The conservative memory frequency on Hawaii cards

Toggle sidebar Toggle sidebar
L

Lepton87

Platinum Member
Jul 28, 2009
2,544
9
81
I've been wondering why AMD chose to clock the memory on its flagship cards so conservatively, after OC I can get the same bandwidth on a card with just 384 bit memory bus. From what I remember their 512 bit memory controller takes less die space then that in tahiti, still from what I heard the memory clocks way better then stock settings. Could someone with 290X check how much performance is gained by overclocking the memory to 1500MHz(7GHz) or whatever is stable without touching the core clock. I would appreciate that. A few other tests besides 3D mark would be appreciated.
I did some similar tests on my Titan

1440 MSAA 8X 876(1GHz)/1500MHz 288.4GB/s 74FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1500MHz 288.4GB/s 81.6FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1650MHz 317.8GB/s 84.7FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1200MHz 230.8GB/s 71.6FPS

BW +37%
FPS +16.5%
That is in uniengine tropics

So I find it hard to believe that more bandwidth would offer just minuscule gains
 
S

SinOfLiberty

Senior member
Apr 27, 2011
277
3
81
So I find it hard to believe that more bandwidth would offer just minuscule gains[/QUOTE said:
The new soon to launch Flagship 390X should be very powerful with high bandwidth speed.

If you talking about 290XT(full Hawaii) which is supposedly coming as well, along with 390X; don't bother.

You can have its(290XT) performance with 290 w/ Aftermarket cooler. 390X is what everyone should be looking at.
Click to expand...
 
3

3DVagabond

Lifer
Aug 10, 2009
11,951
204
106
Lepton87 said:
I've been wondering why AMD chose to clock the memory on its flagship cards so conservatively, after OC I can get the same bandwidth on a card with just 384 bit memory bus. From what I remember their 512 bit memory controller takes less die space then that in tahiti, still from what I heard the memory clocks way better then stock settings. Could someone with 290X check how much performance is gained by overclocking the memory to 1500MHz(7GHz) or whatever is stable without touching the core clock. I would appreciate that. A few other tests besides 3D mark would be appreciated.
I did some similar tests on my Titan

1440 MSAA 8X 876(1GHz)/1500MHz 288.4GB/s 74FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1500MHz 288.4GB/s 81.6FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1650MHz 317.8GB/s 84.7FPS
1440 MSAA 8X 1001MHz(1.125GHz)/1200MHz 230.8GB/s 71.6FPS

BW +37%
FPS +16.5%
That is in uniengine tropics

So I find it hard to believe that more bandwidth would offer just minuscule gains
Click to expand...

Most 290X's will do 6GHz on the mem. You'd have to run a 384 bus @ 8GHz for the same bandwidth. The wider bus allows for lower spec'd/priced mem. It's not just the memory controller that limits RAM speed. It's the RAM itself most times.
 
F

f1sherman

Platinum Member
Apr 5, 2011
2,243
1
0
5 Gb/s may seem low, but that's still whooping 320 GB/s

I've found out that raising RAM speed adds considerably to heat output on my 290,

(and vice versa of course)

As with any VRAM overclocking, you may seem fine with your OC, but verifying that is not so simple.
 
L

Lepton87

Platinum Member
Jul 28, 2009
2,544
9
81
f1sherman said:
5 Gb/s may seem low, but that's still whooping 320 GB/s

I've found out that raising RAM speed adds considerably to heat output on my 290,

(and vice versa of course)

As with any VRAM overclocking, you may seem fine with your OC, but verifying that is not so simple.
Click to expand...

Yeah, that's why I always check if the performance actually improves. Is someone willing to do some tests like I did? Maybe even in uniengine tropics for comparison to my results.
In my Titan increasing bandwidth by 37.5% increases performance by 16.5% so it's nice scaling. I would like to see if 290X scales similarly with bandwidth, some tests with underclocked memory would be interesting to see too.
 
R

rtsurfer

Senior member
Oct 14, 2013
733
15
76
Lepton87 said:
That's terrible scaling, very terrible.
Click to expand...

Its okay.

Unlike Valley, in real world, Higher memory OCs on GPUs seldom offer significant gains in performance.

Especially on AMD's side. Nvidia has more improvement with Higher RAM speed.
But this can just be attributed to difference between the architectures.

Edit: You seem to be using some old Unigine Tropics Bench.
I suggest switching to Unigine Valley.

It is more new & relevant.
And also responds greatly to memory OCs.
 
Last edited:
F

Flapdrol1337

Golden Member
May 21, 2014
1,677
93
91
1440p and 8x msaa :)

When amd determined the clocks I don't think they chose that as their target.

Higher memory speed means more expensive memory, more power use of the memory, so maybe the core clock would have to go down a bit and you'd lose performance on more realistic resolutions.
 
L

Lepton87

Platinum Member
Jul 28, 2009
2,544
9
81
rtsurfer said:
Its okay.

Unlike Valley, in real world, Higher memory OCs on GPUs seldom offer significant gains in performance.

Especially on AMD's side. Nvidia has more improvement with Higher RAM speed.
But this can just be attributed to difference between the architectures.

Edit: You seem to be using some old Unigine Tropics Bench.
I suggest switching to Unigine Valley.

It is more new & relevant.
And also responds greatly to memory OCs.
Click to expand...
Valley
1200MHz 37.6 fps
1700MHz 42.6 fps

core clocks the same

still 13% more fps

nothing like 290X in those tests. well, I looked at individual tests and hitman got 13% more performance from 288 GB/s to 400 GB/s
 
Last edited:
L

Lepton87

Platinum Member
Jul 28, 2009
2,544
9
81
Flapdrol1337 said:
1440p and 8x msaa :)

When amd determined the clocks I don't think they chose that as their target.

Higher memory speed means more expensive memory, more power use of the memory, so maybe the core clock would have to go down a bit and you'd lose performance on more realistic resolutions.
Click to expand...

This is a normal resolution for this sort of card, we're not talking about 4k or anything.
 
Attic

Attic

Diamond Member
Jan 9, 2010
4,282
2
76
Like others, i've found I have to up voltage and deal with much higher temps to get memory much higher than 5200-5400mhz.
 
L

Lepton87

Platinum Member
Jul 28, 2009
2,544
9
81
Attic said:
Like others, i've found I have to up voltage and deal with much higher temps to get memory much higher than 5200-5400mhz.
Click to expand...

5000MHz is the default frequency, so you can get like 4-8% frequency uplift? And yet some cards OC to over 1500MHz like that sample from computerbase.de, but on average going from 1250MHz to 1500MHz only give you 4% more performance so unless you happen to be playing a game that is very sensitive to memory bandwidth like Alan Wake or Hitman Absolution or perhaps the best known memory hog the original Crysis game alas not benchmarked by computerbase.de. I guess how much your memory is going to OC depends on the brand, Hynix is thought to be superior then Elpida. I don't know if they use Samsung memory.
 
I

itsmydamnation

Diamond Member
Feb 6, 2011
3,044
3,831
136
it was covered heaps on launch, there is a big power cost increase going from ~5 to ~7ghz. also the memory phy on the chip is smaller on 290 then it is on 7970 yet it is 128bits wider its much more inline with pitcain in terms of size per bit and it ran 1.2ghz slower then Tahiti .
 
L

Lepton87

Platinum Member
Jul 28, 2009
2,544
9
81
itsmydamnation said:
it was covered heaps on launch, there is a big power cost increase going from ~5 to ~7ghz. also the memory phy on the chip is smaller on 290 then it is on 7970 yet it is 128bits wider its much more inline with pitcain in terms of size per bit and it ran 1.2ghz slower then Tahiti .
Click to expand...

About the size I said about it in the OP
From what I remember their 512 bit memory controller takes less die space then that in tahiti
Click to expand...
 
You must log in or register to reply here.

TRENDING THREADS

COMPANY
RESOURCES
FOLLOW
Top Bottom