Getting a 290x + waterblock for a mining/gaming PC, questions!

[faxon]

So title is straight forward, few questions, some about mining, some about gaming, and some about the watercooling aspect which may need to get taken to the CnC section to get answered later.

1: I plan to mine on the card and overclock it, but also plan to game. I am dubious of powercolor's 2 year warranty and support in general, but am not looking to spend to much extra if i have to. What's the general consensus on whose the best to go for when mining?

2: I see different people getting different results with different brands, is there any 290x in particular which is known (at given equal clocks) to have any kind of advantage? my sapphire 7950 seems to mine a lot faster at the same clocks than my 2 7950 windforce cards from gigabyte, around 70MHs faster. the card is also in a small ITX system with a single quad core to itself, where as the others are in a dual core system with a slow hard drive as well and single channel ram cause its got 1 stick. shouldnt really matter but just a thought

3: Has anyone significantly benefit when mining from watercooling and overclocking, or is there generally a mining stability wall that is hit with air cooling on these cards? i know the 290s and 290x dont seem to overclock very well on stock cooling at all and am wondering what peoples results for both gaming stability and mining stability stats have been. I am looping the card and cpu off a dual output D5 top from alphacool which also has a 2600k thats not overclocked (h77 lameness) with 2 black ice extreme GT2 rads and medium speed high static pressure fans (2 yate loons and two silverstone air penetrators). Should handle everything fine including the OC.

4: I plan to eventually get 3 or 4 more of these and put them in another miner. photos of the supertower are on the last page of the cases and cooling gallery right now so you can get an idea of what it will look like, and i plan to watercool the entire box with whatever my breaker can handle in it. anything i should be aware of in terms of how much heat one of these cards outputs so i can plan my rad capacity accordingly? that loop has a swiftech mcr320qp on it now and i was gonna add an alphacool nexooos monsta which has the same performance as a 4 fan 60mm 4x120 rad in a 3 fan profile, but im concerned that even that + the swiftech wont be enough to cool 4 cards at full mining load on water, especially if i decide to drop in my quad core again and OC it to see what it is capable of in games for some benchmarks (quad currently powering the ITX while the miner runs off a celeron).

5: anyone know of any good sales on ref PCB sapphire gigabyte or MSI cards right now? I would prefer sapphire with how fast their last RMA was for me. under 2 weeks turnover, they shipped me a refurb as soon as my bad one came in and the refurb was a better OCer than the old one and had .2 lower default vcore, but the prices outside california are $700+ still and newegg even worse with tax.

 

[wand3r3r]

1. MSI has one of the best warranties, allowing changing TIM/coolers, while Sapphire and most others will void the warranty.
XFX in the US will allow you to change it if you contact them beforehand.

Basically the 290x needs better cooling to shine.

Watercooling will resolve the basic problems with the reference cooler, throttling and noise. Mining is intensive and will throttle reference cards without additional fan speed (thus noise). I think you will be limited to the reference PCB unless there are custom blocks out? In any case the differing factor will probably just be the warranty.

3. If you game you will not be able to use Stilts BIOS mods (unless you boot and flash them back before gaming). If you end up with Elpida RAM the timings can be off so the best mining speed is actually from underclocking to 900 MHz in my case. I haven't tried to go over 1000core due to heat and keeping the fan noise down to a bearable level. I don't know how elpida cards handle core overclocks. Elpida is pretty bad, it would be a shame to get an elpida card for water (again I haven't OCed my elpida card, it could shine at a higher core).

4. If you go multi-gpu with water you'll be fine, but otherwise heat will be a limitation in a case. PCI-E risers are a great approach if you are going with many GPUs for mining purposes.

They are solid miners (290/x). If you keep the temperatures under 94C they are even more efficient.

 

[Shmee]

So that explains why my 290 doesn't get good hashrates from OCing, even under water. It has Elpida, and downclocking actually helps it whether it is set as a 290 or 290x (it unlocks)

I would love to get better hashrates with a tweaked bios, either as a 290 or 290x, but you say the modded BIOSes don't do games well. Why is that?

 

[Kenmitch]

So that explains why my 290 doesn't get good hashrates from OCing, even under water. It has Elpida, and downclocking actually helps it whether it is set as a 290 or 290x (it unlocks)

I would love to get better hashrates with a tweaked bios, either as a 290 or 290x, but you say the modded BIOSes don't do games well. Why is that?

What hashrates do you get?

 

[Shmee]

Best hashrates are around 835 Kh/s, at about 870-950 MHz core.

 

[Subyman]

Overclocking is almost never good for mining. You will need water-cooling to keep them from throttling in a traditional multi-GPU setup. I ran stock speeds water-cooled on my R9 290. I did not tweak anything at all and simply used GUIMiner's high intensity settings for my card. I ran at 850kh/s and my temps never went over 44C, ambient in my room is around 21C. The R9 290 seems to draw more power than my new 780. My 780 only gets to high 30's under load in the same set up.

I did check my load while mining. I had one monitor connected as well, but I was pulling around 380W from the wall. That includes pump, idle CPU, fans, PSU inefficiency (Platinum though), and the monitor. I would say the card was pulling near 300W at stock settings.

 

[Attic]

So that explains why my 290 doesn't get good hashrates from OCing, even under water. It has Elpida, and downclocking actually helps it whether it is set as a 290 or 290x (it unlocks)

I would love to get better hashrates with a tweaked bios, either as a 290 or 290x, but you say the modded BIOSes don't do games well. Why is that?

The modded bios's from The Stilt knock off half the ROP units AFAIK.


On your card you may try core and clock combinations from 850c and up and 1000m and up. 925c/1250m should play nice there without creating a ton of heat. Can try down to 800c/1000m, performance should not suffer much.


Name of the game for minining is consistency. 800-850khs on the 290 is fine. If it's not in the cards to get 950kh/s i wouldn't sweat it. I have had more favor underclocking my mining cards and keep temps and fan rpm's low as possible.

 

[wand3r3r]

^ Yeah he has cut down some of the computing power which is unnecessary in an attempt to save power. It seems to work too, some users are getting much higher hashrates while using a little less power.

Overclocking is almost never good for mining. You will need water-cooling to keep them from throttling in a traditional multi-GPU setup. I ran stock speeds water-cooled on my R9 290. I did not tweak anything at all and simply used GUIMiner's high intensity settings for my card. I ran at 850kh/s and my temps never went over 44C, ambient in my room is around 21C. The R9 290 seems to draw more power than my new 780. My 780 only gets to high 30's under load in the same set up.

I did check my load while mining. I had one monitor connected as well, but I was pulling around 380W from the wall. That includes pump, idle CPU, fans, PSU inefficiency (Platinum though), and the monitor. I would say the card was pulling near 300W at stock settings.

That blanket statement is simply not true.

Memory & Timings

Generally memory seems to have timings which are set in a range of memory clocks. e.g. 1375-1500 use the same timings so the hashrate increases as you go up to 1500 and then at 1505 it drops off a cliff as the latency increased the next bracket until 16xx or 17xx where it's at the highest speed for that latency. If you can overclock to the high end of the memory latency range you will see benefits and each bracket brings a little more benefits. Read the Stilts explanations for better info.

On my 290x's I found 1250 and 1500 to be the best frequencies (Hynix and Elpida, but quite a difference between them), with 1500 being better than 1250. I've only gone from <1000 up to the AB max of 1625 and I believe the next "high" would be around 1750 based on the pattern.

Both Hynix and Elpida like 1500 memory, but the elpida hashrate is lower at 1000 core and increases as you go down to ~900 core, and even still it's considerably slower then the Hynix.

What the stilts bioses do are "fix" the memory timings so that they are linear from e.g. 1250 to 1375 on the 290/x. The Elpida and manufacturer's timings are very lax and don't progress linearly they are quite screwy according to the Stilt.

My understanding of it would be in DDR where you have the clock speed and timings. If you have 1333 DDR3 with a timing of 10 it's the same "latency" (in ns) as 2666 with a timing of 20. These timings are critical for mining.

Core

Overclocking some cores doesn't seem to benefit much, likely due to some relation to the memory. Certain cores like higher clocks more than others though, my 7950 loves more core speed and increases the hashrate with the core increase nicely. One 290x has to be dropped (probably to align with the memory). I suspect if it were able to OC I would be able to find a point where it hashes faster than 900 core, but it could be unattainable due to thermal constraints (is it 1025 or 1200 Idk).

The biggest issue with overclocking is the effect on power consumption and if you have to overvolt it increases dramatically. Overclocking has a dramatic effect on some cards, but not much on others.

 

[Subyman]

Notice, I said almost. If you want to fiddle with the card for hours for a few more hash then go for it. Typically downclocking is much more beneficial from a hash/power perspective.

 

[faxon]

1. MSI has one of the best warranties, allowing changing TIM/coolers, while Sapphire and most others will void the warranty.
XFX in the US will allow you to change it if you contact them beforehand.

Basically the 290x needs better cooling to shine.

Watercooling will resolve the basic problems with the reference cooler, throttling and noise. Mining is intensive and will throttle reference cards without additional fan speed (thus noise). I think you will be limited to the reference PCB unless there are custom blocks out? In any case the differing factor will probably just be the warranty.

3. If you game you will not be able to use Stilts BIOS mods (unless you boot and flash them back before gaming). If you end up with Elpida RAM the timings can be off so the best mining speed is actually from underclocking to 900 MHz in my case. I haven't tried to go over 1000core due to heat and keeping the fan noise down to a bearable level. I don't know how elpida cards handle core overclocks. Elpida is pretty bad, it would be a shame to get an elpida card for water (again I haven't OCed my elpida card, it could shine at a higher core).

4. If you go multi-gpu with water you'll be fine, but otherwise heat will be a limitation in a case. PCI-E risers are a great approach if you are going with many GPUs for mining purposes.

They are solid miners (290/x). If you keep the temperatures under 94C they are even more efficient.

Risers arent an option, this box is being designed to withstand 85-90f+ days at full load sitting in the hottest room of the house. I have an air cooled farm in the garage which I plan to riser out but I'm concerned for the summer when the garage no longer is the bastion of stored cool air it is during the other half of the year. I will consider the bit about MSI and coolers for when I order more, for now I managed to get a sapphire 290x on sale yesterday on newegg for 599 (from 649) and the block will be ordered this coming week since funds wont be free for it until after this weekend.

Thanks for the tip about the elpida memory, i will be on the look out for that if i get any with it.

 

[faxon]

So that explains why my 290 doesn't get good hashrates from OCing, even under water. It has Elpida, and downclocking actually helps it whether it is set as a 290 or 290x (it unlocks)

I would love to get better hashrates with a tweaked bios, either as a 290 or 290x, but you say the modded BIOSes don't do games well. Why is that?

dude that's the suck, that explains a lot. I remember seeing those on the card when we put the loop together and wondering about it but had forgotten, was gonna hit you up today about it after seeing this. didn't notice you already had (idk how i missed it). You need to hit me back today anyway though, we're supposed to figure out why my 890FX-GD80 rig is throttling all miners put into it remember? We can get it fixed hopefully and then transfer the cards you're buying from me so I can pay off the charge on my card for that 290x. more hashes for everyone!

 

[Coldblackice]

^ Yeah he has cut down some of the computing power which is unnecessary in an attempt to save power. It seems to work too, some users are getting much higher hashrates while using a little less power.



That blanket statement is simply not true.

Memory & Timings

Generally memory seems to have timings which are set in a range of memory clocks. e.g. 1375-1500 use the same timings so the hashrate increases as you go up to 1500 and then at 1505 it drops off a cliff as the latency increased the next bracket until 16xx or 17xx where it's at the highest speed for that latency. If you can overclock to the high end of the memory latency range you will see benefits and each bracket brings a little more benefits. Read the Stilts explanations for better info.

On my 290x's I found 1250 and 1500 to be the best frequencies (Hynix and Elpida, but quite a difference between them), with 1500 being better than 1250. I've only gone from <1000 up to the AB max of 1625 and I believe the next "high" would be around 1750 based on the pattern.

Both Hynix and Elpida like 1500 memory, but the elpida hashrate is lower at 1000 core and increases as you go down to ~900 core, and even still it's considerably slower then the Hynix.

What the stilts bioses do are "fix" the memory timings so that they are linear from e.g. 1250 to 1375 on the 290/x. The Elpida and manufacturer's timings are very lax and don't progress linearly they are quite screwy according to the Stilt.

My understanding of it would be in DDR where you have the clock speed and timings. If you have 1333 DDR3 with a timing of 10 it's the same "latency" (in ns) as 2666 with a timing of 20. These timings are critical for mining.

Core

Overclocking some cores doesn't seem to benefit much, likely due to some relation to the memory. Certain cores like higher clocks more than others though, my 7950 loves more core speed and increases the hashrate with the core increase nicely. One 290x has to be dropped (probably to align with the memory). I suspect if it were able to OC I would be able to find a point where it hashes faster than 900 core, but it could be unattainable due to thermal constraints (is it 1025 or 1200 Idk).

The biggest issue with overclocking is the effect on power consumption and if you have to overvolt it increases dramatically. Overclocking has a dramatic effect on some cards, but not much on others.

Thanks much for this informative post. A couple ?'s, if you don't mind:

1. Seems like memory timings are set up like "fan profiles", so to speak. If so, are memory timings changing within a stage, or between stages? E.g., if one speed "stage" is 1300-1500mhz, will timings be set at a static value through all of 1300-1500mhz, or will timings changing dynamically throughout 1300-1500 (before resetting the cycle for the next stage)? From your post, it seems they move within a stage.
2. If timings do change within a stage, do they then have a "set" curve across all clock stages? E.g., from the first 0-10% clock-speed of a stage, timings set at X; from 11-30%, timings set at Y; 31-60%, set at Z, etc.? Or are timing curves different even across differing clock "stages"?
3. Is it possible to view what memory timings + clock "stages" are set at in a bios (without having to "reverse-engineer" through trial and error by watching where bandwidth increases according to a specific memory overclock)?
4. When tweaking memory timings, is it a matter of tweaking the "stages" (i.e., 900-999mhz, 1000-1349mhz, 1350-1600mhz, etc.), or are you tweaking the actual/specific timings, like you would with DRAM (e.g., CL, CAS, RAS, etc.)? Or both?
5. Are there any tools available for tweaking timings, or is it a matter of hard-editing hex offsets in a bios?

Feel free to answer any/all/none, and however detailed or brief you can afford. I thought it'd be better to condense the questions into one post, rather than sprinkle their related discussion across numerous threads.

 

[wand3r3r]

Thanks much for this informative post. A couple ?'s, if you don't mind:

The Stilt the guy who created a number of alternative BIOSes optimized for scrypt mining, has a good explanation. Take note that this guy was able to produce custom BIOSes for the reference 290/x cards which are signed. Clearly he has expertise and connections in the industry.

(Note this is from the 79xx thread)
https://litecointalk.org/index.php?topic=12369.0
A few excepts:

Against common beliefs there is no magical GPU:MEMCLK ratio (e.g. "0.57") for optimal scrypt mining performance.

There are simply just badly calibrated VBIOSes.

Scrypt mining requires plenty of memory however again against the common beliefs it is not bandwidth intensive.
Instead it can be quite sensitive to memory latency.

..
1375MHz (1251-1375MHz)
1500MHz (1376-1500MHz)
1625MHz (1501-1625MHz)
1750MHz (1626-VCO Max)

So when the default memory clocks of a 7970 card is 1500MHz, the memory controller will drive timings defined by the 1500MHz clock profile (if present). As soon as the user sets the memory clock to 1501MHz (e.g.) the memory controller will start driving timings defined by the 1625MHz profile. The same thing happens if the clocks are lowered to 1375MHz for example.

Since the higher frequency profiles will usually have looser timings a slight increase in clock frequency might actually increase the memory latency instead of lowering it. For example at 1500MHz the GDDR5 timings for Hynix modules are 16-17-17 (tCL-tRCD-tRP, in MEMCLKs). This means that the actual cycle time is 10.66ns for tCL, 11.33ns for tRCD and tRP (1000 / 1500 * delay in MEMCLKs). For the 1625MHz clock profile the timings are 17-18-18. This means that if you overclock the memory from 1500MHz to 1550MHz the actual memory latency will increase while the bandwidth increases: 10.967ns for tCL and 11.612ns for tRCD and tRP. This is one of the reasons why overclocking the memory is not always desireable.

You have to go to the edge of the steps to get the optimal latency. If you go just over into the next latency you will get a performance penalty.

There are three different types of GDDR5 models used on Radeon 7900 and 280X series cards.
Two of the modules are made by SK Hynix (AFR and MFR) and the third one is produced by Elpida.

The ODMs are free to choose between the three products so the used memory type might vary even between the different batches.

The AFR version made by SK Hynix is the only one with properly calibrated memory block (provided by AMD).
In case the card is equipped with AFR modules there are no significant performance limitations.

However in case of MFR version the performance is absolute rubbish.
This is simply because of a poorly calibrated VBIOS memory block which has been provided to all ODMs by AMD.
The bottle neck is caused by excess latency and the performance impact can be up to 200kHash/s.

The Elpida modules are in the middle of the two Hynix modules.
There is a slight performance limitation however it is not nearly as severe as on Hynix MFR modules.

These may not be directly comparable to the the 290/x but you get the picture between the memory types.

On to your questions.

1. Seems like memory timings are set up like "fan profiles", so to speak. If so, are memory timings changing within a stage, or between stages? E.g., if one speed "stage" is 1300-1500mhz, will timings be set at a static value through all of 1300-1500mhz, or will timings changing dynamically throughout 1300-1500 (before resetting the cycle for the next stage)? From your post, it seems they move within a stage.
2. If timings do change within a stage, do they then have a "set" curve across all clock stages? E.g., from the first 0-10% clock-speed of a stage, timings set at X; from 11-30%, timings set at Y; 31-60%, set at Z, etc.? Or are timing curves different even across differing clock "stages"?
3. Is it possible to view what memory timings + clock "stages" are set at in a bios (without having to "reverse-engineer" through trial and error by watching where bandwidth increases according to a specific memory overclock)?
4. When tweaking memory timings, is it a matter of tweaking the "stages" (i.e., 900-999mhz, 1000-1349mhz, 1350-1600mhz, etc.), or are you tweaking the actual/specific timings, like you would with DRAM (e.g., CL, CAS, RAS, etc.)? Or both?
5. Are there any tools available for tweaking timings, or is it a matter of hard-editing hex offsets in a bios?

Feel free to answer any/all/none, and however detailed or brief you can afford. I thought it'd be better to condense the questions into one post, rather than sprinkle their related discussion across numerous threads.

1. & 2. Static. You want to be at the peak of the frequency with the least latency. If you go over that peak it's slower.
3. There aren't any tools for the 290/x.
4. You can't adjust the latency (although you can use The Stilts custom BIOSes which do fix the poor latencies on the cheap/chinsy ram).
5. Nope.

The reason there aren't any tools for the 290/x is because the BIOS has to be signed, which AFAIK is only possible through AMD/partners so far. The only hope is to try the Stilts BIOSes and see if any are stable. It's a fairly time consuming process unless you just get lucky since you have to first determine if your card works, then slowly work up the settings until you can find if it's faster than the stock BIOS.

The Stilts BIOSes are for reference cards only, which includes a few with custom coolers but not with revised PCBs. If you go with these they will have an effect on gaming since they disable some functionality in the interest of efficiency.

The Stilts 290/x bios thread is here.
https://litecointalk.org/index.php?topic=12830.0

As far as my own observations go:

My elpida card mines faster at 900c/1500m then 1000c.
My Hynix card mines faster at 1000c/1500m or 900c/1250m.

It's funny since you'd think the Hynix card at 900c would still like 1500m but ironically it's slower!

I haven't tested any overclocking since my reference cards are thermally limited and I don't feel like using 100% fan. I have tested everything from 800-1000c and 1000-1625m on the different cards. I have a feeling that the Hynix card would do well with more core, but I don't know about the Elpida.