GekkoScience BM1384 Project Development Discussion

[notlist3d]

I had some issues with the second node's reset that took me until after lunch to figure out, but when it lit up I immediately did several pelvic thrusts and then ran into the office to excitedly request multiple high fives. Why wouldn't I want to share stuff that exciting with you guys?

I'd consider pushing this thing to 300MHz/66GH and see how much it beat the U3 by at its own hashrate, but the chips aren't heatsinked well enough. We'll have to wait until I have a better setup.

The highlighted is part of why I love your products, and I am a am a fan and buyer of them.   Look at lot of the companies to see what your  sharing they would want NDA and sue the crap out of you for releasing it.

I honestly am not a fan of NDA's.  But it is great how you share dev process with us.   I get to see awesome product development and see it along it's way to market, without signing my life away.

[1715378086]

1715378086

[1715378086]

1715378086

[AJRGale]

I had some issues with the second node's reset that took me until after lunch to figure out, but when it lit up I immediately did several pelvic thrusts and then ran into the office to excitedly request multiple high fives. Why wouldn't I want to share stuff that exciting with you guys?

I'd consider pushing this thing to 300MHz/66GH and see how much it beat the U3 by at its own hashrate, but the chips aren't heatsinked well enough. We'll have to wait until I have a better setup.

The highlighted is part of why I love your products, and I am a am a fan and buyer of them.   Look at lot of the companies to see what your  sharing they would want NDA and sue the crap out of you for releasing it.

I honestly am not a fan of NDA's.  But it is great how you share dev process with us.   I get to see awesome product development and see it along it's way to market, without signing my life away.

could be the reason bitmain doesn't want to sell anymore chips, "Cant show 'tem peeps how ya make stuff 'n' things!" and since its all 2nd hand stuff, cant NDA it!

but hey, here's some PDF's to show what the pin-out is, and how to talk to it and how to power it, but you cant show them hoe you make a miner with them!

[notlist3d]

I had some issues with the second node's reset that took me until after lunch to figure out, but when it lit up I immediately did several pelvic thrusts and then ran into the office to excitedly request multiple high fives. Why wouldn't I want to share stuff that exciting with you guys?

I'd consider pushing this thing to 300MHz/66GH and see how much it beat the U3 by at its own hashrate, but the chips aren't heatsinked well enough. We'll have to wait until I have a better setup.

The highlighted is part of why I love your products, and I am a am a fan and buyer of them.   Look at lot of the companies to see what your  sharing they would want NDA and sue the crap out of you for releasing it.

I honestly am not a fan of NDA's.  But it is great how you share dev process with us.   I get to see awesome product development and see it along it's way to market, without signing my life away.

could be the reason bitmain doesn't want to sell anymore chips, "Cant show 'tem peeps how ya make stuff 'n' things!" and since its all 2nd hand stuff, cant NDA it!

but hey, here's some PDF's to show what the pin-out is, and how to talk to it and how to power it, but you cant show them hoe you make a miner with them!

I think their reason is based on dollars.  They have a hot new chip they know they can sell it in a miner in pretty big quantity.

I don't think they care unless you show up with a big stack of money.  Then they might find a way to share.  But I suspect until there is another miner out there BM will not even look at selling chips.  Even then they still might they are kinda hard to read.

[elrippo]

Well sidehack and novak dropped the 1384 by 40%
I am exited to see what happens with a 1385 wenn it comes near the 0,1xx's 

[Meech]

Bitmain is unlikely to offer any chips I presume until their U3 is dicontinued.  I feel it died at launch due to batch orders instead of single units but thats due to shipping costs.
Hopefully we hear about some chips by xmas, that would make my day.

[notlist3d]

Bitmain is unlikely to offer any chips I presume until their U3 is dicontinued.  I feel it died at launch due to batch orders instead of single units but thats due to shipping costs.
Hopefully we hear about some chips by xmas, that would make my day.

They seem to keep the U series alive though.  So most likely one day when U3 is discontinued a U4 will be released.   

Now that they released a ant router that controls U3's I think they will eventually have a U4 that plugs into it.  But I still would rather have one from sidehack, so I hope  BM changes mind. 

[AJRGale]

Bitmain is unlikely to offer any chips I presume until their U3 is dicontinued.  I feel it died at launch due to batch orders instead of single units but thats due to shipping costs.
Hopefully we hear about some chips by xmas, that would make my day.

They seem to keep the U series alive though.  So most likely one day when U3 is discontinued a U4 will be released.   

Now that they released a ant router that controls U3's I think they will eventually have a U4 that plugs into it.  But I still would rather have one from sidehack, so I hope  BM changes mind. 

I thought they were attempting to dump the U3 v1 and v2? since they made more then they know what to do with, then revised it, without fixing the original stability issue, (also using the old bm1382) no one really wants them. Being at the cost of $60-150 to buy one of them units, doesn't help.

When sidehack asked for some chips off BM, they offered him the U3s for the chips instead! On that note, if side just used the U3 chips for the compacs, it would of just been a U2 clone.

Now I can understand if they have run out of BM1384 stock, and there is no BM1385 available in stock to sell to the public since they are all been placed on the S7.

Following what I mentioned before, throwing paranoia at this and not knowing what was/is in BMs NDA, are they refusing since the NDA may of stated "you must not show how you use these chips" and sidehack has shown Dev side of things with 2nd hand chips? ignoring the fact they do have the documentation on them chips out in the public.

..or is the NDA just to cover how and where one receives the chips (which may state that one is required to sacrifice a goat over the couriers van)?

back to reality, we just have to wait it out, for sidehack to setup shop to take in dead/damaged/useless S5 boards or BM to start selling off BM1485 to everyone.

[notlist3d]

Bitmain is unlikely to offer any chips I presume until their U3 is dicontinued.  I feel it died at launch due to batch orders instead of single units but thats due to shipping costs.
Hopefully we hear about some chips by xmas, that would make my day.

They seem to keep the U series alive though.  So most likely one day when U3 is discontinued a U4 will be released.  

Now that they released a ant router that controls U3's I think they will eventually have a U4 that plugs into it.  But I still would rather have one from sidehack, so I hope  BM changes mind.  

I thought they were attempting to dump the U3 v1 and v2? since they made more then they know what to do with, then revised it, without fixing the original stability issue, (also using the old bm1382) no one really wants them. Being at the cost of $60-150 to buy one of them units, doesn't help.

When sidehack asked for some chips off BM, they offered him the U3s for the chips instead! On that note, if side just used the U3 chips for the compacs, it would of just been a U2 clone.

Now I can understand if they have run out of BM1384 stock, and there is no BM1385 available in stock to sell to the public since they are all been placed on the S7.

Following what I mentioned before, throwing paranoia at this and not knowing what was/is in BMs NDA, are they refusing since the NDA may of stated "you must not show how you use these chips" and sidehack has shown Dev side of things with 2nd hand chips? ignoring the fact they do have the documentation on them chips out in the public.

..or is the NDA just to cover how and where one receives the chips (which may state that one is required to sacrifice a goat over the couriers van)?

back to reality, we just have to wait it out, for sidehack to setup shop to take in dead/damaged/useless S5 boards or BM to start selling off BM1485 to everyone.

They can always go back and make more chips.  I could be wrong but they are making the antminer router with older chips I think.  So they have access to chips.  It's just they are hard to work with on buying chips.

Also with it being a controller for U3.  I imagine they will try to sell them aswell.  But I could be wrong.

[AJRGale]

...snip...

They can always go back and make more chips.  I could be wrong but they are making the antminer router with older chips I think.  So they have access to chips.  It's just they are hard to work with on buying chips.

Also with it being a controller for U3.  I imagine they will try to sell them aswell.  But I could be wrong.

yes, and if they have a constant foundry running, no problem, but they make say 10K chips, thats about half a silicon bullion, at minimal failures.
if they are using a 3rd party top make the chips, they will have to wait in line for the next person to finalise their product.

If they are doing it themselves, they will have to retool back to the BM1384 over the xx85. that takes time, and if they are doing it themselves, they would focus on the lithographs they have installed now, instead of realigning the old ones. Also, are they not using a smaller lithograph for the 85 over the 84?

As for the R1 as a controller for the U3 (or any other usb based miners) isn't it just a simple Atheros SoC, 64MB ram, 16MB flash. (no idea, never seen specs) with 2x usb off the SoC, and GPIO to the ASIC (probably a BM1384)..

i could make my old WRT54GS into one of these, with a compac and have USB out to control a U3, enough room for a usb hub, power and a stick with a nice fan, running openWRT with Lukes or Cons software.

[sidehack]

When I first asked about BM1384 they offered me U3 I could pull chips off, but I had to correct the CSR who erroneously believed the U3 had those chips on it. A Bitmain rep also sent me a copy of an NDA, to which I responded with a few questions to clarify a section or two, and I never heard back. If they want some NDA going down before they'll work with me any more, they need to tell me and they need to answer my questions regarding said NDA satisfactorily before I'll sign it.

[notlist3d]

When I first asked about BM1384 they offered me U3 I could pull chips off, but I had to correct the CSR who erroneously believed the U3 had those chips on it. A Bitmain rep also sent me a copy of an NDA, to which I responded with a few questions to clarify a section or two, and I never heard back. If they want some NDA going down before they'll work with me any more, they need to tell me and they need to answer my questions regarding said NDA satisfactorily before I'll sign it.

Make sure you get it in writing or email form on the questions with NDA.  I guess I am a little non-trusting.   I would suggest making Bitmain change the NDA parts you question.... but not a chance Bitmain edits that NDA.

I will be surprised if they give you good answers on some of it.  But I hope they prove me wrong.

[AJRGale]

When I first asked about BM1384 they offered me U3 I could pull chips off, but I had to correct the CSR who erroneously believed the U3 had those chips on it. A Bitmain rep also sent me a copy of an NDA, to which I responded with a few questions to clarify a section or two, and I never heard back. If they want some NDA going down before they'll work with me any more, they need to tell me and they need to answer my questions regarding said NDA satisfactorily before I'll sign it.

communication breakdown on their side, some how I'm not surprised, with their size and activity. you did attempt to communicate with them again?

[QuintLeo]

Well sidehack and novak dropped the 1384 by 40%
I am exited to see what happens with a 1385 wenn it comes near the 0,1xx's

 Bitmain ran the 1384 at it's max voltage in the S5, left a lot of room for undervolting it.

 They appear to be running the 1385 in the S7 at closer to it's mid-point on voltage.
 0.20 on a 1385 might be possible, but I'd not bet it can do much better than that.

[VirosaGITS]

Well sidehack and novak dropped the 1384 by 40%
I am exited to see what happens with a 1385 wenn it comes near the 0,1xx's

 Bitmain ran the 1384 at it's max voltage in the S5, left a lot of room for undervolting it.

 They appear to be running the 1385 in the S7 at closer to it's mid-point on voltage.
 0.20 on a 1385 might be possible, but I'd not bet it can do much better than that.

I think someone did the math and it came down to 0.17 J/GHs, so it seem clear that a much better efficiency can be reached with the BM1385.

It would be great to get some of them in Sidehack's hand, but that might take a while. It doesn't look like Bitmain is really inclined to sell those for now, so we'll be stuck with finding more indirect ways of getting that performance for now.

[Luke-Jr]

I've been working on trying to calculate the reg_data for BM1382/4 on demand, and have an algorithm that seems to mostly work...
But there's 6 exceptional cases in Gekko's cgminer table that don't fit it:

For 181.25 MHz, Gekko is using 0e83, but this seems to actually be 187.5 MHz, while 181.25 MHz should be 0e03.

For 412.5 MHz, Gekko is using 1006, but this seems to actually be 206.25 MHz, while 412.5 MHz should be 1005.

In four cases, Gekko is using what appears to me to be an overly-precise divider out of the norm pattern:

• For 143.75 MHz, Gekko is using 1687, but 0b03 seems like a better choice.
• For 168.75 MHz, Gekko is using 1a87, but 0e03 seems like a better choice.
• For 212.5 MHz, Gekko is using 1086, but 0802 seems like a better choice.
• For 237.5 MHz, Gekko is using 1286, but 0902 seems like a better choice.

Can anyone shed some light on these differences?

For reference, here is the code I am using to calculate the register data:

In C...

Code:

bool bm1382_freq_to_reg_data(uint8_t * const out_reg_data, const float mhz)
{
	if (mhz < 100)
		return false;
	float best_delta = FLT_MAX;
	unsigned best_dc = 0;
	unsigned try_list[4], *tlp = try_list;
	if (mhz >= 200) {
		if (mhz > 400) {
			*(tlp++) = 0x0101;
			*(tlp++) = 0x0205;
		} else {
			*(tlp++) = 0x0202;
		}
		*(tlp++) = 0x0406;
	} else {
		*(tlp++) = 0x0403;
	}
	*(tlp++) = 0x0807;
	for (unsigned *tli = try_list; tli < tlp; ++tli) {
		const float d = *tli >> 8;
		const float df = 25. / d;
		const float delta = fmodf(mhz, df);
		if (delta < best_delta) {
			best_delta = delta;
			best_dc = *tli;
			if (delta == 0) {
				break;
			}
		}
	}
	if (!best_dc)
		return false;
	const float d = best_dc >> 8;
	const float df = 25. / d;
	const unsigned di = best_dc & 0xff;
	const uint16_t reg_num = ((unsigned)((mhz / df) - 1) << 7) | di;
	pk_u16be(out_reg_data, 0, reg_num);
	return true;
}

In Python...

Code:

def freq2reg(f):
	best_delta = 99999
	best_d = 1
	d_to_try = []
	if f >= 200:
		if f > 400:
			d_to_try.append(1)
		d_to_try.append(2)
	d_to_try.append(4)
	d_to_try.append(8)
	for d in d_to_try:
		df = 25. / d
		delta = f % df
		if delta < best_delta:
			best_delta = delta
			best_d = d
		if delta == 0:
			break
	best_df = 25. / best_d
	if best_d == 1:
		di = 1
	elif best_d == 2:
		if f > 400:
			di = 5
		else:
			di = 2
	elif best_d == 4:
		if f > 200:
			di = 6
		else:
			di = 3
	else:  # best_d == 8
		di = 7
	return (int((f / best_df) - 1) << 7) | di

[sidehack]

As far as I know, the numbers are taken from the frequency table found on the webconfig for the S5 controller I hacked up for testing. I think it was copy-paste but you'd have to ask Novak.

[klondike_bar]

I've been working on trying to calculate the reg_data for BM1382/4 on demand, and have an algorithm that seems to mostly work...
But there's 6 exceptional cases in Gekko's cgminer table that don't fit it:

For 181.25 MHz, Gekko is using 0e83, but this seems to actually be 187.5 MHz, while 181.25 MHz should be 0e03.

For 412.5 MHz, Gekko is using 1006, but this seems to actually be 206.25 MHz, while 412.5 MHz should be 1005.

In four cases, Gekko is using what appears to me to be an overly-precise divider out of the norm pattern:

• For 143.75 MHz, Gekko is using 1687, but 0b03 seems like a better choice.
• For 168.75 MHz, Gekko is using 1a87, but 0e03 seems like a better choice.
• For 212.5 MHz, Gekko is using 1086, but 0802 seems like a better choice.
• For 237.5 MHz, Gekko is using 1286, but 0902 seems like a better choice.

Can anyone shed some light on these differences?

IIRC (from s1/s3 tuning) its quite possible to have 2+ codes for the same frequency, as long as you still follow the 'rule sets' that define the code

right now im tinkering with the R1, and might try to utilize your codes to test out smaller increments. however, timing code is relayed as such:
6:125:0983 = ~6.35GH (9.5hr sample)
6:150:0B83 = ~6.81GH (2hr sample)
5.5:150:0B83 = ~6.94GH (20min sample)
5:150:0B83 = ~7.62GH (20min sample)

obviously the middle is frequency value (ie 150), and the last is the relvant hex code. what is the first value though? timing?

[RichBC]

I've been working on trying to calculate the reg_data for BM1382/4 on demand, and have an algorithm that seems to mostly work...
But there's 6 exceptional cases in Gekko's cgminer table that don't fit it:

For 181.25 MHz, Gekko is using 0e83, but this seems to actually be 187.5 MHz, while 181.25 MHz should be 0e03.

For 412.5 MHz, Gekko is using 1006, but this seems to actually be 206.25 MHz, while 412.5 MHz should be 1005.

In four cases, Gekko is using what appears to me to be an overly-precise divider out of the norm pattern:

• For 143.75 MHz, Gekko is using 1687, but 0b03 seems like a better choice.
• For 168.75 MHz, Gekko is using 1a87, but 0e03 seems like a better choice.
• For 212.5 MHz, Gekko is using 1086, but 0802 seems like a better choice.
• For 237.5 MHz, Gekko is using 1286, but 0902 seems like a better choice.

Can anyone shed some light on these differences?

IIRC (from s1/s3 tuning) its quite possible to have 2+ codes for the same frequency, as long as you still follow the 'rule sets' that define the code

right now im tinkering with the R1, and might try to utilize your codes to test out smaller increments. however, timing code is relayed as such:
6:125:0983 = ~6.35GH (9.5hr sample)
6:150:0B83 = ~6.81GH (2hr sample)
5.5:150:0B83 = ~6.94GH (20min sample)
5:150:0B83 = ~7.62GH (20min sample)

obviously the middle is frequency value (ie 150), and the last is the relvant hex code. what is the first value though? timing?

I thought that the first figure was the Delay, bigger number for lower frequencies / longer delay with 6 being about 19ms & 5 about 16ms? The frequency I thought was effectively Text that is displayed but does not set anything? And the hex number set up the divider for the frequency?

Rich

[klondike_bar]

I thought that the first figure was the Delay, bigger number for lower frequencies / longer delay with 6 being about 19ms & 5 about 16ms? The frequency I thought was effectively Text that is displayed but does not set anything? And the hex number set up the divider for the frequency?

Rich

pretty much, just not sure how the first value (seems to accept decimals?) calculate to a delay in ms (which could then easily match to a clock frequency)

ps: i hit 8.52GH (3hrs) on my R1 but am afraid to push it further (not because of heat or chip limits, but for the sake of the AC/DC and DC/DC)

[sidehack]

So it kinda looks like "project discussion" is trying to shift to the pod poll thread. I'd just as soon keep most dev details in one place, I guess.

I'm a bit behind schedule on pretty much everything, but here's what it looks like. Tomorrow should be a pretty big day.

We should be ordering parts for the second half of the Compac batch, meaning I'll be able to start meeting those backorders pretty quickly and then have sticks in stock. The parts order will also include stuff for a new server PSU board design, and two other things which are interesting to this thread, as discussed below.

We'll also be sending off for prototype PCBs for the PSU board, the 8-chip pod miner and a fancy-schmancy BM1384 tester board.

The 8-chip pod miner is taking over the "Amita" name and will be built to take a standard socket 115x CPU cooler, which will need to be modified for height as the ASICs are about 1mm thick instead of the approximate centimeter expected from a socketed processor. I'll also have holes to directly mount an Arctic Freezer 7 Pro cooler without requring its base frame. I've got a couple cases of 'em here and it's what I'd sell with the board if someone wanted a cooler and what I'd use to mount to the ones I kept so that's what I'm working around. Since we like having flexible and entertaining gear, the Amita will have both a barrel jack and a 6-pin PCIe for power. Since it'll be able to draw over 100W at the top-end you'll probably want the 6-pin if you're going to push it, but a 12V 5A brick should get you in the 130-140GH neighborhood. It'll also have jacks for both USB-B and mini USB, so it shouldn't be hard to find a cable. No matter how much you ask, no I won't put a micro USB jack on it because I don't like them at all. The Amita has two fan headers, a 3-wire and a 4-wire. Since 3-wire is just power, ground and a tach signal it requires PWM on the power line to vary speed (which also means a 2-wire fan will work). If you have a decent cooler with a 4-wire fan, which has internal PWM implemented via an external signal, that one will also work on the 4-wire header. It's designed with a temp sensor mounted right in the center of the ASIC field, and the sensor's package is a fraction shorter than the ASICs so with a bit of grease it'll have good enough thermal contact with the heatsink to measure temperatures well without risking keeping the heatsink from contacting the ASICs themselves. The onboard microcontroller which monitors tempeatures will also be responsible for fan speed, and setting the core voltage. We're using a bucked string topology on this one, which keeps the regulator current down while increasing its output voltage, both of which will increase the conversion efficiency over a traditional VRM design while maintaining full adjustability. The total string voltage, as well as its supply current, will be monitored by the onboard controller which means you can actually see the power being dissipated by your ASICs in realtime. I'm fairly certain that's pretty cool.


Regarding the fancy new BM1384 test board, I based all the power and controls off the Compac - by which I mean I pretty much copy-pasted a Compac and started hacking at it a bit. There's holes for a QFN56 test socket, which I hope won't require any modification to work with BM1384. The only problem would be if the ground pin in the center overlaps the VDD corner pads, but given the corners are in corners and round pins don't really have corners it should probably be okay. This board should allow me to rapidly test both the input and output of a pulled BM1384. I could make sure it hashes correctly, make sure it relays data upstream and downstream correctly, and if I wanted to I could probably even grade them with moderate precision (by clocking somewhat high and monitoring input current requirements).

Being able to test (and grade) pulled ASICs will be super handy, since if I end up making a couple hundred Amitas (right now it looks like I have interest enough to merit a batch of 200, if I get that many parts) I'll be going through a couple thousand ASICs. Couple that with other design tasks (which, including our own TypeZero boards, is actually three moderately powerful mining board designs) and I'll need a lot of working ASICs to design and build with. Guess-and-check is going to suck on an 8-chip board, but would be even worse on a 20- to 40-chip board. I'll probably have extra PCBs of this test board, and if it works I might be willing to part with some to folks who may want a better ability to test BM1384 for S5 repair and such. The QFN socket I found, unfortunately, runs about a hundred bucks so I don't really expect most folks to want one but hey, if you're gonna be doing stuff anywhere near the bulk of what I'm gonna be doing in the next month or two you'd want at least one of these things.

So that's the news right now. I may have more coming up soon, which I'm always hopeful will include something like "hey guess what guys I finally got samples of a current-gen ASIC to work with" but I can't say that yet. I'm still stuck with BM1384 projects for now, which it's a darn good chip for sure but it's getting harder and harder to justify.