Best ASIC devices for Bitcoin mining in 2018 TechRadar

Technical Cryptonight Discussion: What about low-latency RAM (RLDRAM 3, QDR-IV, or HMC) + ASICs?

The Cryptonight algorithm is described as ASIC resistant, in particular because of one feature:
A megabyte of internal memory is almost unacceptable for the modern ASICs. 
EDIT: Each instance of Cryptonight requires 2MB of RAM. Therefore, any Cryptonight multi-processor is required to have 2MB per instance. Since CPUs are incredibly well loaded with RAM (ie: 32MB L3 on Threadripper, 16 L3 on Ryzen, and plenty of L2+L3 on Skylake Servers), it seems unlikely that ASICs would be able to compete well vs CPUs.
In fact, a large number of people seem to be incredibly confident in Cryptonight's ASIC resistance. And indeed, anyone who knows how standard DDR4 works knows that DDR4 is unacceptable for Cryptonight. GDDR5 similarly doesn't look like a very good technology for Cryptonight, focusing on high-bandwidth instead of latency.
Which suggests only an ASIC RAM would be able to handle the 2MB that Cryptonight uses. Solid argument, but it seems to be missing a critical point of analysis from my eyes.
What about "exotic" RAM, like RLDRAM3 ?? Or even QDR-IV?

QDR-IV SRAM

QDR-IV SRAM is absurdly expensive. However, its a good example of "exotic RAM" that is available on the marketplace. I'm focusing on it however because QDR-IV is really simple to describe.
QDR-IV costs roughly $290 for 16Mbit x 18 bits. It is true Static-RAM. 18-bits are for 8-bits per byte + 1 parity bit, because QDR-IV is usually designed for high-speed routers.
QDR-IV has none of the speed or latency issues with DDR4 RAM. There are no "banks", there are no "refreshes", there are no "obliterate the data as you load into sense amplifiers". There's no "auto-charge" as you load the data from the sense-amps back into the capacitors.
Anything that could have caused latency issues is gone. QDR-IV is about as fast as you can get latency-wise. Every clock cycle, you specify an address, and QDR-IV will generate a response every clock cycle. In fact, QDR means "quad data rate" as the SRAM generates 2-reads and 2-writes per clock cycle. There is a slight amount of latency: 8-clock cycles for reads (7.5nanoseconds), and 5-clock cycles for writes (4.6nanoseconds). For those keeping track at home: AMD Zen's L3 cache has a latency of 40 clocks: aka 10nanoseconds at 4GHz
Basically, QDR-IV BEATS the L3 latency of modern CPUs. And we haven't even begun to talk software or ASIC optimizations yet.

CPU inefficiencies for Cryptonight

Now, if that weren't bad enough... CPUs have a few problems with the Cryptonight algorithm.
  1. AMD Zen and Intel Skylake CPUs transfer from L3 -> L2 -> L1 cache. Each of these transfers are in 64-byte chunks. Cryptonight only uses 16 of these bytes. This means that 75% of L3 cache bandwidth is wasted on 48-bytes that would never be used per inner-loop of Cryptonight. An ASIC would transfer only 16-bytes at a time, instantly increasing the RAM's speed by 4-fold.
  2. AES-NI instructions on Ryzen / Threadripper can only be done one-per-core. This means a 16-core Threadripper can at most perform 16 AES encryptions per clock tick. An ASIC can perform as many as you'd like, up to the speed of the RAM.
  3. CPUs waste a ton of energy: there's L1 and L2 caches which do NOTHING in Cryptonight. There are floating-point units, memory controllers, and more. An ASIC which strips things out to only the bare necessities (basically: AES for Cryptonight core) would be way more power efficient, even at ancient 65nm or 90nm designs.

Ideal RAM access pattern

For all yall who are used to DDR4, here's a special trick with QDR-IV or RLDRAM. You can pipeline accesses in QDR-IV or RLDRAM. What does this mean?
First, it should be noted that Cryptonight has the following RAM access pattern:
QDR-IV and RLDRAM3 still have latency involved. Assuming 8-clocks of latency, the naive access pattern would be:
  1. Read
  2. Stall
  3. Stall
  4. Stall
  5. Stall
  6. Stall
  7. Stall
  8. Stall
  9. Stall
  10. Write
  11. Stall
  12. Stall
  13. Stall
  14. Stall
  15. Stall
  16. Stall
  17. Stall
  18. Stall
  19. Read #2
  20. Stall
  21. Stall
  22. Stall
  23. Stall
  24. Stall
  25. Stall
  26. Stall
  27. Stall
  28. Write #2
  29. Stall
  30. Stall
  31. Stall
  32. Stall
  33. Stall
  34. Stall
  35. Stall
  36. Stall
This isn't very efficient: the RAM sits around waiting. Even with "latency reduced" RAM, you can see that the RAM still isn't doing very much. In fact, this is why people thought Cryptonight was safe against ASICs.
But what if we instead ran four instances in parallel? That way, there is always data flowing.
  1. Cryptonight #1 Read
  2. Cryptonight #2 Read
  3. Cryptonight #3 Read
  4. Cryptonight #4 Read
  5. Stall
  6. Stall
  7. Stall
  8. Stall
  9. Stall
  10. Cryptonight #1 Write
  11. Cryptonight #2 Write
  12. Cryptonight #3 Write
  13. Cryptonight #4 Write
  14. Stall
  15. Stall
  16. Stall
  17. Stall
  18. Stall
  19. Cryptonight #1 Read #2
  20. Cryptonight #2 Read #2
  21. Cryptonight #3 Read #2
  22. Cryptonight #4 Read #2
  23. Stall
  24. Stall
  25. Stall
  26. Stall
  27. Stall
  28. Cryptonight #1 Write #2
  29. Cryptonight #2 Write #2
  30. Cryptonight #3 Write #2
  31. Cryptonight #4 Write #2
  32. Stall
  33. Stall
  34. Stall
  35. Stall
  36. Stall
Notice: we're doing 4x the Cryptonight in the same amount of time. Now imagine if the stalls were COMPLETELY gone. DDR4 CANNOT do this. And that's why most people thought ASICs were impossible for Cryptonight.
Unfortunately, RLDRAM3 and QDR-IV can accomplish this kind of pipelining. In fact, that's what they were designed for.

RLDRAM3

As good as QDR-IV RAM is, its way too expensive. RLDRAM3 is almost as fast, but is way more complicated to use and describe. Due to the lower cost of RLDRAM3 however, I'd assume any ASIC for CryptoNight would use RLDRAM3 instead of the simpler QDR-IV. RLDRAM3 32Mbit x36 bits costs $180 at quantities == 1, and would support up to 64-Parallel Cryptonight instances (In contrast, a $800 AMD 1950x Threadripper supports 16 at the best).
Such a design would basically operate at the maximum speed of RLDRAM3. In the case of x36-bit bus and 2133MT/s, we're talking about 2133 / (Burst Length4 x 4 read/writes x 524288 inner loop) == 254 Full Cryptonight Hashes per Second.
254 Hashes per second sounds low, and it is. But we're talking about literally a two-chip design here. 1-chip for RAM, 1-chip for the ASIC/AES stuff. Such a design would consume no more than 5 Watts.
If you were to replicate the ~5W design 60-times, you'd get 15240 Hash/second at 300 Watts.

RLDRAM2

Depending on cost calculations, going cheaper and "making more" might be a better idea. RLDRAM2 is widely available at only $32 per chip at 800 MT/s.
Such a design would theoretically support 800 / 4x4x524288 == 95 Cryptonight Hashes per second.
The scary part: The RLDRAM2 chip there only uses 1W of power. Together, you get 5 Watts again as a reasonable power-estimate. x60 would be 5700 Hashes/second at 300 Watts.
Here's Micron's whitepaper on RLDRAM2: https://www.micron.com/~/media/documents/products/technical-note/dram/tn4902.pdf . RLDRAM3 is the same but denser, faster, and more power efficient.

Hybrid Cube Memory

Hybrid Cube Memory is "stacked RAM" designed for low latency. As far as I can tell, Hybrid Cube memory allows an insane amount of parallelism and pipelining. It'd be the future of an ASIC Cryptonight design. The existence of Hybrid Cube Memory is more about "Generation 2" or later. In effect, it demonstrates that future designs can be lower-power and give higher-speed.

Realistic ASIC Sketch: RLDRAM3 + Parallel Processing

The overall board design would be the ASIC, which would be a simple pipelined AES ASIC that talks with RLDRAM3 ($180) or RLDRAM2 ($30).
Its hard for me to estimate an ASIC's cost without the right tools or design. But a multi-project wafer like MOSIS offers "cheap" access to 14nm and 22nm nodes. Rumor is that this is roughly $100k per run for ~40 dies, suitable for research-and-development. Mass production would require further investments, but mass production at the ~65nm node is rumored to be in the single-digit $$millions or maybe even just 6-figures or so.
So realistically speaking: it'd take ~$10 Million investment + a talented engineer (or team of engineers) who are familiar with RLDRAM3, PCIe 3.0, ASIC design, AES, and Cryptonight to build an ASIC.

TL;DR:

submitted by dragontamer5788 to Monero [link] [comments]

A look into the future regarding Decentralization,ASIC resistance and Vertcoin and other crypto currency (Long Post)

Warning: this post is lengthy because it includes details to understand the current development of Crypto and ASIC resistant Cryptos.
I. Decentralization is the fundamental assumption in the block chain security model:
I am glad that the recent Vertcoin price hike have brought more people to the awareness of crypto-currency decentralization. As decentralization is an assumption in satoshi's white paper, and hence the fundamental aspect in block-chain's security model. It appears that the block-chain security model is not complete. As you can see, there is an obvious concentration of computing power appears in bitcoin where one or two ASICs manufactures are controlling more than 51% of the network hash power. In satoshi's white paper, the assumption of 1 CPU,1 vote, does not hold indefinitely. Just 5-6 years after the inception of blockchain, we appear to have such machine based on ASIC, and the phenomenon of 1 ASIC, 1*103 or more votes, and the magnitude is only seem to be increasing.
Centralization defeats the entire security model of any crypto-currency based on block-chain and its variant. As of the time of the writing the bitcoin network and its public ledger's survival is not based on its invulnerability to rewrite, but based on the fact that the ASIC computing powers that secure the network currently lacks incentive to destroy it. When such incentive arrives the result can be catastrophic. As whoever controls the 51% hash power control the power to modify the block chain. In the Segwit 1 fork, there is worry that the bitcoin chain can not survive. (reference this article for a variety of possibility during a fork where miner controls the majority of hash power: https://medium.com/@jimmysong/uasf-bip148-scenarios-and-game-theory-9530336d953e ). In segwit 2X fork, some miners wants to make their own copy of of the chain, and in the process destroy the original chain. This upcoming fork is much more threatening than every single bitcoin fork comes before it.
II. CPU/GPU vs FPGA vs ASIC - you must understand the differences to understand the ASIC resistance movement
The decentralization problem is not fully solved yet. the crypto community and its developers are left to fill in the question.
As you can see the current approach is to make hashing algorithm to be hard to realize in ASICs. To fully discuss this approach, we must look at the currently available computing hardware architectures. the list go like this:
(CPU and GPU)->FPGA->ASICs.
The list go from the most general purpose,flexible computing hardware to the least flexible, and specific task computing hardware.
The list also go from the worst raw performance(you can say hash power for crypto) to the best raw performance, given a specific task.
CPU, and to a extend GPU are general purposed hardware that can be programmed to perform all tasks, while ASIC(Application Specific Integrated Circuits) can only perform a specific task. FPGA(Field Programmable Gate Arrays) - sits somewhere in the middle, it can be reprogram to perform a specific task better than CPUs and GPUs but the performance and durability is worse than ASIC.
In therms of computing speed,optimization and hence raw performance on a specific task, the list goes in reverse, this is because hashing algorithms and its calculation can be optimize thru parallelism(I have 10 workers to do 1 task 10 times quicker) and pipe-lining (think factory production pipeline with sequential work stations). CPU and General-Purpose GPUs in our computers exploit parallalism and pipe-lining to a degree, But because they are general hardware, the exploitation is limited because they must accommodate all types of possible computation. ASICs, are develop to only accommodate the required computation in a task, and exploit parallelism and pipe-lining to the extreme, this gives rise to ASICs such as AntMiners, where the performance is more than 3 magnitudes better than CPU and GPU.
III. ASIC resistance, and the movement to keep the crypto decentralize
The ultimate goal of alt-coin development is to fill in the void of satoshi's block-chain security model. The void is , How to keep the network decentralized in terms of hashrate/s?
The obvious answer, the first approach, would be to let the most abundant hardware to perform as well as the least abundant hardware. Thus, make an hashing algorithm so that either a CPU can perform as well as ASICs, or make an algorithm so that it is very very hard(cost prohibited) to develop ASICs for.
It appears that this approach is the most successful at the moment, some memory hard algorithms such as Vertcoin's very own Lyra2REv2 has no ASICs currently available.
But on the longer time frame, the profit driven development of ASICs is a definite trend, ASIC resistance is a constant Spear vs Shield game. Being ASIC resistance is not necessarily equivalent to being decentralized.
There are several ramification of being ASIC resistant. First the algorithm is necessarily more complex and cost more electricity on CPU/GPU to perform. Secondly, Developing ASIC for algorithm such as Lyra2REv2 is hard. Because of this hardness, there are fewer people who can develop this than the amount of people who can develop SHA256*bitcoin ASICs. Maybe in the not too distance future bitmain's monopoly over SHA256 ASICs would end and more of us can purchase a bitcoin ASIC, thus the bitcoin network becomes decentralized again. But because it is harder to develop Lyra2REv2 ASICs, once developed the ASIC monopoly can remain for a very long time enough to destroy the network. Because fewer people can do it, it will be more centralized once developed.
This does not mean that Vertcoin's security model is not good. In fact it is very promising. First the hardness to develop Lyra2REv2 ASIC can be to the point of such extrem that no one is able to figure out over an very long period of time. Second, once developed, the devs promise to hard fork the network again with a new algorithm in their tool bag. because the tool bag is unknown, the ASIC development cycle repeats, possibility over a long time.
So the Vertcoin's hashing algo Lyra2REv2 is among the best of all crypto. combining with the fact that a promised evolution of hashing algo once ASIC appear, I dare to say that the security/decentralization model is the best in crypto.
IV. Further discussion regarding ASICs and Network decentralization and security. paradigm switch regarding ASICs
It is in the profit driven nature that an ASIC would apear,Bitcoin already fell, for a memory hard algo, Scrypt and Scrypt-N is thought to be resistant enough, but ASIC appear, thus LiteCoin and The old Vertcoin falls. Vertcoin later forked and adapt to Lyra2 , and sub sequently Lyra2REv2 and remain the most secure coin.
For the ones used by GroestleCoin(Groestl), Decred(Blake256), SteinCoin(Stein256) , although there is no ASICs, but over an infinite horizon, the ASIC will appear this coins can all flop over night, if they do not adapt to the changes , Like what Vertcoin can do.
I think in the infinitely long term, there are 2 solution.
1st the same as Vertcoin, Keep ASICs out, and keep evolving the unknown puzzle bag for replacement if ASICs appear.
2nd, Amend the algorithm so that the theoretical upper bound in the speed up from ASIC is low. This requires making most calculations sequential and none-associative, with a slow bottle neck. thus parallal and pipe-lining machine can not take too much advantage. After that make ASIC development an open source, community movement, so that the entire community is guarantee to enjoy the advancement in ASICs. This would guarantee that the advantage from a new novel asic is small compare to what the community have, and limit the degree of concentration of hash power. ASIC can also benefit the network by reducing power consumption and increase transaction speed.
V. Conclusion
The current security model of Bitcoin is flawed and Vertcoin's solution is the current best at tackling the security concern. The promise of evolution of Vertcoin's Lyra2REv2 can be a viable long term solution to the Spear vs Shield game of ASICs. Nonetheless, I think we are making good progress of filling the void. I hope the future decentralization solution of Vertcoin can evolve past the paradigm of strictly ASIC resistance, and considering community driven and fair distribution of ASICs. I hope everyone in crypto can participate in this discussion.
Disclosure: I hold Vertcoin, 100% of my porfolio :).
submitted by bntyjx to vertcoin [link] [comments]

I believe it's considered a scam when...

If there was a slot machine at a casino that promised "99% payout" and when it was audited it was determined it only paid 98%, and the casino had set it like that on purpose, would you consider that a scam?
If you were led to believe a new business venture you were going into with a partner had a 20% chance of you making 10x your money, and an 80% chance of losing all your money, and your business partner had a 20% chance of losing all their money and an 80% chance of making 2.5x their money, and it turned out your business partner had knowingly reversed the odds, would you consider that a scam?
I believe most people would generally consider all of these scenarios a scam.
If a party offering an opportunity knowingly misrepresents the risk and/or reward, they are scamming you.
So.. is butterfly labs a "scam"? Even if they ship ASICs in the next few days/weeks/months?
I say: Absolutely.
Since last June they have been offering this value proposition to potential customers: give us your money now, and within a relatively short time you will receive a device that will pay for itself in two months and continue generating immense returns.
They been aggressively pushing (banner ads everywhere, the booth at CES) their investment opportunity (nobody buys an FPGA/ASIC for any reason other than to make money), and they market it as relatively low risk with a relatively high reward.
In reality though, it's been a very HIGH risk investment (due to the opportunity cost of tying up your bitcoins for approaching one year), and low reward (due to them not being first to market and the zero-sum game mining is in general).
But perhaps they acted in good faith, and simply bit off more than they could chew.
I would say that's not very likely. Let's compare them to Avalon.
Avalon took a limited number of pre-orders, just enough to get the funding necessary to get the first batch made. He did no real advertising and worked hard to get the machines actually made and started shipping before he accepted preorders for another, slightly larger batch.
BFL on the other hand has poured a ton of the money they got from taking unlimited pre-orders into advertising, so that they could get even MORE pre-orders.. estimates are that they've now literally taken tens of thousands of them... tens of millions of dollars worth!
When they finally do ship, and I believe they will eventually, those who ordered will find themselves with a machine that uses more electricity than promised, mines fewer coins than promised, started mining coins later than promised, and cost orders of magnitude (due to spending bitcoins when they were ~$10 vs ~$100) more than promised.
Scam!
submitted by zhoujianfu to Bitcoin [link] [comments]

FPGA Mining

A FPGA opensource miner has just been released running at 80Mhps but at a cost of $585. The efficiency is stated below quoted from a post in the thread.
At 80 MHps, I will need at least 3 of these to achieve a single 5830 hashrate. That is $595.-x 3 = $1785.- at full price, vs. $190.- for the 5830.
Giving the 5830 is consuming $11.- a month in electricity, and assuming this board will consume zero electricity, it will take more than 145 months, or 12 years to recover the investment, always comparing to a 5830.
BUT:
In this thread, someone mentioned he is doing 210Mhash/sec after some optimization but he will cease public posting of his development.
Apologies but no more development information will be posted. I've been offered a 25% share from someone that owns 2 FPGA clusters. If you haven't seen that type of hardware before think a 156 FPGAs per machine.
From those posts what we can understand is that the factors that affect FPGA now are high procurement cost, low running cost and ease of scalability . What this means is that with the increasing total hash rate of the network (30Ghash/day last difficultly adjustment) the question becomes when would the difficulty render GPU inefficient in contrast to running cost?
Remember to take into account FPGAs are usually run in clusters and even though it would not be beneficial to buy one outright, those who have access to FPGA are the first movers and eventual dominant forces of the mining market.
Of course, in the end, ASIC is where it's at. Anyone? =D
Edit: read more stuff, added info.
submitted by Coz131 to Bitcoin [link] [comments]

Are YOU Expanding In 2020? GPU, ASIC or FPGA Mining? Bitcoin mining 4 DE2-115 Altera cyclone 4 FPGA The Outlook on Cryptocurrency Mining - GPU vs ASIC vs FPGA ... FPGA based Bitcoin Miner What is the Difference Between an FPGA and an ASIC - (Part 2, Ch 1)

These figures are staggering, but Bitmain’s monopoly of the Bitcoin ASIC market may come to an end, following the release of Halong Mining’s DragonMint 16T ASIC. About the DragonMint 16T Designed with brand new technology and boasting 16 TH/s per miner, the 16T is the most powerful and efficient Bitcoin miner to date. ASIC Miner vs GPU Miner Before working on a particular mining project, make sure that what you are aiming for and the legitimacy of the energy you use are worth your time. Search the Internet, view the market size, and access online communities associated with it. CoinTerra’s TerraMiner IV is a Bitcoin earnings machine with the following features: ASIC-equipped water cooling. Delivered with 4U add-on with power supply. Equipped with 4 ASIC chips, each of which delivers up to 500GH/s, which means that the total power of TerraMiner IV is 2TH/s. At the moment, it is quite difficult to earn Bitcoin, and ... Then they started to use graphic processors and FPGA (field-programmable gate array), but soon they also ceased to be quite useful. ASIC miner is the best bitcoin miner of the latest generation. It provides a higher cryptocurrency mining speed, the machine heats less and consumes less electricity. FPGA vs. GPU vs. ASIC ... FPGA (Field Programmable Gate Array) ... Bitcoin: A Total Turing Machine. Craig Wright (Bitcoin SV is Bitcoin.) M-Pesa: A Case Study in Financial Inclusion.

[index] [32883] [33770] [30615] [4109] [5346] [26854] [7948] [6719] [22600] [1252]

Are YOU Expanding In 2020? GPU, ASIC or FPGA Mining?

FPGA Miner for Cryptocurrency Mining: Why Use FPGA for Mining? FPGA vs GPU vs ASIC Explained - Duration: 7:56. FPGA Guide 626 views My Favorite Book - "The Bitcoin Standard": https://geni.us/NH5spNd Email - [email protected] ... Why Use FPGA for Mining? FPGA vs GPU vs ASIC Explained - Duration: 7:56. FPGA Guide 1,123 ... What are the TOP 5 coins to mine for best profitability in 2019 with ASIC miners?! Today we will review Bitcoin, Ethereum, Dash, Litecoin, Zcash, and Grin mi... FPGA Miner for Cryptocurrency Mining: Why Use FPGA for Mining? FPGA vs GPU vs ASIC Explained - Duration: 7:56. FPGA Guide 666 views VoskCoin livestream on the Outlook on Cryptocurrency Mining - GPU vs ASIC vs FPGA with Q&A. Text version of todays video - http://bit.ly/2LaZA5R -- The lands...

#