Thanks, asic_king. Slide 10 on MRAM opportunity looks interesting. According to Wikipedia, MRAM is denser than SRAM while not much slower. Would that be useful in fitting C32 into 16 nm? Or is this type of memory still too esoteric / immature?
It looks like MRAM is currently only supported in 28nm or higher nodes, not at any 16nm node. Also, it looks like MRAM, while fast, is still several times slower than SRAM.
It is interesting that you picked out MRAM because there is a recent wave of announcement by Samsung and Intel about the availability of MRAM option on 22/28 nm nodes. (https://www.eetimes.com/document.asp?doc_id=1334410)
MRAM is being embraced by the AI community. One example is Gyrfalcon Technology Inc. who built two versions of neural accelerator chips, one with SRAM and one with MRAM using TSMC process. (https://www.eenewsanalog.com/news/tsmc-embedded-mram-key-gyrfalcon-ai-chip) A close friend of mine who is a memory expert worked on this project. One problem with MRAM is it has a slightly higher bit error rate due to things like stray magnetic fields. AI would be a good candidate for MRAM because CNN has a certain amount of tolerance for error. Here is a link that describes MRAM in more details for the hard core enthusiast (https://nepp.nasa.gov/files/24256/12_124_JPL_Heidecker_MRAM%20Technology%20Status%20jpl%20pub%2013_3%202_13%20rec%204_15_13.pdf)
My view is that it is just a matter of time that MRAM will become available on 16nm node and beyond. If I have to guess it might take one year to two years.
So does Cuckatoo Cycle. Consider all possible bit flips during edge trimming, where N=2^32 is the number of edges in C32:
edge bit flips from alive to dead: results in loss of expected number of 42-cycles going through this edge; which is negligible at ~ 1/N.
edge bit flips from dead to alive: may result in finding fake 42-cycle, with probability 1/N, but this is trivially excluded on verification
node bit flips from from 0 to 1: might delay the trimming of a leaf edge by 2 rounds, which is totally harmless.
node bit flips from 1 to 0: might kill a connected edge, same effect as in first case of edge bit flipping to dead.
Even if the total fraction of bit flips during one graph is 1% of N (which is HUGE), the fidelity should still be over 79%.
Stuck bits (that are always 0 or always 1) are more of a problem, behaving like repeated bit flips at particular locations, but even something like 0.1% of stuck bits should be tolerable, unless the trimming is required to eliminate all but say, 0,01% of edges.
Even if the total fraction of bit flips during one graph is 1% of N (which is HUGE)
Is it though?
I can totally see an asic that does a few rounds of removing all nodes with connection counts under a 2^int; all super waterfalled with no control logic before passing off the ugly but much smaller graph to proper logic
EDRAM couldnāt be made on 16nm/10nm chip directly. TSMC only provided EDRAM on 90nm~40nm.
https://www.tsmc.com/english/dedicatedFoundry/technology/edram.htm
A common way for integrating EDRAM with CPU/ASIC is MCP (multi-chip package).
P_semi, welcome to the board. We see that you joined about 24 hours ago. What happened to the post on the TSMC 7nm yields? I believe it was the N7 yields that you posted. Which is very out of date. If the Moderators removed it, thank-you for adhering to a code of conduct and setting expectations that this community does not want a reputation of posting confidential information. It can spoil it for every other honest manufacturer. It is also our understanding TSMC has already moved on to N7+. Attached is an article from almost 12 months ago, that mentions how the N7+ SRAM yields are performing better than the N7. This is public domain so it adheres to the governance that we abide by. https://www.anandtech.com/show/12677/tsmc-kicks-off-volume-production-of-7nm-chips
Potentially relevant in the context of this discussion: Slean mining - mining AT3x using low memory GPUs
1 Like
Dear everyone ā
Iām a lawyer/miner and I provided $20k advance payment to Obelisk for 10 DCR-1s in April 2018. Obelisk never delivered these devices (itās been a year!). I requested a refund within 2 weeks after paying and then I took them to small claims court. Obelisk hired Cooley (one of the most expensive firms in the states) to defend their contract of adhesion. Obelisk won.
Itās been 11 months since I paid for the devices, 10.5 months since I requested a refund ā not only do I have no equipment, but Obelisk won in small claims so I have no recourse.
Iām going to be taking this on as a significant consumer affairs issue. I want to expose David Vorick for his scammery and under no circumstances should a company like Obelisk be entrusted for any pre-sale. The company has stolen from me and has ripped off several others.
Hereās my Medium article from last year. Thereās more public materials coming soon (e.g. a recording of what the attorneys said in court).
Obelisk is a problem. They are a bad actor. Stay away. Iām coming for them.
1 Like
Obelisk has shipped the majority of the DCR1 units, with the final units set to go out in the next two weeks. If you havenāt received yours yet, itās because you are in the final group. Weāve put forward a compensation plan for all of our customers to make up for the fact that we shipped late. The compensation plan is that we will be giving all customers credit equivalent to the amount of cryptocurrency that they would have mined (assuming free electricity and no other expenses) had we shipped on time. Weāve done everything we can to make the situation right.
You lost in small claims court because Obelisk has been fair to its customers in this situation. Weāre as conscious of our obligations to our customers as we are of our obligations to the communities that we build hardware for. Some unfortunate things happened with the later batch SC1s and DCR1s, and weāve taken big steps to make things right. Cooley has been Obeliskās primary council since the inception of the company, they were not hired merely to handle your court case. They handle all of our legal, and they always have.
We havenāt yet released the details to the public, but the major shipping hurdle that we hit was due to a big, unexpected mistake on the part of our chip design team (who is not involved with the GRN1 and will never work with Obelisk again). For the SC1 and DCR1, our chip team maintained full control of the relationship with TSMC. We were never allowed to interact with the chip manufacturer or review orders that were going out to the foundry, despite our best efforts to get involved and have more visibility into the manufacturing process.
About 6 months ago we received the final shipment of chips from TSMC. We had full confidence in the shipment because we had already received working parts from TSMC. We knew that the chips worked and that there was nothing to worry about.
Our chip design team had placed an order for the wrong parts. We got back broken chips because our chip team had put the wrong serial number down on the order form. We took steps to make it right, but there is nothing you can do to get around bad parts. We placed a new order as fast as we could, including taking direct control of the relationship with TSMC. But the fastest that we could get new parts was 6 months after receiving the bad parts.
Weāve already blogged about it, but following the SC1 and DCR1 we made massive improvements to our manufacturing process. We are now working with more transparent teams to design and build our rigs, we have direct relationships with all of our manufacturers and suppliers (including TSMC), and we have many more steps of diligence and double-checking that we have added to our processes to ensure high quality, low mistake, fast turn-around machines for our customers.
One major step we took this time around with our chip team is that we brought in auditors. I have already posted a link to the audit earlier in this thread, but we paid a third party auditing team to come in and verify that our current chip team (ePIC Blockchain) was using best practices and representing truthfully their capabilities. This is just one step of many that we have taken, but the audit was highly encouraging and demonstrated to us that our new process was working well.
David, thank you for the response. I also received a āshipping noticeā today from UPS that my DCR-1s are shipped and expected to come on April 1. The problem with that is this:
- The units donāt perform as you said they would when they were pre-sold;
- You said youād ship them by 9/30, and at April 1, thatās 6 months late.
Additionally, your attorneys made representations in court about Obelisk as a crowdfunding platform, something thatās completely wrong and misleading to consumers.
Itās not too late to do the right thing. You can (and should) refund me and any other Obelisk consumer thatās similarly situated. You should also adopt a practice for future purchasers (e.g., of GRN miners) to allow for refunds if you donāt deliver what you promise.
You canāt just ship these to me now and say itās ok. The units arenāt as you described, theyāre late, and thereās also the obvious point that I donāt have a data center at the house that you shipped it to and I lost additional money from having reserved and paid for data center space in anticipation of September.
My issue with you and your company does not end with Small Claims. In fact, having lost in Small Claims Iām taking this to a much broader consumer complaint platform. Weāll get to know each other pretty well.
For the community, hereās my court papers. I also have recordings of the attorneys and am transcribing those for everyoneās benefit.
Iām starting to speak publicly about my interest in Blockchain ā one of your attorneys laughed in court when I said my interest was in governance. Itās not a joke. [hereās recent vid ā pls forgive the family nature my daughter did this). In the future this issue is going to be at the center of my policy work in blockchain. The way youāre mishandling and misleading consumers and beating them up with contract is a governance issue for blockchain and one that must be addressed.
You should really stop and figure this out. I am not going away and there are many other people that are similarly upset. Thereās a class action. There are these consumer complaints. Lots more to come. All follow the same theme of your company promising one thing and failing to deliver it.
The blockchain community needs the private sector to participate with us to develop solutions. You should be lauded for doing that but shamed for not living up to your promises. The good news is that you can change your policy.
Grin community ā this is the misrepresentation that worries me and why David Vorick should not be entrusted to make any statements about his future intentions with these rigs. Vorick makes a āmea culpaā admission here in the forum because he wants Grinās trust:
Compare: the position above about refusing a refund and not shipping for six months. You say here you were fair. But you told us earlier, repeatedly, that you would give us refunds. By any objective standard your statements here are disingenuous and cannot be reconciled.
āThe US has very strict consumer protection laws, if we are late in delivering the units, customers have the right to request a full refund, and the law will enforce their ability to receive that refund.ā ā David Vorick (in a forum chat), June 2017
"You can legally demand a refund if we produce units that do not perform according to the specifications we promise in the presale, at least if you are a US citizen consumer protection laws are very real." -- David Vorick (forum chat), November 2017
āThe FTC has very harsh rules. If you miss performance targets, you have to give everyone refundsā ā David Vorick (forum chat) Jan 2018
then . . .
"Itās no secret that Obelisk does not have enough money to refund all customers. We are not usre [sic] how many refund requests we will get, however we are quite confident it will be beyond our financial means to provide refunds to everyone who request. We will figure out how to provide refunds after we know the total number of refunds that must be issued and after we know how much money remains after all units have been built." ā David Vorick (forum chat) Aug 2018
Source: screenshots of above available here.
In short, you stated repeatedly that we could order in confidence and that you would return the money if things didnāt go well. Here youāve now admitted the reason why things didnāt go well, that those reasons were under your control ā and you reversed the prior promise you made to provide refunds.
These are promises you made, and you, sir, are not a man of your word.
It was very clear during the fundraising for the SC1 and DCR1 that the money we collected was going to be used for development and NRE in addition to manufacturing. It was also clear in our terms of service that all specifications and shipping dates were estimates, and all purchasers were required to consent to the terms of service before they were allowed to place an order.
We provided a compensation plan which pays for the full revenue which the miners would have received had they been mining. You canāt claim to have lost additional money for having reserved data center space because you have a compensation plan that is equal to all of the revenue that your miners would have made had that datacenter space been put to use.
Thatās not at all how the quotes you posted read. The quotes reference what my understanding was at that point of the law. Hereās another quote from me from the same thread:
āSame thing would happen to us as happens to any company that runs out of money - weād go bankrupt.ā
At no point did we assert that the money would be available to provide refunds.
Weāve done our absolute best to make the situation right. Including providing full compensation for every bit of mining that you missed out on due to not having your unit. That is above and beyond what any other manufacturer in this space has done.
David, I wonāt respond to your points here because I realize that youāre just defending your current position. Thereās nothing thatās changed in the law or in business from the time that you repeatedly reassured customers that youād give a refund and the point now where you have refused. Iāll let the community and the court of public opinion judge for themselves.
I must say that I am particularly grateful that you have engaged here because youāve provided additional context, including an admission that your error caused a 6 month delay. Youāve shifted this to the consumers and whatever compensation plan youāre referring to is not the refund that you had promised.
Itās early days in our dispute, and even though I lost Small Claims I gained a full record, transcript, and now, additional information for this next phase. There are many more quotes and representations to share. We know what the Boston Municipal Court says, weāll now have a chance to see what consumer agencies think about your practice.
One more comment: Itās ironic that weāre having this discussion on a thread that includes reps from two major ASIC providers, one from the USA (Obelisk) and another from China (Innosilicon). One of the motivating reasons for my purchase of the Obelisk units was frustration with delays from Innosilicon (just of a few days) and complications with customer service. Even with those frustrations, I can definitively conclude, based on my own experience, that the customer service experience and the technology from Innosilicon are far and above what Obelisk is offering. In retrospect, I can hardly believe that my concern about a few daysā delay in shipment from a Chinese manufacturer drove me to embrace a new, untested American company. That was a mistake that I wonāt make again.
[EDIT 3/28] Adding audio recording of small claims hearings (about 1 hr) available here.
Thank you for sharing this PSA. This topic is somewhat forked from the tech issue addressed in this thread, but it brings up a more practical point: track record on theory vs. execution. If GRN buyers end up like papacabeza, our community will be very unhappy.
Vorick has made beautiful, eloquent descriptions of technology and challenges, heās convincing to people. When he was selling pre-orders he reassured users about refunds during the time that he was locking down pre-orders (all those quotes and promises for refunds), but once the pre-orders were secured and the specs changed, he came out with all these new details on the changed specs, delays in shipments, and suddenly decided not to give refunds anymore. I donāt even know what to say.
I was planning on ordering and wonāt now,. Currently, Obelisk is advertising GRN-1 to have a final design in March with delivery in October. Their projections are incredulous (if measured by past promises ā all missed by months, some still unattended). The volume of litigation described could very well bring Obelisk down well before the GRNs are manufactured. If there were such a thing as a license for a presale, then Vorickās license for that should be revoked.
Anyway this information Vorickās technical proposal and his credibility for execution in new context.
We were quite up-front about everything. Had we gone the refund route, everyone would have received a fraction of what they paid, and Obelisk would have gone fully bankrupt. And this was apparent throughout the sale as well, it was well established that the money we were collecting would be used for research and development, and manufacturing, and it was also well established that we did not have separate money to fund the R&D and manufacturing in the event that something went wrong.
We were at that point ill prepared for the challenges associated with firmware development, with chip simulation and specification, and we had never built hardware before. That is all different this time around. You can see this blog post for more information about what weāve changed: Building a Better ASIC Company. How we continuously improve Obeliskāsā¦ | by Zach Herbert | Obelisk Blog | Medium
We are sorry that we did not make our original shipping date nor specification estimations. However, we provided a full compensation plan for the missed shipping date, which includes paying out IN FULL the revenue that you would have received had you received the machines on time. This is even better than receiving the machines on time, because it means that you donāt even have to pay for electricity, maintenance, or other operation expenses.
And we feel that this is fully fair towards our customers. Obelisk does not and has never had the ability to give out refunds for the original sale, nor was it ever represented that we would have enough money to give out refunds, so we did the next best thing and put together a compensation plan which goes above and beyond what any other manufacturer has done to compensate for a missed shipping date.
This is the equivalent of saying āgive us your money now, weāll build something for you, but when we change our mind on specs, or with some third party screw up the order and deliver an inferior product, or donāt deliver it all, since our contract says you take on all the risk, this situation is even better than receiving the machines on time because our low-quality, late product would never have worked right anyway.ā
The description youāre provided, and all the coupons, upgrades, bait & switch and new technical descriptions (while teetering on bankruptcy), are characteristic of an unstable business model, to say the least. Hereās my suggestion: instead of taking money from consumers, shafting them on product and telling them they should be happy for the shaft (thatās how I read what you said), how about taking a pause from any presales at all until you catch up with your promises. This includes issuing a full refund to any Grin prepurchase customers that have coupons for the devices in this thread.
Itās never too late to do the right thing.
1 Like
big upvotes to the developers