Wait, I don’t get it, isn’t that inherently dishonest by NVidia? They promise a “feature” (driver-based hash rate limiter) that they then don’t implement? And are the CMP cards that yet have to be released pointless now?
What am I missing? I have a hard time believing the whole thing was a fraud to begin with.
My assumption is that the release of a new driver without the announced mining brake happened by mistake (note that it is a 470-family driver, i.e. different branch than the 460-family drivers currently in use).
Correct and you won’t find it available in any Nvidia repo.:
If I didn’t know that NVidia has done a bad job in releasing drivers before (at least Linux), I’d totally say: this was intentional! It’s just too convenient to basically say: oh, we just forgot! We just made a mistake.
Or how about with dummy plugs:
In general I think any type of “downgrading” or “nerfing” or “disabling” is always a bad idea, at least in software – I’m not just talking about NVidia. But also NVidia. For years they were never supporting the current release of Fedora, they were always some two version numbers behind. But now things broke all the time, because the current Fedora always ships with a gcc version that couldn’t be used with CUDA. Now people either a) fudged the version number in a config file, and that was pretty unsafe, but in most cases still worked well, b) needed to several serious contortions to their installations to install an old gcc (hacks, manual installs, copra repos, asf.). I complained about this on various online forums and usually got replies along the lines of “well, it’s difficult with the many distros …”, well, generally: yes, but you are NVidia. A global company, billions in sales, and can’t afford Linux release managers that can cover all distros. They make it sound as if they were some poor mom and pop shop. I admit things got better in the more recent past, about the time the HPC package was released. But for years they didn’t care about fixing things: not supporting current gcc, or even fixing bugs in the CUDA samples. I’ve reported a bug here in this forum, and it’s still not fixed in the current CUDA release. It seems nobody from NVidia is reading these forums. VERY negligent in every way.
So, if this is really compay-internal dysfunctionality obsession (as it manifests otherwise), this may have been a “legit” case of “oh, we forgot”. Ironically, the “innocent” version of human error is actually credible here, because it’s a) driver release at NVidia, b) Nvidia as company. If it weren’t a driver release at NVidia I’d call it a ruse in the disguise of a human error.
Which means: what does that mean for the CMP cards? It seems the 470 drivers are easily available on the web, permanently. Now, what’s the point of the CMP cards, especially if they’re gonna be expensive?
So… Can HPC/ML run on the HX GPUs or not?
Are they simply GeForce with no display out leading to better compact cooling?
If so, than the answer would be yes, making a decent secondary market for them once they are done mining?
- Ron
Give it a try and let us know :-) Given that mining software is based on CUDA, I would expect any CUDA software to run, but that is pure speculation on my part. I am the kind of user who has been buying systems based on Xeon CPUs and Quadro GPUs for twenty years now, so no hands-on experience with CMP hardware.
The first product, CMP 30HX, supposedly just shipped. It is said to be based on a TU116 chip, with performance comparable to a GeForce GTX 1660 Super. First reports mention a retail price of around $700. Given that the GeForce GTX 1660 Super at one time retailed for something like $250 (I think, corrections welcome), this doesn’t look exactly like a steal to me …
update: now they’re trying it on the 3060 with new silicon:
My personal opinion: I think the goal with these crippling attempts is already wrong. “You are allowed to use your card for gaming but not for mining”. Who says that a graphics card may only be used for gaming? People buy things because of the entirety of its feature set, and its usefulness. And I’m sure there are gamers out there that are accepting higher prices for higher features because they know they can recoup some of the money with mining – what would be wrong with that thinking?
To me that sounds very much like big-guvnmint type of mindset that wants to command upon us what we can use stuff for and for what we can not use it, trying to “reason” that some people do “wrong” things with it (and again, I see nothing wrong with using GPUs for mining). Like banning hypodermic syringes for private households, because there are some people who use them to inject illegal drugs. I defy all that and use hypos to refill my ink cartridges with ink from bulk ink bottles that I bought – MUCH cheaper than buying all those ink cartridges all the time – overpriced, and “less trash” should be a popular belief these days, so I also make Greta a bit happy with that. And that usage wasn’t the intention of the syringe manufacturer, and hypos in private hands are illegal in most states, but I don’t care. Just because some people inject drugs with hypos doesn’t stop me from using them to refill my ink cartridges pretty cheaply. So I don’t think NVidia should be allowed to tell us what we can use the 3060 for and what not to use it for. The intention is already wrong.
Within legal limitations, companies are free to change the features of their products however they like. Consumers get to vote on changed feature sets with their wallets. This seems like a storm in a teacup to me.
However, if the reported rumor is true, there is a worrisome implication. None of the cryptocurrency mining nonsense would have measurable impact on general availability of GPUs if production could be increased to keep up with demand. Additional effort put into creating a mining brake could reasonably be speculated to be due to a forecast that chip shortages are not going away by this summer, as was (to my knowledge) the original semiconductor industry sentiment, but are here to stay for longer.
This would jibe with a news item I came across the other day that stated that the production of some of the machinery needed to increase global chip production is hampered by a shortage of certain microcontrollers (embedded CPUs). Quite ironic.
… and the Samsung chip production plant in TX was delayed because of the TX storm / power outage.
Yes, it will take a while until the factories are even ready for production. The Samsung chip plant in Austin is supposed to cost 17 billion. So that won’t happen quickly. And my guess is that the chip production shortages are the reason that the newly-announced Grace server CPUs won’t ship until 2022. So far Grace is all talk. And NVidia would be free to change all of that or cancel or improve or do a totally different server CPU. Somewhat reminds me of the Rock announcement by Jonathan Schwartz. He made a big fuss on some conference about holding one physically in his hands … but then it was canceled shortly after the take-over.
And would that also mean that new CUDA releases will get a slow-down? 11.3 is now Ampere-ready, and with Lovelace not happening in 2021, there seems to be no driver for CUDA improvements at least from the hardware side. That would give room though for improvements on existing CUDA code – I guess. It’s not that I know that, it’s a hopeful question. Like fixing the bugs in the Eclipse plugins. It seems the Eclipse developer tools are always treated like an unwanted step-child. Some bugs never get fixed – although I admit that Fedora support has vastly improved ever since the HPC suite came out. Some things suddenly worked out of the box immediately after install.
Given the overall situation in the semiconductor industry, I expect the impact of a GPU mining brake on GPU supply to regular users to be minimal. From news reports this past week:
[…] Colette Kress, executive vice president and chief financial officer of NVIDIA. […] “Overall demand remains very strong and continues to exceed supply while our channel inventories remain quite lean. We expect demand to continue to exceed supply for much of this year. […]”
Semiconductor companies can take some short-term steps to alleviate some of the pain, Pat Gelsinger said in an interview, adding that Intel aims to boost production of automotive chips within six to nine months. But a full solution to the problem will take much longer, he said. “We do believe we have the ability to help,” said Gelsinger, who recently took over as CEO of the United States’ largest-by-revenue semiconductor company. But “I think this is a couple of years until you are totally able to address it,” he said.
TSMC CEO C.C. Wei offered a similarly dire estimate to investors on Thursday, saying that the Taiwan-based company hoped to “offer more capacity” for meeting retail and manufacturing demand “in 2023.”
On the chip shortage and a follow-up on my own last comment: with the ARM deal now at only 10% (and further hurdles from other European countries and China), which I think will make it a “no”, I think this can only further slow down chip production for NVidia, and I’m sure it will have an impact on Grace, and conceivably on CMP. Without the deal, NVidia would pay 2b for a 25b total addressable market (TAM), a 10:1 benefit, with 38B saved. With the deal NVidia would spend 40b to capture a 125b TAM, which is only a 3:1 benefit. It’s hard to forecast the repercussions of NVidia not getting ARM, but for sure it can’t accelerate anything. At best, there is no production speed impact for NVidia. Just my 3.6e-7 bitcoin.