easde

Oh, I definitely agree that the most prudent decision a few years ago was to avoid Tesla like the plague. Why would Apple, the most profitable company in the world, want a nearly-bankrupt car manufacturer on its books?

Nevertheless, with the benefit of hindsight, it was likely the wrong decision from both a financial and strategic viewpoint. It's the kind of wrong decision Mark Zuckerberg avoided when Facebook bought Instagram and WhatsApp (despite being ridiculed at the time). Over the next decade, Tesla will likely be one of the most serious threats to Apple's dominance in high-end consumer tech goods (the other being companies working on AR/VR).

How so? Apple could have bought a big chunk of Tesla and flipped it back to the public market for ~10 times the price today, for a cool $500B+ profit. Even if they wanted nothing to do with the company's operations.

Of course, such an operation would typically be done by a hedge fund or private equity firm. But guess who is responsible the one of the world's largest hedge funds [1] with $200B+ AUM...

[1] https://en.wikipedia.org/wiki/Braeburn_Capital

Keep in mind, Tesla didn’t design or build the computers.

Tesla has shipped their custom SoC in every new car since April 2019.

Only for the first-gen Roadster, all the way back in 2008. Total sales: 2450.

If you don't want the chip to throttle, the simple answer is to get the Macbook Pro, not hack together a crappy laptop from a Mac Mini :)

Even then, the throttling on the Air seems pretty minor. About ~30% drop in performance after it levels off, which takes several minutes. Less if you get one of those pads with a fan in it that blows on the case.

Aside from what other people have mentioned already, there's one more trick. The Chase Freedom Unlimited gives 1.5 points per dollar on all purchases, which can be transferred to the Reserve and spent at 1.5 cents per point. Effectively a 2.25% cash back card if you don't mind using Chase to book travel, which is among the very best.

Until mid 2021 several other things also count for the travel credit such as gas and grocery stores. Chase probably did not want a wave of card cancellations over the $300 credit being mostly useless during the pandemic.

Not just masks. Early projections (March / April) by supposedly reputable groups like IHME were wildly inaccurate. Even a cursory reading of their methods showed that they had no idea what they were doing. Oversimplifying a bit, they were fitting the CDF of a Gaussian to the cumulative deaths over time. Not a reasonable way of modeling exponential growth.

Nevertheless, IHME was kept in the spotlight by the government and media because they deemed the experts as an independent health research center at a reputable university. Their predictions drove many poor policy decisions, culminating in the infamous claim by Trump that the pandemic would be over soon after Easter. Many epidemiologists, statisticians, physicists, mathematicians etc. called their conclusions into question and built better models (mostly using SEIR / SEIS). It took several more months for IHME to improve their own model, but by then the damage was already done.

Many of the tech companies are in the "bad" neighborhoods. Most notably Uber, Twitter and Square are all headquartered right next to one of the worst parts in the city (Civic Center / Tenderloin).

I remember walking to one of those companies for a job interview and seeing someone pooping in broad daylight on the 8th and Market intersection. Not a great first impression.

TorchScript JIT (torch.jit.script) is similar for PyTorch.

Back then a software engineer at a well-known big company would make around $30-50k (not inflation adjusted). Combined with 1993's interest rates (~7%) $300k was quite steep.

Today the same engineer at a FAANG would make about $300k. Houses are about $2M and interest rates are about 2%, so relatively speaking things are cheaper. However, unlike in 1993, today's salaries are unlikely to grow by 10x over the next 30 years while the mortgage is paid off!

Hardly any of this matters in userspace or device-independent kernel modules, which is what 99% of developers work in.

I like Apple's approach of having more narrow vector units rather than a few extremely wide ones. Narrower vector units are generally easier to program and easier to support on small cores (like the efficiency cores). It's a good tradeoff for cores used outside of niche applications that do a ton of numerical calculations.

Compare this to a design like AVX/AVX2 that has wider 256-bit vectors, but almost always forces the programmer to treat them as two 128-bit vectors glued together. All because Intel realized they had to power down half of each vector ALU to save power at idle.

Of course, the future is hopefully more flexible vector ISAs like SVE. AVX-512 is not too bad either today, but Intel seems almost pathologically opposed to making forward-looking decisions (what about when they want to double the vector width again??).

This is an oversimplified statement. Extra decode width only helps if the rest of the CPU can actually process so many instructions in parallel without blocking on other resources like memory. x86 CPUs also have their own features to alleviate the limited decode width such as micro-op caches.

Case in point is the Samsung Exynos M3, which despite its 6-wide decoder is barely competitive with the 3-wide ARM Cortex-A76. The Exynos would fare very poorly against any recent 4-wide Intel or AMD CPU.

The crazy thing about the M1 is that every structure in the CPU is huge - the decoders, the ROB, the number of execution units, even the caches. And apparently all of it is implemented very efficiently.

I don't think I've encountered a single smartphone by a reputable vendor that has had significant issues sleeping, no matter the chip vendor or OS.

Meanwhile people do have serious issues with sleep even on Intel Macs running MacOS and Thinkpads running Windows. For whatever reason PC chip / hardware vendors just cannot get sleep to work properly. Apparently the problem is so low-level that even a company like Apple cannot paper over the problems without throwing out Intel entirely and using their own SoC in the new Macbooks.

Old server hardware can be very cheap, although it's not necessarily the best value. Typically it only makes sense if 1) you need a lot of memory, 2) you don't mind a slightly higher power bill and 3) you don't care about single core performance.

I've used a 12-core Haswell Xeon as a desktop for about five years (got it used for around $500, plus $200 for a brand new LGA 2011 v3 motherboard). At the time it was twice as fast as a high-end desktop 4-core CPU. It's still reasonably competitive on multithreaded workloads versus a recent 6 core desktop CPU, but fares poorly on single-threaded tasks. DDR4 was a bit expensive back then, but quad channel memory helps in that you can buy lower density DIMMs. Today, even DDR4 ECC RDIMM server pulls are dirt cheap. I picked up 4x16GB DDR4-2400 RDIMMs for just $90 last year.

I'll probably replace it with a 5900X soon, but it's served well over the past five years. Certainly much better than a i7-4770K box from the same time.

Yes, consider vector floating point fused multiply-add (FMA). On a typical AVX implementation like Skylake, this instruction has a latency of 4 cycles and a throughput of 2 instructions per cycle. To avoid stalls, you'd need 4 * 2 = 8 instructions to run independently, and 8 architectural registers to simply store the results. You could store the results onto the stack and reuse the same architectural registers, but usually you want to use the values immediately in the next loop iteration (eg. matrix multiply) so this would be expensive. You probably want a few more architectural registers (at the very least 2, up to 16) to hold the inputs as well.

The study compares how many people in the experimental group and control group develop symptomatic COVID-19. In this case 11 people in the experimental group developed COVID-19 versus 185 in the control group. Thus the vaccine is approximately 185 / (185 + 11) = 94% effective, assuming both groups behaved similarly.

If the control group intentionally avoided exposure more than the experimental group, then the vaccine is even more effective than the study result. This is a blinded study though, so participants don't know which group they're in unless they go out of their way to find out (eg. taking an antibody test).

You're making completely the wrong comparison. On the left you have Geekbench 5 scores for the A12Z in the Apple DTK, and on the right you have Geekbench 4 scores for the Ryzen.

The M1 has leaked scores of ~1700 single core and ~7500 multicore on Geekbench 5, versus 1200 and 6000 for the Ryzen 4750U.

Refurb 12" Macbook?

Wish Apple would reintroduce those with the M1. Awesome form factor, held back by Intel's anemic Core M CPUs.

Dell even sold a 4GB (!) version of the XPS 13 for $999 up until a few months ago.

They're lithium-ion batteries with carbon fiber electrodes.

Doesn't matter what kind of battery it is though, once you start cutting through them you are very likely to short the electrodes together. Which usually ends in fires.

Apple's perf/watt is much higher than AMD. Anandtech has the A14 within a few percent of the 5950X on single threaded workloads. Power consumption is 5 watts (entire iPhone 12) versus 49 watts (5950X package power).

Normalizing for performance (say A14 vs 5950X at 4.5GHz instead of 5GHz) would close the gap somewhat, but it's still huge. Perhaps 4x instead of 10x - those last 500MHz on the AMD chip cost a ton of power.

Of course, none of this is particularly surprising considering Apple is using both a newer process and gets nearly 60% higher IPC.

[1] https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de...

Anandtech posted some comparisons of the A14 against Zen 3 today, which may be an interesting comparison: https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de...

Seems like the A14 is within 10-20% of the desktop 5950X in single threaded workloads. The M1 will probably close the gap with higher clock speeds. AMD will probably still be ahead on multithreaded workloads until Apple releases a chip with 8 high-performance cores.

BTW, Stripe has already left San Francisco for South San Francisco (which is a different city).

It's true that depthwise convolutions are bandwidth bound, but most networks that use them in combination with a kernel size 1 "convolution". If those two operations are tiled and fused together, the result is often compute bound again. This is usually not the case in most ML frameworks and libraries though, including CuDNN.

AB5 passed with approximately 3/4 of the vote from both houses of the state legislature, so even a 3/4 threshold would have placed Prop 22 at immediate risk of being overturned by the legislature.

These changes are more about protecting the average user from shady repair shops that use fake parts. There are some particularly egregious examples such as cheap LCD replacements for OLED iPhone displays. Someone getting their iPhone fixed or buying one from the secondary market would likely be quite unhappy to end up with such a display on their expensive phone.

It would be nice if Apple made genuine replacement parts more generally available, but that's kind of an orthogonal issue - a shady repair shop might still go with the aftermarket LCD for $40 rather than the genuine replacement for $100.

An old $100 PC with an RTX 2060 6GB would be a lot faster than a Jetson NX and only about $400 total.

You'll want tensor cores, so I would stick to the 2000 or 3000-series Nvidia cards. Not to mention the headaches of using AMD cards with any ML framework.