0030000001B11333DC0303A17
~236GB/s (~214GB/s Stock) +10% (OclMemBench)
1375 - 1500 - 1625 Mix
No EDC Errors Mem@2100Mhz (1000mv) / Core @1350Mhz
Tested with 3DMark ,Valley,BF1,FC Primal

Stock 2000Mhz Strap = ~49fps
Uber-Mix Extreme = ~53fps
+10% bandwidth = +8% fps

Edited by robnitro - 4/13/17 at 8:59pm
post #2042 of 3328
A picture is worth a thousand words. This is a shoddy hackjob, so maybe 500 words? Y is bandwidth, X is time, the rest is self explanatory.
Roided midgit
(10 items)
 
Midgit
(12 items)
 
 
CPUMotherboardGraphicsRAM
3770k Asus P8Z77I-Deluxe MSI Radeon RX 480 GAMING X Crucial Ballistix XT 
Hard DriveCoolingCoolingCooling
Crucial M500 Noctua NH-D14 Noctua NF-F12 PWM Noctua NF-F12 PWM 
CoolingCoolingPowerCase
Noctua NF-B9 PWM Noctua NF-B9 PWM Corsair SF600 Fractal Node 304 
  hide details  
Reply
Roided midgit
(10 items)
 
Midgit
(12 items)
 
 
CPUMotherboardGraphicsRAM
3770k Asus P8Z77I-Deluxe MSI Radeon RX 480 GAMING X Crucial Ballistix XT 
Hard DriveCoolingCoolingCooling
Crucial M500 Noctua NH-D14 Noctua NF-F12 PWM Noctua NF-F12 PWM 
CoolingCoolingPowerCase
Noctua NF-B9 PWM Noctua NF-B9 PWM Corsair SF600 Fractal Node 304 
  hide details  
Reply
post #2043 of 3328
Quote:
Originally Posted by robnitro View Post

Bandwidth is a combination of clock bus width and timing. That's why you get more gb/s in oclmembench/aida.

Timings can and do change the bandwidth, in the same way on a broadband connection, if you have latency you can't max out your line speed. That's why they had those tcp tweaks.

So we have to say bandwidth "number" vs useable bandwidth. On my ddr3, I can't go below 9-9-9 on 1600, so I gain by doing 10-10-10 on 1866. Some people can run 8 8 8 1600 which ends up being better on even memory benches. Less overhead means more useable bandwidth.

That's why I wondered what's better, best timings w minor oc Or stock/mid timings with a higher oc. I guess it depends on the ram and what the imc can handle??


Edit: see the confusion about what "bandwidth" means even on the first post:
It depends of what you are using cards for

Anyone here want to help improve card bioses for mining?smile.gif
post #2044 of 3328
Timings are very good for mining , but i am not sure about the performance impact on gaming.Definitely there should such an improvement, but i am not sure how much it would be compared to raw speed.

Later i will do some benchmarks, comparing :

2000mhz/ stock timings 2000/1750 timings 2000/ Uber Light timings
2050mhz/ stock timings 2050/1750 timings 2030/ Uber Light timings
2100mhz/ stock timings 2100/1750 timings cant go over 2030 on Uber timings without getting tons of ERRORS.

Will share the results here.
post #2045 of 3328
Thread Starter 
Quote:
Originally Posted by bardacuda View Post

I never said tighter timings don't help performance. Only that they do not affect bandwidth. Bandwidth is purely a function of bus width and frequency.

Timings are actually one of the best ways to increase mining performance.

Memory subsystems are not 100% efficent

The theorical max Bandwidth is a function of bus width and frequency...the max practical bandwidth is limited by timings.

tighter timings increase the real bandwidth/efficency, like in the system ram.

Real bandwidth < theorical bandwidth
Edited by generaleramon - 4/14/17 at 8:00am
post #2046 of 3328
Thread Starter 
Quote:
Originally Posted by robnitro View Post

Bandwidth is a combination of clock bus width and timing. That's why you get more gb/s in oclmembench/aida.

Timings can and do change the bandwidth, in the same way on a broadband connection, if you have latency you can't max out your line speed. That's why they had those tcp tweaks.

So we have to say bandwidth "number" vs useable bandwidth. On my ddr3, I can't go below 9-9-9 on 1600, so I gain by doing 10-10-10 on 1866. Some people can run 8 8 8 1600 which ends up being better on even memory benches. Less overhead means more useable bandwidth.

That's why I wondered what's better, best timings w minor oc Or stock/mid timings with a higher oc. I guess it depends on the ram and what the imc can handle??


Edit: see the confusion about what "bandwidth" means even on the first post:
Personaly for my system i like to have the lowest ram access time possible and than look for max bandwidth. So i don't like tight timings and very low ram speed because of the max theorical bandwidth limit.

My card can't oc mem past ~2125Mhz... So how can i gain some perfomance? I can only use tighter timings...
I know gpus need a lot of bandwidth ,so i tend to use(if possible/stable) the max frequency and then work on timings.

Ps. With bandwidth i always mean "real bandwidth". Oclmembench show this very well
post #2047 of 3328
I hope this should help you find the best performance combination ratio between speed and timings.

Aida64 Memory Benchmark :

GPU Clock - stock 1305mhz

Memory | Timings | Memory Copy MB/s

2000 | Uber-Light | 201896
2050 | Uber-Light | 206186
2100 | Uber-Light | 212534

2000 | 1750strap | 192314
2050 | 1750strap | 198423
2100 | 1750strap | 203420

2000 | 2000strap | 188731
2000 | 2000strap | 194549
2100 | 2000strap | 200206
Edited by ziXs - 4/14/17 at 1:00pm
post #2048 of 3328
Quote:
Originally Posted by generaleramon View Post

Memory subsystems are not 100% efficent

The theorical max Bandwidth is a function of bus width and frequency...the max practical bandwidth is limited by timings.

tighter timings increase the real bandwidth/efficency, like in the system ram.

Real bandwidth < theorical bandwidth

Yes, this!

When I hear the term bandwidth, I assume we are talking about the theoretical max. I thought that was conventional to avoid ambiguity.

It's good to have it clarified so that people can be accurate in what it is specifically that they're asking about / trying to accomplish.
Quote:
Originally Posted by generaleramon View Post

Personaly for my system i like to have the lowest ram access time possible and than look for max bandwidth. So i don't like tight timings and very low ram speed because of the max theorical bandwidth limit.

My card can't oc mem past ~2125Mhz... So how can i gain some perfomance? I can only use tighter timings...
I know gpus need a lot of bandwidth ,so i tend to use(if possible/stable) the max frequency and then work on timings.

Ps. With bandwidth i always mean "real bandwidth". Oclmembench show this very well

Yes, this! Well-put, sir!
Edited by bardacuda - 4/14/17 at 9:01am
post #2049 of 3328
Thread Starter 
Quote:
Originally Posted by ziXs View Post

I hope this should help you find the best performance combination ratio between speed and timings.

Aida64 Memory Benchmark :

GPU Clock - stock 1305mhz

Memory Timings Memory Copy MB/s

2000 Uber-Light 201896
2050 Uber-Light 206186
2100 Uber-Light 212534

2000 1750strap 192314
2050 1750strap 198423
2100 1750strap 203420

2000 2000strap 188731
2000 2000strap 194549
2100 2000strap 200206

I'll post the ubermix v2.2 in the next days. Stay tuned.wink.gif
post #2050 of 3328
Quote:
Originally Posted by megax05 View Post

Still looking for help here I just want to know which value in the timing strap is responsible for memory bandwidth??

For anyone trying to improve mining performance through timings, try selecting one of your stock straps (either 1500, 1625, or 1750), and putting it in this tool after the ?:

http://doreymills.ca/cgi-bin/strapmod?

Example:

I copied the 1750/2000Mhz mix timings from the op and put it in the address like this:

http://doreymills.ca/cgi-bin/strapmod?777000000000000022CC1C00106A6D4DD0571016B90D060C006AE70014051420FA8900A0030000001E123A46DB354019

It then changes a few values which are important for Ethereum mining and gives an output like this:
Quote:
Rx strap detected
Old, new RRD: 6 , 5
Old, new FAW: A , 0
Old, new 32AW: 7 , 0
Old, new ACTRD: 1E , 0x10
777000000000000022CC1C00106A6D4DD0571016B90D060C006AE70014051420FA8900A0030000001E123A46DB354019
777000000000000022CC1C00106A5D4DD0571016B90D060C0060070014051420FA8900A00300000010123A46DB354019

The last line is your new strap. Give it a try and see if it helps performance! Start with your 1750 strap and if it's stable and improves your Mh/s at 2100MHz (or whatever is max stable at stock), then try the 1625 strap, and so on..

All credit for this belongs to nerdralph over on the bitcointalk forums but please don't spam him with noob questions.
Edited by bardacuda - 4/14/17 at 9:19am
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: AMD/ATI
Overclock.net › Forums › Graphics Cards › AMD/ATI › Polaris Bios Editing ( RX5xx / RX4xx )