HashFast announces specs for new ASIC: 400GH/s

[klee]

oh they dont exist, but im saying dont start comparing shit when its not out yet. asshole.

Another biased shithead here! surprise surprise
Fuck ur mother anytime

im not biased you dick head. i have orders with knc, xcrowd and bfl (still waiting)

can we just stop while we are ahead.

Biased AGAINST HF!! what a retard

[ASIC-K]

im not biased! i wanted to order a baby jet but didnt have the BTC. ugh. ok im quitting anyway.

[Vagnavs]

oh they dont exist, but im saying dont start comparing shit when its not out yet. asshole.

Another biased shithead here! surprise surprise
Fuck ur mother anytime

im not biased you dick head. i have orders with knc, xcrowd and bfl (still waiting)

can we just stop while we are ahead.

Biased AGAINST HF!! what a retard

you sure, come across as a prick.

[ProfMac]

im not biased! i wanted to order a baby jet but didnt have the BTC. ugh. ok im quitting anyway.

Same here. 

[cypherdoc]

im not biased! i wanted to order a baby jet but didnt have the BTC. ugh. ok im quitting anyway.

Same here. 

do a bank wire.

[ASIC-K]

im not biased! i wanted to order a baby jet but didnt have the BTC. ugh. ok im quitting anyway.

Same here. 

do a bank wire.

nah i dont wanna blindly send them money i cant get back.

[ProfMac]

im not biased! i wanted to order a baby jet but didnt have the BTC. ugh. ok im quitting anyway.

Same here. 

do a bank wire.

I need a group buy.  I'd like to bring in another $1,500.

[HashFast]

We've posted a new update on our blog:
https://hashfast.com/countdown-to-tapeout/

-John

[Vbs]

I'd like to write today about a small piece of why we are confident our product is better than KnCs.

So today's topic: Our silicon design is superior.

Both are 28nm designs, but HashFast's is far more powerful and energy-efficient.

Let's look at KnC's 28nm ASIC, and some basic details as we can pull from their documentation. https://www.kncminer.com/news/news-25

First let's calculate the hash rate per square millimeter of silicon. This is a measure of the efficiency of the design.

Honestly, we don't need much to estimate this. The lid size for their chip is enough to make some good estimates.
  
KnC's diagram shows their chip has a 41.2mm lid, and implies that the silicon under that lid may be between 30mm x 30mm, and 36mm x 36mm. (The additional space is needed for decoupling capacitors and such.)
    Let's use those two numbers as bounds for the size of the silicon under the lid. If the die(s) take up just 30x30mm of the space under the lid, then:
     30x30mm = 900mm^2
     100 GHash / 900 mm^2 = 0.11 GHash/mm^2

    Or if the die takes up a bit more of the space under the lid,
      36x36mm = 1296mm^2
      100 GHash / 1296mm^2 = 0.077 GHash/mm^2

HashFast's Golden Nonce chip: I don't have to estimate the size because I work at HashFast.
   One 18x18mm die is able to do 400 GHash (nominal - more overclocked**)
   Hashing per square mm:
      18x18mm = 324mm^2
      400 GHash / 324mm^2 = 1.23 GHash/mm^2
  
Let's compare those numbers, for the high and low values for KnC's chip:

      1.23 / 0.11 = 11.2
      1.23 / 0.077 = 16

So HashFast's chip is between 11 and 16 times more efficient, in hashing per square mm, than KnC's chip.

This has an impact on how fast we can deliver units to customers. One wafer of HashFast's chips has the same capacity as 11 to 16 wafers of KNCs. The initial engineering run from TSMC is limited to 12 wafers, out of which KNC will be able to satisfy 11 to 16 times fewer customers than HashFast will be able to. You'll get your units faster once production starts from us.

In addition, the HashFast chip operates much more efficiently. You get four times the hash rate for the same amount of power (250W). That's 1.6 GHash/W for HashFast, and 0.4 GHash/W for KnC.

Calculations such as this are a small part of why we are confident that we are delivering a quality product to our customers.

We figure it's time to start sharing such information.

Amy Woodward

VP Engineering
HashFast

** P.S. Simon made me put in the line about overclocking. But no one would ever do that to our beautiful chips, right?

Interesting so, KnC is a 250/900 = 0.28 W/mm^2 chip and HashFast is a 350/324 = 1.08 W/mm^2 chip, requiring a cooling solution that can transfer 5.6x the heat per mm^2. I really hope your cooling solution holds up 24/7!

Out of comparison, an ATI 7970 is a 250/365 = 0.68 W/mm^2 chip and an nVidia GTX Titan is 250/561 = 0.45 W/mm^2 chip.

[Simon Barber]

Interesting so, KnC is a 250/900 = 0.28 W/mm^2 chip and HashFast is a 350/324 = 1.08 W/mm^2 chip, requiring a cooling solution that can transfer 5.6x the heat per mm^2. I really hope your cooling solution holds up 24/7!

Out of comparison, an ATI 7970 is a 250/365 = 0.68 W/mm^2 chip  and an nVidia GTX Titan is 250/561 = 0.45 W/mm^2 chip.

Hashfast's chip at nominal (400GH/s) is expected to consume about 250W of power, not 350W. i.e. 0.77W/mm^2. Overclocked Sandy Bridge E in shipping commercial products with the same cooling system runs 350W with a 425mm^2 die, 0.82W/mm^2.

[Vbs]

Interesting so, KnC is a 250/900 = 0.28 W/mm^2 chip and HashFast is a 350/324 = 1.08 W/mm^2 chip, requiring a cooling solution that can transfer 5.6x the heat per mm^2. I really hope your cooling solution holds up 24/7!

Out of comparison, an ATI 7970 is a 250/365 = 0.68 W/mm^2 chip  and an nVidia GTX Titan is 250/561 = 0.45 W/mm^2 chip.

Hashfast's chip at nominal (400GH/s) is expected to consume about 250W of power, not 350W. i.e. 0.77W/mm^2. Overclocked Sandy Bridge E in shipping commercial products with the same cooling system runs 350W with a 425mm^2 die, 0.82W/mm^2.

OK, so two questions:

1) https://hashfast.com/shop/babyjet/ states 350W power draw (+/- 20%), but the chip only consumes 250W?  100W for cooling/misc?

2) An overclocked Sandy Bridge-E doesn't "run" at 350W, for example check this (pic below). The 349W here is System Peak Power, a very different metric (system vs CPU and peak vs sustained).

Subtracting the system idle power (85W@4.7GHz), even at peak usage it would run at 264W (and it would die from electromigration if run like this 24/7), so 264/425 = 0.621W/mm^2, still 20% below 0.77W/mm^2.



Edit: Let's not forget that a Sandy Bridge-E uses a high quality heatspreader (IHS) with fluxless solder, so the actual contact area with the cooler is much bigger than the die size, reducing the W/mm^2 requirements by a large amount.

[Simon Barber]

OK, so two questions:

1) https://hashfast.com/shop/babyjet/ states 350W power draw (+/- 20%), but the chip only consumes 250W?  100W for cooling/misc?

2) An overclocked Sandy Bridge-E doesn't "run" at 350W, for example check this (pic below). The 349W here is System Peak Power, a very different metric (system vs CPU and peak vs sustained).

Subtracting the system idle power (85W@4.7GHz), even at peak usage it would run at 264W (and it would die from electromigration if run like this 24/7), so 264/425 = 0.621W/mm^2, still 20% below 0.77W/mm^2.

Edit: Let's not forget that a Sandy Bridge-E uses a high quality heatspreader (IHS) with fluxless solder, so the actual contact area with the cooler is much bigger than the die size, reducing the W/mm^2 requirements by a large amount.

350W is a rounded up number for the whole system, including the power lost in the 2 stages of power supply, and fans etc. The chip itself draws 250W @ nominal.

Depends on how hard you cool and overclock your Sandy Bridge E. The company that is assembling our systems specializes in overclocking. They run Sandy Bridge Es overclocked to 350W (CPU power alone, not whole system), using the same cooling system we are using. We are also using a heatspreader.

Metal migration is a well understood phenomenon. We have followed all the fab's rulesets for electromigration so that the current levels we're going to see will not be a problem (even current distribution, and thicker metal layers). Currently in the simulator for EM our chip passes the test for a 5 year lifetime, but fails the 11 year test - and that is running somewhat overclocked, at about 540GH/s.

[jspielberg]

Thanks for the details Simon.

If the chip is rated for 5years at 500GH, why is the warranty only 10 to 30 days?

[Simon Barber]

Thanks for the details Simon.

If the chip is rated for 5years at 500GH, why is the warranty only 10 to 30 days?

Because with open source software, and overclocking we don't know what people will try and push these things to do.

[DeathAndTaxes]

1) https://hashfast.com/shop/babyjet/ states 350W power draw (+/- 20%), but the chip only consumes 250W?  100W for cooling/misc?

Not sure why this is still a question (goes beyoond just HF).  Remember even excluding water cooling pumps, fans, and controller boards there are a lot of efficiency loss between between the wall and the chip.  The chips doesn't run on 12V they run at ~1VDC (exact voltage will depend on ASIC design) and no ATX PSU supplies 1VDC so ANY ASIC will need to convert the 12VDC to ~1VDC used by the chip and that usually means at least a 10% loss right there.  Of course most people don't have 12VDC outlets so you need to convert the wattage at the wall to the 12VDC uses by the ASIC board and that is another 10% loss (depends on PSU rating and load).

So a hypothetical chip (any chip) which uses 250W @ ~1VDC will require at least ~277W @ 12VDC not 250W. ~27W is lost as heat not at the chip die but at the DC to DC PSU.  Now for the ATX PSU to supply 277W @ 12VDC will require 308W @ 120VAC and another 31W will be lost at the ATX PSU.  Add in 20W for pump and radiator fans plus a 10 to 20W safety margin and ~350W system wattage with 250W chip wattage makes perfect sense.  Once again this isn't unique to just HF, it applies to all electronic devices.

Case in point.
Avalon chip wattage: 1.86W @ 1.15 (282 MH/s nominal)
Avalon total chip wattage 1.86W x 212 chips = 394W
Avalon wattage at the wall: 620W
Total system wattage is >57% more than chip wattage.

https://en.bitcoin.it/wiki/Avalon#Power
http://garzikrants.blogspot.com/2013/02/avalon-miner-power-usage.html

[Simon Barber]

Will there be a safety mechanism that shuts off the system if the chip gets too hot?

There is a temperature sensor on the chip, and as long as the open source software has not been incorrectly modified, it will limit the temperature. In fact the software will automatically hash at the fastest possible speed, limited only by temperature. If you deliver better cooling, or colder air, it will go faster.

[jspielberg]

Thanks for the details Simon.

If the chip is rated for 5years at 500GH, why is the warranty only 10 to 30 days?

Because with open source software, and overclocking we don't know what people will try and push these things to do.

Another quick question.

Regarding the on chip thermal control listed in the blog: https://hashfast.com/thermal-control/
Is this thermal control done inside the chip automatically, or does it require the driver/cgminer to throttle down the chip?

edit:  I see my question was already answered by you right above my post. TL;DR - Sensor and driver work in tandem to seek peak performance of the chip.

[Kouye]

If it 90% efficient then to supply 277W @ 12VDC requires 308W @ 120VAC.

Talking about serious PSU, here. The kind that cost more than HF chips.

[DeathAndTaxes]

If it 90% efficient then to supply 277W @ 12VDC requires 308W @ 120VAC.

Talking about serious PSU, here. The kind that cost more than HF chips.

?

[mrb]

Interesting so, KnC is a 250/900 = 0.28 W/mm^2 chip and HashFast is a 350/324 = 1.08 W/mm^2 chip, requiring a cooling solution that can transfer 5.6x the heat per mm^2. I really hope your cooling solution holds up 24/7!

Out of comparison, an ATI 7970 is a 250/365 = 0.68 W/mm^2 chip  and an nVidia GTX Titan is 250/561 = 0.45 W/mm^2 chip.

Hashfast's chip at nominal (400GH/s) is expected to consume about 250W of power, not 350W. i.e. 0.77W/mm^2. Overclocked Sandy Bridge E in shipping commercial products with the same cooling system runs 350W with a 425mm^2 die, 0.82W/mm^2.

Simon, regardless of what you think is doable, you are taking unnecessarily high risks to develop such a high-TDP chip. It would have made a helluva lot more sense to go with, say, 4 chips of 63W each. The die would be 1/4th the size. 1/4th the cost too since you would have 4x more chips per wafer. A little more PCB space would be used (not a big deal). But this would have taken away a lot of the risk of cooling a single hot chip.

Bitcoin mining being an embarrassingly parallel algorithm, there is absolutely no point in trying to made a die as large / as hot as possible, when you can spread computations across chips with virtually no downside.

I see this as a very poor strategic decision made by Hashfast. (And Cointerra, and KnCMiner). Only BFL seems that they will get it right, since their 350W Monarch card should split the workload across 10-30 chips.

If I was an investor in Hashfast, I would have asked you "why do you take the risk of aiming at the highest TDP possible for your gen1 chip?". Please explain your choice.