The Prompt API with audio input requires a GPU.

Significant Trainium2 capacity is coming online in Q2 and scaled Trainium3 capacity is expected to come online later this year

over one million Trainium2 chips to train and serve Claude

We have signed a new agreement with Amazon that will deepen our existing partnership and secure up to 5 gigawatts (GW) of capacity for training and deploying Claude

over one million Trainium2 chips to train and serve Claude

over one million Trainium2 chips to train and serve Claude

chips from different generations running at different speeds still matched the ML performance of single-chip-type training runs, ensuring that even older hardware can meaningfully accelerate AI training.

With increasing levels of hardware failure, Decoupled DiLoCo continues to deliver a high level of 'goodput', or useful training, while that of other approaches nosedives.

By customizing and co-designing silicon with hardware, networking and software, including model architecture and application requirements, we can deliver dramatically more power efficiency and absolute performance.

Chinese labs, for their part, are not purely idealistic: Open-source is not only free advertising but also a shrewd workaround. Without access to cutting-edge chips restricted by US export controls, releasing models openly accelerates the cycle of external feedback and contributions that compensates for constrained compute.

We calculate the aggregate amount of compute (in H100-equivalents) held by Amazon, Google, Meta, Microsoft, and Oracle, as a share of the global total each quarter.

The H100-equivalent unit uses a chip's highest 8-bit operation/second specifications to convert between chips. The actual utility of a particular chip depend on workload assumptions, so H100e does not perfectly reflect real-world performance differences across chip types.

The H100-equivalent unit uses a chip's highest 8-bit operation/second specifications to convert between chips. The actual utility of a particular chip depend on workload assumptions, so H100e does not perfectly reflect real-world performance differences across chip types.

A 606 MiB model at ~49 tokens/s consumes ~30 GB/s of memory bandwidth, close to the c6i.2xlarge's DRAM limit. No amount of SIMD tricks will help when the CPU is stalled waiting for model weights to arrive from DRAM.

All training runs are conducted on an Ascend 910B cluster, and our setup follows the growing evidence that large-scale model training on Ascend clusters is feasible in practice.

E2B LoRA：8-10GB 显存即可训练

单张 24GB 4090 直接部署 32B LLM

Where training a language model took 167 minutes on eight GPUs in 2020, it now takes under four minutes on equivalent modern hardware.

Where training a language model took 167 minutes on eight GPUs in 2020, it now takes under four minutes on equivalent modern hardware. To put this in perspective: Moore's Law would predict only about a 5x improvement over this period. We saw 50x.

Create multilingual experiences that go beyond translation and understand cultural context.

TriAttention enables OpenClaw deployment on a single consumer GPU, where long context would otherwise cause out-of-memory with Full Attention

TriAttention enables OpenClaw deployment on a single consumer GPU, where long context would otherwise cause out-of-memory with Full Attention

At this rate, the phone in your pocket will run today's frontier models before you upgrade it.

The bundle includes four models, including Gemma's first MoE model, which can all fit on a single NVIDIA H100 GPU and supports over 140 languages.

We train and run Claude on a range of AI hardware—AWS Trainium, Google TPUs, and NVIDIA GPUs—which means we can match workloads to the chips best suited for them.

Owning a $5M data center

Cloud computing is essentially local computing with extra, quite pricy steps today for consumer use scenarios. Unless the economics of local hardware truly does fall off a cliff somewhere down the line, I can't see Bezos' vision of a cloud-only future coming true any time soon — even for casual PC users.

There's a hard cap on the amount of chips the human race can physically produce at any one time, at least as of writing. With nation states effectively printing money to outbid consumer tech companies on basic components, I'm not sure demand will come down any time soon. That is, of course, unless those investors and nation states stop believing AI can deliver anything more impressive than repackaged reddit answers, memeslop, and blog posts for the vast majority ...

That means DRAM, but also increasingly other components too. SSD storage is the next component expected to hit a shortage, battering consumer prices hard in the process.

Is it really so far fetched to imagine that most people would most likely be "fine" with renting their full computing solutions from companies like Microsoft and Amazon?

TDP

Flash storage is prone to failure if power is cut while writing or modifying stored data. Basically because an SD card or SSD is a tiny computer that writes files and updates its map of file locations. If this is interrupted, bad things result. When this happens to storage-only, you can repair the file system and usually come away ok. When it happens to the boot drive, the system can be un-bootable. Also other failures can suddenly cause SD card failure, but the most common is power interruption during file writes. The system drive is frequently updated for normal operating system tasks, making a power loss event more like russian roulette for data loss. Tl;dr you may wish to have a periodic task and a separate flash drive whose only role is to sit idle, then get a copy of any files changed recently. Rsync is a great tool to do so in the Linux/raspberry pi world. This way the really expensive part of your work is preserved, even if your system's SD card fails.

I personally read every single email.

Arbib and Seba explain this by categorising human civilisation into two fundamentally intertwined complexes: the production system, encompassing all the foundational systems by which we meet fundamental material needs across energy, transport, food and materials (corresponding to ‘hardware’); and the organising system, encompassing how the former systems are governed, regulated and managed by society through economic, political, military, cultural and ideological structures and values (corresponding to ‘software’)

The ‘hardware’ is a configuration of matter which harnesses energy from its environment with surprising efficiency and dissipates it as waste back into the environment.

A wired connection to headphones, receivers, or powered speakers

But this position is increasingly untenable. If the EU Chips Act achieves its goal of 20% global production by 2030, emissions could increase up to eightfold, surpassing those of other emission-heavy industries. The climate issue will therefore become ever more pressing.

Plaza Tire Service’s Exclusive Premium Alignment starts as low as $229.99. With the Premium Alignment, you only pay one time for alignments to be performed on your vehicle for the entire time you own it. Just visit at the required intervals to keep your vehicle driving straight. One-time wheel alignments start as low as $129.99.

Color Space Support: 10 bit (8 bit + FRC)

Brightness: 400

Volkswagen is working to identify a technical solution that will continue to offer some telematics services.

For consumers, the equivalent of "build or buy" could be called "ads or nerds". "Ads" meaning ad-supported services, like consumer Gmail or Facebook. "Nerds" meaning hobbyist services based on free software and commodity hardware.

500 cd/m2 max brightness (typical)

From Jim Keller on Lex, there’s three fundamental types of compute CPU: add, multiply, load, store, compare, branch (nothing can be known about anything) GPU: add, multiply, load, store (when things happen is known, but the addresses aren’t) DSP: add, multiply (everything is known except the data) Neural networks are DSPs. All the loads and stores can be statically computed, which isn’t even possible for GPU workloads, never mind CPU ones.

The HoverBar Duo’s base is stable and heavy enough to support Apple’s biggest 12.9” iPad Pro.

Mona also makes good use of the iPad’s support for keyboard shortcuts.

You can make an antenna to better receive KOPN! FM antennas work best when they are sized for the wave length you are trying to receive. KOPN’s wave length at 89.5 MHz is 10.99 ft. A folded dipole antenna can be made from a 5 ft. 6 in. length of 300 ohm twin lead antenna cable. Strip both ends of the cable. Twist the two wires together at each end and solder the connection. Cut one of the wires in the exact middle of the cable and strip back both ends. Connect these two ends by twisting and soldering to whatever additional length of twin lead cable you need to reach the 300 ohm antenna connection into your FM receiver.

I know it's not in fashion, but I will suggest that renting physical servers is a very good and under-appreciated compromise. As an example, 45€/month gets you a 6-core AMD with 64GB of RAM and NVMe SSDs at Hetzner. That's a lot of computing power!Virtualized offerings perform significantly worse (see my 2019 experiments: https://jan.rychter.com/enblog/cloud-server-cpu-performance-...) and cost more. The difference is that you can "scale on demand", which I found not to be necessary, at least in my case. And if I do need to scale, I can still do that, it's just that getting new servers takes hours instead of seconds. Well, I don't need to scale in seconds.In my case, my entire monthly bill for the full production environment and a duplicate staging/standby environment is constant, simple, predictable, very low compared to what I'd need to pay AWS, and I still have a lot of performance headroom to grow.One thing worth noting is that I treat physical servers just like virtual ones: everything is managed through ansible and I can recreate everything from scratch. In fact, I do use another "devcloud" environment at Digital Ocean, and that one is spun up using terraform, before being passed on to ansible that does the rest of the setup.

Highlights

Towards open source patents: Semi-automated open hardware certification from MediaWiki websites

Executive Summary

Equitable Research Capacity Towards theSustainable Development Goals: The Case forOpen Science Hardware

Here are two products that are basic rectangular boxes with a rounded edge (the one on the left also has some unpleasant drafted walls, but that’s another article about how to become a hardware design snob). Look at the beginning and end of that rounded edge on the main surface. See how there’s a sharp shift in highlight? That’s the result of tangency.

The template for what personal computing could become was really obvious by the end of the 1970s. If you look at Engelbart, it was obvious in 1968. But it did take quite a while for the computer chips to be powerful enough and inexpensive enough to make the kinds of things that billions of people use today.

#Software #Hardware #Technology #Programming #Internet #Web Development #Data #Beginner Tutorial #Computer Science #Computers #Engineering

If you put a bunch of research into designing a really great product and it succeeds but gets effectively copied by low-cost clones, you’ll be sad. I am not sure how to defend this, and I think it is probably the weakest point of this business model;

For Nvidia, the speed of the 3080 package makes for a solid sales pitch: This cloud PC is probably faster than your home system, so cloud gaming is worth it. Cloud gaming will always present a latency tradeoff, but that latency is easier to accept if you're getting otherwise-unattainable graphics quality along with it.

The most notable are Remarkable, Onyx Boox, Boyue Likebook, Supernote, and now Kobo.

I have an iPad Pro, a 2018 model that I bought earlier this year because I felt that it was important to have an understanding of the iOS ecosystem even if I was using Android with my smartphone.

85,013

63,968

il faut envisager ici l’échelle de la machine en adéquation avec celles des énergies terrestres et concevoir des structures qui, dans le temps, épouseront les cycles de l’ère géologique que nous habitons

the logical and physical page addresses are decoupled. A mapping table, which is stored on the SSD, translates logical (software) addresses to physical (flash) locations. This component is also called Flash Translation Layer (FTL).

NAND flash pages cannot be overwritten. Page writes can only be performed sequentially within blocks that have been erased beforehand.

For example, if one looks at write latency, one may measure results as low as 10us – 10 times faster than a read. However, latency only appears so low because SSDs are caching writes on volatile RAM. The actual write latency of NAND flash is about 1ms – 10 times slower than a read.

Another important difference between disks and SSDs is that disks have one disk head and perform well only for sequential accesses. SSDs, in contrast, consist of dozens or even hundreds of flash chips ("parallel units"), which can be accessed concurrently.

SSDs are often referred to as disks, but this is misleading as they store data on semiconductors instead of a mechanical disk.

Volkswagen, the world’s largest car maker, has outspent all rivals in a global bid by auto incumbents to beat Tesla. For years, industry leaders and analysts pointed to the German company as evidence that, once unleashed, the old guard’s raw financial power paired with decades of engineering excellence would make short work of Elon Musk’s scrappy startup. What they didn’t consider: Electric vehicles are more about software than hardware. And producing exquisitely engineered gas-powered cars doesn’t translate into coding savvy.

For example, the one- pass (hardware) translator generated a symbol table and reverse Polish code as in conven- tional software interpretive languages. The translator hardware (compiler) operated at disk transfer speeds and was so fast there was no need to keep and store object code, since it could be quickly regenerated on-the-fly. The hardware-implemented job controller per- formed conventional operating system func- tions. The memory controller provided

Unfortunately no - it cannot be done without Trusted security devices. There are several reasons for this. All of the below is working on the assumption you have no TPM or other trusted security device in place and are working in a "password only" environment.

Ainsi, par le recours aux outils informatiques, la pratique architecturale se voit coupée en deux (opérant par là le renversement d’un paradigme accepté par la discipline depuis Alberti) : une partie hard dont l’architecture est responsable, et une partie soft rendue aux habitants

Połączona sieć komputerów działających w ramach inicjatywy Folding@Home przewyższyła swoją mocą obliczeniową najwydajniejsze siedem superkomputerów na świecie. Dobrze przeczytaliście: połączone w ramach F@H urządzenia dysponują mocą obliczeniową na poziomie 470 PetaFLOPS - wyższą od siedmiu najwydajniejszych superkomputerów na świecie razem wziętych! To trzy razy większa wydajność od tej, którą dysponuje najwydajniejszy obecnie superkomputer SUMMIT (149 PFLOPS).

Use a USB-condom. This is a device that plugs in between your normal cable and the computer and blocks the data lines

Use a USB cable with a "data switch". This cable is normally power-only, which is what you want 90% of the time. However there is a button ("Data Transfer Protection On/Off Switch") you can press that will enable data. An LED indicates the mode. This kind of cable is much safer and secure, plus more convenient for the users. It follows the security principle that if you make the defaults what you want users to do, they're more likely to follow your security policy.

espite the potential of emerging technologies to assist persons with cognitive disabilities,significant practical impediments remain to be overcome in commercialization, consumerabandonment, and in the design and development of useful products. Barriers also exist in terms of the financial and organizational feasibility of specific envisionedproducts, and their limited potential to reach the consumer market. Innovative engineeringapproaches, effective needs analysis, user-centered design, and rapid evolutionary developmentare essential to ensure that technically feasible products meet the real needs of persons withcognitive disabilities. Efforts must be made by advocates, designers and manufacturers to promote betterintegration of future software and hardware systems so that forthcoming iterations of personalsupport technologies and assisted care systems technologies do not quickly become obsolete.They will need to operate seamlessly across multiple real-world environments in the home,school, community, and workplace

Mantenimiento

independent

Your hardware modular rig is completely integrated,

The biggest benefit may be from substantially increasing the amount of physical memory in a server: 2 socket Xeon systems can hold up to 3TB of RAM, but 24TB of Optane, and 4 socket systems support up to 12TB RAM, but 48TB Optane. This could be a huge boost for applications that need truly enormous quantities of memory.

Unlike flash, which physically wears out due to the stress placed by erases, 3D XPoint writes are non-destructive. This gives the drives much greater endurance than NAND of a comparable density, with Intel saying that Optane SSDs can safely be written 30 times per day, compared to a typical 0.5-10 whole drive writes per day.

3D XPoint is a new kind of persistent solid state memory devised by Intel and Micron. Details on how the memory actually works remain scarce—it's generally believed to use some kind of change in resistance to record data—but its performance characteristics and technical capabilities make it appealing for a wide range of applications

Two 140mm fan (front)

paradoxically drive more people to purchase Apple hardware.

In addition to the improved performance, Big Sur is far more versatile and efficient than the off-the-shelf solutions in our previous generation. While many high-performance computing systems require special cooling and other unique infrastructure to operate, we have optimized these new servers for thermal and power efficiency, allowing us to operate them even in our own free-air cooled, Open Compute standard data centers.

Developing code for an FPGA is actually describing how the above pieces are configured and interconnected.

lowRISC is producing fully open hardware systems. From the processor core to the development board, our goal is to create a completely open computing eco-system. Our open-source SoC (System-on-a-Chip) designs will be based on the 64-bit RISC-V instruction set architecture. Volume silicon manufacture is planned as is a low-cost development board. There are more details on our plans in these slides from a recent talk lowRISC is a not-for-profit organisation working closely with the University of Cambridge and the open-source community.