Posts by Fido 34 publicly visible posts • joined 29 Aug 2023
At least the new CEO is not talking about zetascale supercomputing next year. At the same time the world does not need any more low-performance high-core-count CPUs for running web servers. Instead it would be nice to have an affordable processor that could run a 500 billion parameter LLM locally at reasonable cost and speed.
To me it seems likely AI combined with traditional numerical computation will form a sum greater than the individual parts. For this reason I'd look forward to huge on-package matrix accelerators with unified HBM memory and selectable precision from 4 to 64 bits. If the promise is for Intel to listen, then one can dream that the voice of one user might be heard.
At the university here we repurposed the old scientific computing cluster as a cloud using VMware. The nodes each have a total of 32 cores so I'd naively expect a 72/32=2.25 increase in licensing costs.
Since we're also looking at a possible 10 percent general budget reduction due to lack of planning, it seems possible the solution will be to turn everything off but not back on again. For educational institutions Microsoft offers extended Windows 10 support at $1 per seat for the first year doubling each year after that. Will Broadcom offer similar discounts on legacy VMware for schools.
I'm not involved in the IT decisions, so my only problem is as a user of the cloud. Hopefully the discounts are worked out and operations will continue normally.
I'm interested only in the technology aspects of Signal, not the specifics of this case.
If a hostile foreign government were able to compromise the mobile phone networks of a targeted nation, would it then be possible to secretly install remote access software on a government issue phone?
To avoid giving away the fact that end-to-end encryption is compromised, a foreign spy might use the remote access software to add a third party to a private chat who is likely to leak the information. When the third party leaks the private information, the foreign government does not get blamed yet benefits from the leaked information.
I don't know how real spy operations work. However, it seems to me that this technique could be effective at circumventing end-to-end encryption without revealing who did it.
Somehow the disadvantages of having two distinct programming languages in a single code base has to be outweighed by one of them being Rust. Having to maintain separate bindings for each language and synchronise them by hand sounds more error prone to me than the danger C code in production for more than a decade possesses a hidden use after free, bounds error or null pointer dereferences. At the same time, I'd get very excited to see a nontrivial GPU, network or USB device driver written in Rust that avoids the unsafe keyword.
A year ago there was a US government directive to stop using C and switch to memory safe languages for essentially everything. While the directive looks as well thought out as the long since abandoned idea of using ADA for everything, it's possible adding Rust to the Linux kernel is based primarily on the need to demonstrate progress being made towards use of memory safe programming practices.
Such a benefit, though nontechnical in nature, could be enough to offset the interoperability and maintenance disadvantages of two languages in the same project.
The same considerations could be the main motivation behind both Microsoft's and Google's reported progress on the adoption of Rust.
My impression is the Altra, Altra Max and One processor designs are slow compared to recent Graviton, Intel, Epyc and likely the next generation Loongson processors. Although slow is not an ideal feature during the transition to compute intensive AI, the conjecture I've heard is Ampere has an engineering team with enough experience to quickly integrate a massive TPU into existing ARM designs.
I think at this point it's clear AI is not a fad that will fade like skinny jeans, but a tool which greatly increases productivity for workers in certain technical and creative fields.
When the dot-com bubble burst, online shopping didn't go away. Neither will the need for economical AI inference go away if the AI bubble breaks. Whether Ampere turns out to be a valuable acquisition for SoftBank is a different matter.
I think the latency--time between pressing the key and a program recieving a character--was greater with a USB keyboard. Since it makes no difference when playing Nethack or Dwarf Fortress, never mind.
I find it an interesting question whether a non-GNU licensed version of GNU core utilities would eventually become a closed-source proprietary component of Ubuntu Linux. I don't have any insight into Ubuntu's roadmap, but even if I did, the management can change or the company be sold.
It's been a few years since I experimented with Rust. Back then it did not have a shared library runtime. If that's still the case, each command in Rust core utilities might load hundreds of thousands of extra bytes. The resulting slowdown may be difficult to measure with micro-benchmarks; however, the large executable sizes might thrash both disk and CPU caches.
If I understand what end-to-end encryption is, there is no way to add a backdoor without an observable change to the software running on the user's device. As criminals and spy agencies constantly reverse engineer software updates to look for security related patches, actually adding a backdoor to UK phones would be a much more noticeable canary compared to making no changes in the code that implements the encryption.
From this point of view the least malicious form of compliance could be disabling the feature rather than spelling out in code exactly what was asked for.
My impression is people who buy Apple trend not to be as concerned about money compared to those who buy a Windows PC. Since saving money is not an objective, the Apple customer is not unhappy when older kit is obsoleted because that provides a reason to get the latest greatest.
Here the university has been purchasing new iPads, iPencils and portfolio keyboards for every incoming student since the epidemic. I tried to figure out how students could use this technology for programming tasks related to numerical computing and computer science. The Apple point of contact recommended Swift playgrounds.
Swift playgrounds features a 3D rendered version of Karol the Robot for toddlers but is actually much more: One can also write, compile and run real programs. Unfortunately the stack size is small, no optimiser is available and I could not find any scientific libraries.
Today it appears the main purpose of the iPad is to create a cyberpunk version of a medieval university where the students use iPencils to transcribe the text being lectured on as it is written in chalk on a blackboard. Traditions die hard.
On a modern note, the iPads do satisfy the digital equity objective of providing underprivileged students the same ability to TicToc, Facebook and Instagram instead of studying.
Since the original design of C was to eliminate as much assembly language as possible, if C were eliminated in favour of Rust, would that imply a need for more assembler in the kernel again?
If not, could the use of unsafe in Rust be more unsafe than writing the same routine in C for those OS things where unsafe is anyway necessary?
Deploying commercial single-vendor IT infrastructure forms a long-term business partnership with a vendor that is generally based on trust. The difficulty occurs when a takeover or buyout changes the nature of the partnership.
A similar thing happens at smaller scales with blog sites and free software. The owner sells to a third party that subsequently monetizes the readership or inserts adware.
The difficulty is that further regulations could make things worse. Consider, for example, when SoftBank went in to ARM-architecture monetisation mode after not being able to sell to Nvidia. All sorts of things can suddenly change the nature of a business partner.
Since IT embeds itself so deeply into a company's business process, relying on single-vendor software does not provide the robustness needed for long-term planning. While decision makers with golden parachutes may not worry much about the future, protection of the business from disruption caused by changes in software licencing and development is an argument for open or at least owning the source.
Said another way, switching from VMware to another commercial offering is a fool-me-twice decision.
Imagine if you suddenly had a return of 1,000,000 on every $ or £ deposited in your bank account. The conclusion: what separates the high flying business person, diplomat and government spy from Homer Simpson is less than a factor of a million.
Given the way cyber crime scales once the malware is written, it is not unreasonable to imagine millions could be automatically targeted by a known but unpatched vulnerability. From this point of view, it's not possible to disappear into the multitude. Since it takes more insignificant users for a significant effect, the insignificant might be more likely to be targeted.
Since the telecoms carrying SMS based second factor authentication were already compromised by foreign agents, I can see how FIDO2 helped T-Mobile avoid being as easily hacked.
Said another way, if your adversary can already read your text messages, then not only do they know the second factor but you have also told them when and where you are logging in.
My recollection is CrowdStrike pushed a configuration file--not even code--for their Falcon platform and put Delta Airlines out of business for a week. Others customers were affected for a total of over 5 billion direct losses.
Needless to say corporations who had opted out of automatic updates due to having their own mission critical review process for high availability weren't pleased that CrowdStrike pushed the update everywhere anyway.
As much as the security business seems overrun by clowns, if automatic updates are not supposed to happen then it's clear updates shouldn't happen automatically, even if the updates are good.
While the gaming community may be larger than engineers who develop and run numerical codes on their desktops, the time stepping in a scientific simulation is quite similar to in-game dynamics but without the game. In fact, since a numerical run has no graphics-card bottlenecks, the X3D series CPUs with full-width AVX512 are astonishingly fast for small-scale computations.
Woohoo! If more science and debugging can get done faster on the desktop, that leaves leadership-class supercomputing available for larger runs.
When I read the Tom's Hardware interview with Jack Huynh, AMD's senior vice president and general manager of the Computing and Graphics Business Group, I initially had the feeling that leadership was clueless and lacking.
After refection I decided trying to sell what the engineers made is better than promising zetascale compute
I think AMD unified the engineering behind Radeon video cards with the Instinct GPUs but badly presented this as deprioritising gaming.
Right now AMDs CUDA equivalent--ROCm--is not well supported on Radeon video cards. As games may soon include a nontrivial element of AI inference, getting ROCm to work on all AMD GPUs seems essential to me. Unfortunately, this unification seems to have led to the next generation Radeon cards not being competitive for high-end gaming.
Although a senior vice president can only sell what the engineers can make, I would have characterised the lack of a flagship gaming card as a necessary unification of two diverging GPU architecture so ROCm driven AI libraries can run everywhere.
The expense of training a model is likely to go down, just as a Raspberry Pi is faster and cheaper than the original Cray supercomputer.
One important security aspect of regular open source software is verifiable binary executables. For example this makes it possible to check that unexpected changes weren't made to a program while creating the packages in a Linux distribution.
A typical software engineer does not participate much in verifiability; however, the ability to make modifications to the source goes hand in hand with the need for verifiable binaries in open source.
Given the even greater possibility of hiding things in generative AI, a class of verifiable models may be needed before people can safely use open source LLMs. While there are admittedly randomised elements in stochastic gradient descent and other aspects of training, requiring that the original training data be available is an important step towards verifiability.
Note also that the more expensive it is to train a model, the more important it is being able to verify the result.
The difficulty I see is that more bugs are likely when simultaneously using two languages in the kernel compared to using only one. Thus C mixed with Rust will have more security issues than only C, just because of the mixing, even if Rust might be better.
I also find it surprising that Redox OS has been in development for 9 years and still not available for production use on any cloud platform, but that's a different story.
The article states "Although some of the betas offered Wayland as an option, it's gone from the final release, which uses just X.org." If the Deepin desktop is as widely used as suggested, sticking with X11 may say something about direction of the window fracture in Linux.
As a contrast Raspberry Pi OS switched to Wayland by default last year. Though tied to a particular vendor's hardware, in my opinion that OS is also trend setting.
From what I understand, the efficiency core is not a performance enhancement but a setback that scales up to support more VMs while still meeting data center thermal and power requirements. As likely as not, eight to 32-core CPUs are chasing single thread performance and anyway don't have much trouble meeting the power limits of even the oldest racks.
I think a similar book written by the same author about RISC-V would be appropriate whether or not this takedown from Softbank ever happened.
In particular, RISC-V does not require any malice towards ARM to succeed. Since Linux, GCC and LLVM have reached the level where ARM is practical in the data center, so RISC-V and even Loongson are equally practical.
In the end it's about who has the better and faster hardware. ARM is causing trouble for companies with architectural licenses to innovate, see the Qualcomm Nuvia lawsuit, while RISC-V is so easy no license is required.
For me the main take away message is how important the high-bandwidth memory Xeon Max will be in high-performance computing. This is explained in the quote by Grider that "it hardly ever happens in computing that you can move to a new system and see huge gains without changing the codes, but the switch from Trinity to Crossroads will do just that." Said another way, crossroads combines the magic of Riken's Fugaku supercomputer with the more familiar x86 architecture and execution environment.
My experience with AntiX-19 is that it was one of the few distributions which ran on older 32-bit processors such as the Athlon Thunderbird and Pentium III that lack SSE instructions. Alpine used to run on such systems and was less political, but the last round of updates broke that for me. Does anyone know whether AntiX-23 works on a Pentium III processor?
Another example of a model city built in the desert is Fatehpur Sikri founded by Akbar in 1571 and abandoned in 1610. Presumably, it was abandoned due to lack of water, however, it's possible that Akbar simply lost interest.
Losing interest is also possible in California, especially if the whole point is for certain people to enjoy spending a share of the proceeds. From that point of view, it would be more efficient to sell shares of a gold mine that doesn't have any gold or perhaps an AI startup.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
