exged

Just wondering, what makes NN-512 superior to Intel's oneAPI DNN library?

Almost everything has at least an 8-bit MCU now. 32-bit MCUs are also extremely common now, and even simple ones like the ARM Cortex-M0 are competitive with a 286/386 (albeit with less memory / no MMU).

A rather surprising number of devices run very powerful application processors. An amusing example is Apple's Lightning to HDMI adapter, which has an ARM SoC with 256MB of RAM and boots a Darwin kernel in order to decode a H.264 compressed video protocol. Depending on what exactly they put into it (wouldn't be surprised if they borrowed the Apple TV chip for a relatively low-volume product like this) it may be more powerful than a fairly recent computer.

Aside from very simple use cases, you'd probably struggle to render 800x480 or 1024x768 frames on an MCU. At 1024x768 especially, the pixel rate (47 MP/s at 60Hz) and framebuffer size (2.3MB at 24bpp) are just too big even for large MCUs.

APs are the right tool for the job. It's true that for hobbyists, documentation is lacking and DSI is complex to get started with. The best solution would probably extend DSI with something like EDID. DSI displays can then be plug-and-play (like HDMI / DP) using an universal driver.

Another option is eDP, but it's not so common on cheap APs or screens.

Ehh, not really. I've used both Macs (old 15" rMBP, 12" Macbook, M1 Macbook Air) and Thinkpads (T450s, X1C G7, T14 AMD) and Lenovo just frequently drops the ball on things like display, speakers, touchpad and battery life.

Apple has also made some poor decisions in various areas in the past like poor thermals, power-sucking dGPUs, and of course the butterfly keyboard. The M1 Macbook is very nearly perfect though, and I don't know how any of the PC vendors are going to top it.

Garbage trucks should be possible. They are heavy but no heavier than a semi truck, with a tiny fraction of the range (150 miles vs 500+). Plus, low speed city driving is much more efficient than highway driving with an EV drivetrain.

I get that all else being equal, it's better to put everything in RAM, but what point is it just poor software design? 1TB of RAM is nothing to sneeze at. Is there really such a great need for sub-100us read latencies that commodity NVMe SSDs are insufficient?

The shortage is actually a direct outcome of Intel's problems. Historically Intel had by far the highest capacity for leading edge fabrication processes - just for their own CPU production. It didn't make sense for companies like TSMC to invest in additional fabs as long as Intel had the lead. AMD taking the lead and moving ~25% of x86 CPU production to TSMC is a huge shift.

It's not just some gold coins, it's also the matter of $12.7 million invested in his research endeavor. If you raised $12.7 million to start a company like Theranos you'd also be looking at a steep prison sentence for securities fraud.

Some other studies [1] have shown that N95 valved masks are as effective at filtering exhaled droplets as unvalved cotton masks, and considerably more effective than improvised things like bandannas and neck gaiters. It makes some sense - the valve lets a lot of droplets out, but the rest of the N95 mask is probably a far better filter than a cotton mask.

It doesn't contradict the claim in this article, but it's an interesting thing to ponder about the risk of the general public having seemingly all standardized on cotton masks.

[1] https://advances.sciencemag.org/content/6/36/eabd3083

Reticle limited on Samsung 8nm means plenty of room is available on TSMC 5nm.

Also, GDDR does not have appreciably higher latencies than DDR memory when measured in nanoseconds. It's just more expensive than DDR and much more limited in terms of capacity.

If Apple does go the "huge SoC" route I'd expect to see HBM2 memory with socketed DDR4 or DDR5. It'd provide the best of both worlds - extremely high bandwidth and low latency for a small portion (say 32-64GB) of the memory, and high capacity for the rest (say 1-2TB), all without compromising the unified memory concept.

This is not without precedent - recent Xeon Phis, for all their other shortcomings, have had a similar memory hierarchy.

If you have some time, you can MITM the 802.1x auth packets [1] and use a less crappy router. I run this with a VyOS router and the same 5268ac that you have, but it works with things like Ubiquiti routers too. The only catch is you need three NICs on your router, but a cheap USB 10/100 one will do for the port that connects to the 5268ac.

Another option is getting the 802.1x certificate out of a hacked router, but it's not possible as far as I know on the 5268ac. You could buy a hackable ATT router but they're not cheap. Some sellers even sell the key by itself.

Mysteriously, doing this fixed an issue I previously had where SSHing into AWS would fail.

[1] https://github.com/jaysoffian/eap_proxy

I'd expect most things to work and have native versions by next year, with the notable exceptions of virtualization software and package managers like Homebrew.

Virtualization software likely will never be fixed if you expect an x86 guests to work. The hardware obviously cannot natively virtualize x86, nor can Rosetta cannot emulate privileged x86 code. x86 Docker images will also not work.

Homebrew itself will probably be ported to ARM by then. However, there will likely be a long tail of packages that won't have ARM builds for some time.

Minor nit, but what you're describing is loss of color depth. Gamut is how saturated the colors get, which doesn't change much in this scenario.

FB generates around $2 million in annual revenue and $600k in net income per employee (not just engineers). If the only thing you consider is value generated by each employee, then FB should be paying far more. Maybe 400k for a new grad and 1M for a senior engineer?

Below average for someone joining as a new E5 employee, especially if stock refreshers and equity appreciation is taken into account.

That's true, but it really depends a lot on the kind of software you're writing. If you're Google or Facebook and make fractions of a cent per pageview despite doing an enormous amount of computation then optimizing your code and infrastructure is very important indeed. On the other hand someone like Airbnb is very likely to make tens or hundreds of dollars off a handful of pageviews, so cloud costs are much less of a concern.

EEMBC is a very poor benchmark for anything but very small embedded workloads (the datasets generally fit in a desktop CPU's L1 cache), but Micro Magic's claims are even more ludicrous. To quote the person in the article (Andy Huang, apparently their advisor):

"Using the EEMBC benchmark, we get 55,000 CoreMarks per Watt. The M1 chip is roughly the equivalent of 10,000 CoreMarks in EEMBC terms; divide this by eight cores and 15W per core, and that is less than 100 CoreMarks per Watt."

Almost every claim in this is wrong. There doesn't seem to be a published score for the M1, but looking at some Intel/AMD CPU scores on the EEMBC website [1] suggests a score of 50K points per core for comparable CPUs like the Ryzen 9 3900X. So the M1 is more likely around 200-300K Coremarks, not 10K. The M1 also consumes around 15W for the entire SoC, not 15W per core. The real Coremarks/watt of the M1 is probably closer to 20K than 100.

Suddenly 55,000 Coremarks per watt for their chip doesn't sound so impressive when you realize the M1 also contains a full LPDDR4X memory controller, GPU, neural net accelerator, etc. and has 10x the IPC even on this extremely simple benchmark...

[1] https://www.eembc.org/coremark/scores.php

The market is efficient in the sense that it aggregates all knowledge between market participants. Nevertheless, no one actually has a crystal ball to predict events far in the future. Tesla's valuation and volatility can both be explained by some market participants expecting huge growth over a long period of time, with a correspondingly huge uncertainty over the actual outcome.

Just to put some simple numbers on things, a commonly held expectation is that Tesla will grow revenue at ~50% a year until 2030 while maintaining high margins. At that point Tesla's revenue would be approximately 1T with perhaps 100B in net income. A fair valuation would then be around 3T, assuming there is still some future growth left.

This largely explains today's 500B valuation - if there is a 30% chance of this scenario playing out (and Tesla goes bankrupt in the other 70%) then this is a reasonable bet compared to buying other stocks.

As for the volatility, imagine if the average market participant changes their mind and believes that Tesla is only likely to achieve a 40% growth rate over the next decade. This would drop 2030 revenue and profits by more than 40% compared to the previous scenario, and today's stock price would fall significantly as well.

It's "just" 4 GPUs and 64 CPU cores. Nvidia has a flavor of Linux for it, but it's basically just Ubuntu with Nvidia drivers and tools. You can rent a machine pretty similar to this on AWS (P3 instances) for about $12/hour, but it has the last-gen V100 GPUs.

You'd usually run CUDA applications that can take advantage of the GPUs. Deep learning is all the rage these days, and the new A100 GPUs are particularly well-suited. Many HPC applications like computational fluid dynamics simulation take advantage of GPUs nowadays. There are also a number of visualization applications that GPUs excel at.

That seems pretty likely. Most (non-software) engineering disciplines have a ton of Windows-only GUI software.

There aren't any free PCIe lanes on the SoC, so even if they added a bridge chip for a SATA / M.2 SSD it would have to be connected over USB.

Besides, the device is $70. A $10 32GB eMMC chip would be a far more reasonable ask at this price point if greater reliability is desired.

If you want to upgrade you might as well just buy a regular RPi4 and plug in your own keyboard.

To be fair, it's not uncommon for a ML researcher / engineer to use tens (~$10/hr on cloud, $100k from Nvidia) or even hundreds (~$100/hr on cloud, $1M from Nvidia) of GPUs to speed up their iteration time. If there was a way to spend half as much on hundreds of AMD GPUs instead that would be a huge win, well worth even months of the researcher's time.

The catch is that ML software stacks have had hundreds if not thousands of man-years of effort put into things like cuDNN, CUDA operator implementations, and Nvidia-specific system code (eg. for distributed training). Many formidable competitors like Google TPU have emerged, but Nvidia is currently holding onto its leadership position for now because the wide support and polish is just not there for any of the competitors yet.

Totally agree that most PCs have terrible analog sound quality, often plagued with EMI noise and other distortions. I still find BT to be a pain on Linux and Windows (unlike Mac / iOS, where Apple has really figured out the experience). I ended up buying a bunch of Apple USB-C to 3.5mm adapters for $9 each. Fantastic audio quality and tiny enough to leave permanently attached to headphones. The only catch is they aren't great with very high impedance headphones unless you add an external amplifier.

Mozilla is already irrelevant. Firefox market share on desktop is already lower than Safari [1] and continues to decline.

[1] https://gs.statcounter.com/browser-market-share/desktop/worl...

Turing Pi is mostly useful for educational purposes. For anything performance critical, a cheap 8-core x86 box will run circles around 7 Raspberry Pis in pretty much every way, no matter how they are networked.

The "lite" versions don't have eMMC, so you would be free to put an SD card slot or similar on the carrier board.

Correct me if I'm wrong, but the biggest CA income tax deduction available to most people is the mortgage interest and property tax deductions. So yes, you can deduct a lot of your state taxes - provided you're spending even more on an expensive home.

And yes, the tax rate is progressive, but the average rate is still around 9% at that income level (~36k). Marginal rate is just over 11%.

My overall point is just that almost all of the cost of living delta in CA is from taxes and housing. Didn't mean to get too deep into the numbers.

Taxes and rent are still bigger. For a senior FAANG engineer making $400k a year, CA income taxes are nearly $40k/year, and rent on a house is perhaps another $60k/year (alternatively, ~$100k/year for a mortgage). All other basic needs are a rounding error compared to that, even with a family.

Yes, the ESP32 uses a ton of power relative to most MCUs (external flash comes at a cost!) and has relatively poor peripherals, documentation and tooling. But it's cheap and not bad if you need WiFi or Bluetooth.

Most non-networked projects would probably be better served by one of the numerous Cortex-M microcontrollers out there.