Ya you guys deserve a heads up and some insight into how things work in China.

A lot of the price changes you see are driven by the trading market in Shenzhen.

Let's take an example:  We list an old model ( say 850)  at  $250 on the website ( just an example ). A customer will come in and order 4000 and pay immediately.  The next minute they are all over alibaba, wechat, etc, being adverstised for  $275  in Shenzhen. Traders basically. there is no stopping them. What it means basically is that we got it "wrong" when we thought the market price was $250.

We try to calculate a market price based on BTC prices and difficulty,  but the market has different expectations for the future. the market may think MOON! and snap up all the product, reselling to make a quick buck.  It's rather vicious and defies everything you might think from a western perspective. Of course things go the other way, when traders think MOON and the opposite happens.  think 2018.

So we try to pick a price where the stuff wont sell too slowly and not too quickly. We also have to decide how much to pre sell and how much to save for future spot market selling.

Lets say  I have   X machines projected to be built in sept.   do I want to pre sell all of them now? or do I want to sell some now at a discount and save some portion in case BTC continues to rocket to the moon?

The best way to look at this is that the miner manufacturers are the longest  betters in bitcoin. When we bet the cash to build a machine  that bet wont be sellable for 90 days ( 60- 90 depends)  if BTC moons, then its all good; if it crashes  there is no way we can exit the position as a trader could with a stop loss. The cash to build a machine gets laid out in advance. we are commited deep and hard.

As BTC continues up expect the same kind of craziness we saw in the end of 2017 when supplies were really tight.

This is just bitcoin. When it moons the hardware manufactures have no way to respond rapidly; we will always lag the moon because  it takes time to make machines. The shortage of machines  then works to make it moon harder and  if difficulty does not  keep  up to moderate profitability  then the feedback for even more demand is set in place. and it moons harder.!   If the hardware manufacturers  overbuild (2018), then eventually  difficulty will kick in to moderate profits and then reduce profits... and then the hardware costs come down... just in time for the slow movers to jump on the train and make it even more difficult.. or it comes down so hard that miners exit..

Long ago I used to work in modelling.

I tried to model this.

still trying.

Bottom line.  It is selling out pretty quickly so that now most large customers are pre purchasing  Sept/Oct.

capacity at the fabs  ( TSMC and Samsung) is pretty tight which means  NO ONE will be able to build a huge wack of machines  ( unlike Q1 and Q2 of 2018 )

With Upstream supply tight and BTC mooning  even old shitty stuff is flying out the door. We just started selling 850! which is basically 851  that didnt met the spec. Moving like crazy.

it looks to me like we will have a repeat of end of 2017  when demand was 10X  the supply. If BTC holds or goes higher then you can expect shortages.

life in bitcoin, never dull.

I will try to reserve some quantity for smaller customers and put some aside for Nick at blokforge. he stuck with bitcoin and us through thick and thin.

haha  no not our first. It was in previous versions, just not exposed. but now that the patent issues are resolved...

machine operates up to 40C, been testing from -20 to 40C.

Controller is in the miner, on chip. Our AI chip runs cgminer

hi guys I am back.

we have  completed our enviromental testing up to 40C,  for all the modes of the miner and yes they can take the heat.

More samples should be going out shortly , we had some software tweaks and  wanted to test the first few thousand machines in all the modes, especially turbo.

We  will open up the ordering. Demand is running pretty high, and manufacturers are light on inventory so first come first serve

https://canaan.io/shop/

Ok,  we only had an hour or two in the lab and the customer wanted to point the machines at their pool after we did our initial tests. We were in such a rush across beijing in traffic to get to the lab that I totally forgot about alerting you.

We were also surprised by getting 38TH as we predicted 37 based on the voltage setting of the PSU. There is even more headroom in frequency, ( say at least 33%) but we may hit cooling limits and we  hit the limits of the power strip when we pushed it beyond 2500.

I need to do some more structured tests and longer times at temperature, Needs some prep time and we will give you a heads up  when we head to the lab again.

It was just a link to a youtube vid of the test at the enviromental lab. I emailed it to Matti.
its on my youtube channel.

Finalizing pricing now. It's rather difficult because most of the pricing is done on a  RMB per TH  basis in china. Irrespective of efficiency.

In short, the domestic market doesnt care about efficiency that much.

Lets  compare:

38  TH   @  2500  watts    .065J/GH    @ 3 cent power   =   $5.79  per day     2 years = 4226
31  TH   @  1800  watts    .058J/GH    @ 3 cent power   =    $4.89  per day.   2  years = 3569

Same machine:  same cost.  and everyone sells on RMB per TH.. well which   performance do you pick to price it?

Now compare M10

33  TH  @  1850 Watts    .056 J/GH    @ 3 cent power   = $5.26  Per day          2 years  = 3839

Now compare

33  TH  @  1850 Watts    .056 J/GH    @ 6 cent power   = $3.93  Per day          2 years  = 2868
38  TH   @  2500  watts    .065J/GH    @ 6 cent power   =   $3.99  per day       2 years =  2912

Of course guys here know how to do this math. The math on having a reliable system is harder. The math on how many machines actually deliver the promised spec ( or higher) is also harder to estimate. The math on reducing your number of network connections  because of daisy changing is probably easier.

So, its trivally true that the lower your power price the  more likely you are to trade efficiecny for hash rate. The vast major of huge customers are below  6 cents, all in. So they have to look at all the various modes the machine can run in. After seeing what Phillip was doing, I kinda pressed on engineering to make on the fly frequency changes a lot easier.  There is STILL a long way to go on what I would call proper specifications for systems: in particular I would like to see independent lab testing of machines and like to finally publish some hard data on how performance changes as a function of temperature. Who knows with a built in neural net asic maybe somebody will discover that the machine can learn to control itself. The clock cores already adaptively adjust  their clock rate as a function of DH.

Any way,   a few other updates.  machines will probably be built in a mix of with PSU  and W/O PSU. You can run the machine  with an integrated PSU ( it has voltage control over I2C)  or just attach your own PSU. And if you get it with the integrated PSU  you can always switch it.

Price on the website is  just a stand in price.  It will be below that once we start taking orders. While we have a bunch of machines in test, Lily and I want to get test results from the first  Trial run. You should see those system hit the pool any day now.

I will email you some other stuff  that the mod deleted.

A miner consists of 2 hash boards and a  MMU.  the PMU is gone. The MMU ( miner management unit ) has two  versions, and this will quickly transition to 1 version.

The first version uses the old style rasb pi  to AUC.  My best estimate is that only a few machines will be built in that style.

In parallel we are finalizing test on a new style MMU  that has  RJ45  instead of the old 5 pin connector. with this version, the PI will be gone, the AUC will be gone and machines will daisy chain over ethernet. no real limit on how many you can chain..

This MMU is up and working, final gerber released and so  mass pro will switch over  ASAP.  If you get a AUC  version, we will probably have a  replacement MMU ( with RJ45)  to upgrade. I ordered the parts to do this.

---

The  AI chip is here

https://kendryte.com/

https://kendryte.com/downloads/

it's not scary.

---

I was running at 38TH earlier today  at 40C in the enviromental test lab. fun times.

---

Nope its 16nm  because the wafer cost per TH on yeilded die is lower on 16nm than 7nm.

No controller required in the RJ45  version, cgminer runs on the AI chip,  dual core  risc V.

I think the option with the non integrated PSU will probably rule the day, especially when folks figure out that if you have the right PSU, you can run it   above 31TH.. (31 TH is running about 1/2 the max clock) you'll take an effciiency hit that more than pays for itself at lower power prices.  

The purchasing decision and operating decisions become a bit more complex when you look at running the machine in max effciency mode ( say 24TH ) versus running it in max hash rate mode ( 38+ today) And how do I just put one number on the spec sheet? and then you have to factor in how efficiency changes with ambient temp.

Its almost like an AI problem. just kidding. But I do foresee customers who will now have fine grain control of the frequency that is tied to the actual operating conditions of their farm.

Well Lily and I are back in charge of setting Specs and  firmware requirements!

Trying to get super finegrain control of the power comsumption of machines so that freq can be set and reset on the fly.
In testing on the 841/851   should  migrate  easily.  

On the  10 series there will be TWO  options:

1.  Option with "intgrated" PSU.  basically this means a PSu with I2C control. There will be no PMU. Upside, miner controls
     the PSU and  you dont have to worry about your PSU burning up the PMU ( if the PSU supplies too low of a voltage for ex)
2.  Option for  regular PSU.  The hash rate will take a little hit.

Also the AUC is  gone.  Machines  daisy chain via  internet connection.   more details follow

and now for some really exciting news....

Just open sourced  the 851

http://ehash.org/

I've sent you a mail with an introduction to the local guru. The voltage setting are just one way to change the clock rates.
if you have the cooling and know what you are doing 14-16+ TH is achievable.  Basically  setting zero  gives you
a clock rate of around 750,  but the max is 1G, so theoretical max is 16.25. Do you have the instructions on bypassing
the standard voltage settings?

Finally,  I should probably invite a luck guy to come to beijing to write a overclocking/underclocking guide.
All expenses paid  + free miner. If anyone  on the forum is interested  write me at steve @ canaan.io

That's the intent.

the miner will operate up to 40C,  but above 30C the performance (efficiency) decreases  for SOME  machines.
So,  we would have speced,  40C  and then increased the +-  on the power consumption,  

There is  no issue about the warranty. you wont void the warranty.

---

Moderator's note: This post was edited by frodocooper to correct erroneous quote formatting.

  Ya,  One of the things I am working toward is a manufacturing test report that shows the ACTUAL distribution of machines
power efficiency  during our burn in test.   For the 841   75%  of the machines are over 13.5TH,  but spec wise
we call that  "13".  specksmanship   kinda sucks   so what do I call the 841  Average 13.5?  or Minimum 13?
and how do you get other suppliers to do the right thing?   Any way,  I hope to have the same kind of real measurement
actual distribution of power efficiency performace  at final test.

haha.  we should do a  blockchain for this  with machines writing their ID and peformace into a blockchain

Verdict is in. Samsung.  Explains the yeild issue and the  variance in power efficiency.

---

Moderator's note: This post was edited by frodocooper to remove an unnecessary quote.

  It looks like an S9 but remember that you can find all the design files for an S9 in shenzhen if you know where to look.

  And you dont even need the design files to reverse engineer  it.  you just need cash and connections

  10nm.  

   The first thing is most customers do not know the right questions to ask a mining company to ferret out if they are bullshit or not. So, I encourage everyone to ask the right questions about chips, fabrication,  etc.  If the guy doesnt know what he is talking about it will be clear right away.  

Speculations:  In the beginning a bunch of folks were saying it was a 16nm chip.  And some speculated that it was just
and overclocked S9. Hmm, dunno. The problem with it being a  16nm  chip from a new company is that TSMC would not give a new customer any wafer capicity.

How about  TSMC  10nm?  Well if they answer TSMC  10nm,  then that   fuels another speculation that buzzed around on twitter
that Halong  is  just a Bitmain  subsidary  and  it's Bitmain's method for selling chips. It also helps them clean up their
act  and  diffuse concerns about the concentration in mining. Information vacuums  are never good as they fuel speculation.


Maybe TSMC 12?  performance is somewhat consistent with that.
 
Intel?   Nope.  


Samsung 10nm.   From a performance and shipping standpoint samsung 10nm makes sense because  there is an apparent  yeild issue ( there always is )  and because there is  what looks to be a voltage leak problem which could lead to excessive power consumption. You'll get chips that run at  sub  .09J/GH  and  devices that run above this.  If you can just run the thing through a variety of frequencies and see what happens to the power consumption that will be  a good clue. Maybe they can tweak the  the process to reduce the portion of chips that might have this issue.  Ideally,  they would do  different SKUs,  those that run between 15 and 16 and those that run faster. Hard to do unless you have big volume. In any case  if it is samsung 10nm, then the best thing to do is just to address the performace variations and explain what you are doing to reduce/improve the situation. If they are tweaking the process to run at a fast corner (for example), just tell folks. We are tweaking the process.
If you build chips  it is what you do.

In the end someone somewhere will decap  chip or take it to their dentist for an xray and post the dimensions. I should
do a bounty on that.  If you know the dimensions of the bare die that will tell alot. Has anyone decaped a Bitmain chip?

Finally, the other thing is some guys  use  chip specs  to derive machine specs.  That means you might spec  16TH .07  from
the  single chip performance  ( averages of lots) but when you actually build a machine  its  15-16TH  and  .1  at the wall.
what we do is a little different.  We spec machines  after they are built and tested. So  13TH,  actually means a tested
MINIMUM  13, and empirically  75% of the machines  run at 13.5 or higher.   Ideally there would be some kind of standard
benchmarking  but I dont think the industry is ready for that.

Anyway,  more information from everyone, open information from everyone is always a good thing.

---

Moderator's note: This post was edited by frodocooper to trim the quote from NODEhaven.

Congrats  !

It's simple enough.

10nm

which fab?

how many masks  for that 10nm?

Answers to those two will explain a lot

---

Moderator's note: This post was edited by frodocooper to tidy up quote formatting.

I'm old school, bro. Customer comes first. I wish I could do more. At some point once I have enough guys from this forum trained. I will set them lose to train others.


Thanks. I have  tested this out with a few people and  no news is good news. Once the repair guide and troubleshooting guide is done. Then I will feel better about shipping free spare parts  with every order.

Once I have accurate stats  on failure rates and modes ( stats iwas one of my side gigs-- field failure on autoparts)
I will start to include free spares on larger orders..

small sample stats is hard..LOL

The options  would be.

1.  I ship customers the spares they will  need In advance.  basically the in field failure rate is  1%. so I can just ship units with the order.

2.  Stock parts at Distributors.

3. Stock parts at  Certified technicians. ( hey I stole your idea!)

The Key  of course was explaining to people that  its not optimal to send a whole unit back  IF you design a machine  to be  in field repairable.


The key is designing for reliability.  If you design for reliability then the hardest problem is solved. Marketwise we ship about 25% of the hash rate  last year. Every Bitmain customer that comes to me complains about the same thing: reliability. They report failure rates from 5-30%, so i dont know about the "attention to detail". They key is making a device that is so robust that you cant screw up the manufacturing. They also complained about the downtime  from  shipping back products. I guess the kind of untold story is that our product is designed to be  easily assembled and disassambled  with no specialized training. That means I can ramp production  10X with minimal investment and time. The other untold story is that its easy to repair. When matti and Sky are done with the guide, I hope folks will understand. If not I can always send matti to your farms to train you.  No yacht, sorry, all business here.