really !
Topic: Can KNCMiner really deliver 28 nanometers? - page 6. (Read 11965 times)
really !
I feel rather sorry I brought up this Structured ASIC hypothesis now as its pretty clear from KNCMiner statements that this is wrong (though I was only quoting what had been said elsewhere). I do stand by my comments that their timescale is extremely ambitious and the lack of a test strategy at wafer/package level is very worrying. (And my thanks to ChipGeek for his comments above.)
You make your idea on available information. Pretty much everything pointed at structured ASIC.
I'm still not convinced about the September release (maybe they'll sell the sample chips in a miner as retail
) 
But its been hinted that KNCMiner are doing a 28nM Structured ASIC, ie a HardCopy(tm) of an Altera FPGA design. That's much cheaper in up-front cost (the mask set is just the metalization/vias so much cheaper), but the per-unit cost is higher. It won't perform anything like a full custom ASIC.
https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
My concern is that they're making very lofty performance claims but if it's just an ASIC copy of the FPGA it's not likely to outperform the FPGA crazy significantly. I'm not sure where they're getting all this extra performance per watt from, are they solely depending on the fact that it's 28nm? I'm just worried it'll be another repeat of BFL's power estimates. They might be able to get 350Ghash but I doubt it's going to be come at the low watts they're quoting. https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
I agree. A Structured ASIC whould have little performance gain over the base FPGA part (its essentially the same device, but with the programming hard-wired in), but would have lower unit cost. A full custom ASIC is an entirely different beastie, and at 28nm would expect at least a factor of 2 performance gain over BFL (not quite sure how it scales, ChipGeek will know).
I feel rather sorry I brought up this Structured ASIC hypothesis now as its pretty clear from KNCMiner statements that this is wrong (though I was only quoting what had been said elsewhere). I do stand by my comments that their timescale is extremely ambitious and the lack of a test strategy at wafer/package level is very worrying. (And my thanks to ChipGeek for his comments above.)
But its been hinted that KNCMiner are doing a 28nM Structured ASIC, ie a HardCopy(tm) of an Altera FPGA design. That's much cheaper in up-front cost (the mask set is just the metalization/vias so much cheaper), but the per-unit cost is higher. It won't perform anything like a full custom ASIC.
https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
My concern is that they're making very lofty performance claims but if it's just an ASIC copy of the FPGA it's not likely to outperform the FPGA crazy significantly. I'm not sure where they're getting all this extra performance per watt from, are they solely depending on the fact that it's 28nm? I'm just worried it'll be another repeat of BFL's power estimates. They might be able to get 350Ghash but I doubt it's going to be come at the low watts they're quoting. https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
But its been hinted that KNCMiner are doing a 28nM Structured ASIC, ie a HardCopy(tm) of an Altera FPGA design. That's much cheaper in up-front cost (the mask set is just the metalization/vias so much cheaper), but the per-unit cost is higher. It won't perform anything like a full custom ASIC.
https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
My concern is that they're making very lofty performance claims but if it's just an ASIC copy of the FPGA it's not likely to outperform the FPGA crazy significantly. I'm not sure where they're getting all this extra performance per watt from, are they solely depending on the fact that it's 28nm? I'm just worried it'll be another repeat of BFL's power estimates. They might be able to get 350Ghash but I doubt it's going to be come at the low watts they're quoting. https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
that's why i hedged and got 2 saturns.. figured i'd be able to plug in one of them at the least and sell the other if i must
crazy how much cloudhashing just said they are buying.. 35TH of the stuff.. if the power is way off that scales off the charts
But its been hinted that KNCMiner are doing a 28nM Structured ASIC, ie a HardCopy(tm) of an Altera FPGA design. That's much cheaper in up-front cost (the mask set is just the metalization/vias so much cheaper), but the per-unit cost is higher. It won't perform anything like a full custom ASIC.
https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
My concern is that they're making very lofty performance claims but if it's just an ASIC copy of the FPGA it's not likely to outperform the FPGA crazy significantly. I'm not sure where they're getting all this extra performance per watt from, are they solely depending on the fact that it's 28nm? I'm just worried it'll be another repeat of BFL's power estimates. They might be able to get 350Ghash but I doubt it's going to be come at the low watts they're quoting. https://bitcointalksearch.org/topic/m.2433194 (and subsequent posts).
If you don't get your unit until middle/end of august, you won't see your money back. That's not great odds.
Lol, where do you get this idea o_O The Jupiter miner would still make a marginal profit if the difficulty were 1 billion. 
It is a challenge to both kncminer and us.
If you don't get your unit until middle/end of august, you won't see your money back. That's not great odds.
BS.
both o' yous
If you don't get your unit until middle/end of august, you won't see your money back. That's not great odds.
BS.
If you don't get your unit until middle/end of august, you won't see your money back. That's not great odds.
No.
And we're talking about a 7mm*7mm chip, a 70mm*70mm one like for KNCMINER. That thing is so huge than the yield will probably never be very high (I think they would consider themselves lucky at 70% - rear end fugee number) and that why they HAVE to have a defect management solution in software, otherwise they would just throw away a lot of the ASICs.
Can't you just modularize the unit and plan to have a certain break-down on average? I mean, ASICs like this Jupiter device are already going to have a looot of chips in it anyway. That creates a certain arbitrage opportunity of sorts. Say you can reasonably expect a 70% yield, then you promise a device with 70% of X number of chips theoretical capacity. That way you're virtually guaranteed to get the device you want.This is not like trying to create an Intel CPU where the end-user wants one damn good chip and only really one and high yield is especially important.
Then, as I said, you modularize the chip, building it in units such that if a unit fails you simply turn it off and route around it so that the entire chip itself isn't sour. Sony did this with the PS3 Cell-chip, built in 8 units and assuming that 1 would be failed typically, meaning an 87.5% yield expected. Some chips with 8 units working still deactivated one unit just to remain consistent! But with an ASIC you wouldn't bother with that, you'd just take the extra performance. And the result is their promise +/- 10% was it, or 5%? Anyway, that's my uneducated guess of how things are going to go.
For this initial revision this is what I came away with understanding.
If you buy now, you will get what was promised, but immediately after the design will be revisited and improved upon.
Future buyers will be given greater efficiency and more bang for their buck as Marcus, Michael, Yann and Henrik will demonstrate with Mars over the coming days, but for those now, it's about beating hashrate and delivering in September.
The guys are very confident they have minimised risks as there is no room for error with this initial waffer!
And we're talking about a 7mm*7mm chip, a 70mm*70mm one like for KNCMINER. That thing is so huge than the yield will probably never be very high (I think they would consider themselves lucky at 70% - rear end fugee number) and that why they HAVE to have a defect management solution in software, otherwise they would just throw away a lot of the ASICs.
Can't you just modularize the unit and plan to have a certain break-down on average? I mean, ASICs like this Jupiter device are already going to have a looot of chips in it anyway. That creates a certain arbitrage opportunity of sorts. Say you can reasonably expect a 70% yield, then you promise a device with 70% of X number of chips theoretical capacity. That way you're virtually guaranteed to get the device you want.This is not like trying to create an Intel CPU where the end-user wants one damn good chip and only really one and high yield is especially important.
Then, as I said, you modularize the chip, building it in units such that if a unit fails you simply turn it off and route around it so that the entire chip itself isn't sour. Sony did this with the PS3 Cell-chip, built in 8 units and assuming that 1 would be failed typically, meaning an 87.5% yield expected. Some chips with 8 units working still deactivated one unit just to remain consistent! But with an ASIC you wouldn't bother with that, you'd just take the extra performance. And the result is their promise +/- 10% was it, or 5%? Anyway, that's my uneducated guess of how things are going to go.
I think that's what they are doing. With chips that big they will probably have a lot of defects and they have to deal with it at runtime. i guess the idea is either to make very small chips and trash the non-functioning ones or build them very big and switch off the bad bits.
Depending on the expected fab yields, one way might be better than the other.
It's the wonderful wait and see time. Alea jacta est, nothing we can do about it now.
And we're talking about a 7mm*7mm chip, a 70mm*70mm one like for KNCMINER. That thing is so huge than the yield will probably never be very high (I think they would consider themselves lucky at 70% - rear end fugee number) and that why they HAVE to have a defect management solution in software, otherwise they would just throw away a lot of the ASICs.
Can't you just modularize the unit and plan to have a certain break-down on average? I mean, ASICs like this Jupiter device are already going to have a looot of chips in it anyway. That creates a certain arbitrage opportunity of sorts. Say you can reasonably expect a 70% yield, then you promise a device with 70% of X number of chips theoretical capacity. That way you're virtually guaranteed to get the device you want.This is not like trying to create an Intel CPU where the end-user wants one damn good chip and only really one and high yield is especially important.
Then, as I said, you modularize the chip, building it in units such that if a unit fails you simply turn it off and route around it so that the entire chip itself isn't sour. Sony did this with the PS3 Cell-chip, built in 8 units and assuming that 1 would be failed typically, meaning an 87.5% yield expected. Some chips with 8 units working still deactivated one unit just to remain consistent! But with an ASIC you wouldn't bother with that, you'd just take the extra performance. And the result is their promise +/- 10% was it, or 5%? Anyway, that's my uneducated guess of how things are going to go.
Nice Video 
It's a young Sting in the film Brimstone and Treacle.
Believers.... believers..... let us pray for the skeptics.... thy miners will be done.
That image is just way too big. Have a thought for us poor souls browsing on a Raspberry Pi
with so many skeptics and non believers, how many would bet they will not deliver in September and how many believe they can do it and deliver in September.??
Lets bow down to KNCMiner, for they deliver us in September.
Believers.... believers..... let us pray for the skeptics.... thy miners will be done.
Nice Video
with so many skeptics and non believers, how many would bet they will not deliver in September and how many believe they can do it and deliver in September.??
Lets bow down to KNCMiner, for they deliver us in September.
Believers.... believers..... let us pray for the skeptics.... thy miners will be done.
Jump to: