MLPerf benchmark suite for machine learning announced

This is awesome to see. The state of deep learning benchmarking is still dreadful, and I think most observers don't know even the most basic details about how it should be done properly. For more, see Wave Computing's press release.

Unfortunately there are some big names not yet listed among the supporters. That includes NVIDIA, unsurprisingly.

About Moore’s Law — it’s dead

I've been waiting for someone of sufficient stature to publicly convey this. If you’re not sure what all this means, look at the graph.

While Intel’s 10nm was the canary in the coal mine, it has taken a couple years for the industry to fully grasp the sheer wall it has hit, and how the other foundries would hit it just the same. Cannon Lake’s extreme delay and Apple’s middling A10X and A11 single-threaded performance improvements (despite what it did with the latter's core) were leading indicators.

We're still getting shrinks, but they aren't timely enough to double transistor count every two years anymore.

While there are other areas that can be advanced, we really need materials breakthroughs to be able to push per-core performance again. Until then, we’re mostly stuck.

"Spectre/Meltdown Pits Transparency Against Liability: Which is More Important to You?"

All hardware is degrees of broken. I've unfortunately found, however, that many vendors are happy to advertise their silicon as fully functional despite shipping broken implementations or disabling IP or features outright.

And in the case of the Meltdown and Spectre vulnerabilities, most modern CPUs were deliberately designed with what ultimately proved to be a poor balance between security and performance in regards to their speculative execution implementations.

"HiSilicon Kirin 970 — Android SoC Power & Performance Overview"

If you want to learn about the state of mobile chipsets, AnandTech’s power and performance overview of HiSilicon’s Kirin 970 is a good place to start. It’s not intended to be a comprehensive overview of the SoC, but it’s also easier to read than such a piece.

While you may not care about Huawei or Chinese silicon vendors — you should! — it’s important to follow HiSilicon’s SoC implementations and results, as they among others serve as a barometer of the state of the ARM ecosystem and all mobile SoCs. The company's best effort to date was Kirin 950, which was a very well-implemented chipset.

Real analyses also can and should convey the things that really matter when it comes to SoCs: interconnect implementation, memory access latency, whether the power management works at all, etc. These are some of the things that most often go wrong, or are the most challenging to implement well. The average mobile observer probably only thinks about CPUs and GPUs, but they’re not even remotely the only things that matter.

Andrei is the only person writing publicly who knows how to measure power properly at the rails (or even publish fuel gauge figures), so you can trust these power figures, unlike everything else you may find on the internet. Additionally, his figures provide a good overview and recap of the state of mobile chipsets in recent years.

This is also the best public data we have on 10nm at the moment, and the results echo what I could gather about the node early on (see: Updates). I particularly appreciate his compiling SPEC2K6 for readers' benefit, since that’s a genuine pain to do.

“An increase in main memory latency from just 80ns to 115ns (random access within access window) can have dramatic effects on many of the more memory access sensitive tests in SPEC CPU. Meanwhile the same handicap essentially has no effect on the GeekBench 4 single-threaded scores and only marginal effect on some subtests of the multi-threaded scores.”

This section as well as the other commentary on benchmarks should sound very familiar to subscribers. SPEC is not exactly the best benchmark in the world in terms of real-world representativeness (read: understatement), but it’s the best we’re going to get publicly.

I should note that interpreting benchmark results beyond the basics is tough. You need to really know what’s going on on a low level. Deep learning benchmarking is also complicated, and I’m not a machine learning researcher so I’ll refrain from commenting about that.

The upcoming SoC to watch right now is Exynos 9810. Samsung’s System LSI division really needs to deliver following recent disappointments that failed to live up to the solid Exynos 7420.

Lastly, if you’re hoping for greatness from 7nm, I would argue that it would probably be better to start accepting that Moore’s Law is dead. I’m not the person to write about that, though.

About Fuchsia

I've mentioned this before for subscribers, but I may as well say it here: it seems obvious to me that we'll see Fuchsia/Zircon devices this year.

Think months, not years.

David Kanter on Intel's 22FFL process

I heard Intel say many promising things about 22FFL at TechCon 2017, but it has quite a lot to prove when it comes to SoC processes and winning clients for Custom Foundry.

In-depth public analysis of foundry technology is rare, so I'm grateful that David has written about this important topic.

Treble and Fuchsia

If you had to combine Android and Fuchsia, how would you do it?

This article is available for subscribers on Patreon.

Ending subscriptions

I have decided to end subscriptions for Tech Specs. Existing subscribers for this month will be refunded at the end of the month, and all subscriber articles will still be accessible until the end of the month.

Tech Specs will not necessarily end as a blog, but as of now I don’t know what its future will hold. I want to reenter the tech industry for full-time work, and understandably employers generally don’t allow employees to blog. If I can continue to write in some capacity, I will, but I hope you will understand that it is unlikely to be at the same level of depth or weekly commitment.

I deeply, deeply appreciate the support of all my subscribers to date. If not for your support, I would have stopped blogging many months ago, and I honestly kept going as long as I could justify. The comments and questions on subscriber articles were especially great, so my thanks for all of those. Please do feel free to continue sending me any questions or comments via email or Twitter. I always try to respond eventually.

I know $10 a month was not cheap, and I debated launching new pricing options for many months. Ultimately, unless offering cheaper options would have increased the number of subscribers by an enormous multiple, it would have done very little to move the needle. There was no realistic path towards being able to justify the continued time and especially financial expense. I do feel like I could have eventually made the numbers work, at great effort, though it would have taken many years. That kind of time is unfortunately something I do not have.

Public writing was not something I had anticipated doing much of, as it was the benefit of circumstance. While my writing is not very good, I aimed for an intermediate-level of technical depth that was hopefully not too difficult to follow. It’s hard to say if that was the right call. My one regret is that I didn’t make time to write any introductory articles, or even a technical article or two. 

Above all, I’m sorry to disappoint everyone.

I will, however, write at least one more article for subscribers on one particular topic. I’ve put it off for a long time because it requires much more research than anything I’ve written about to date. You can probably guess the topic.

My thanks to everyone for reading, and hopefully this is not the end.

"Evolution of the img tag: Gif without the GIF"

Everyone loves GIFs, but they're technically awful. This replacement implementation is awesome. The WebKit/Safari team's work has been absolutely amazing in recent years, especially in regard to energy efficiency improvements.

Also:

Aside from not having a formal standard, animated WebP lacks chroma subsampling and wide-gamut support.

"Process Technology Limbo"

These are some highlights from Greg Yeric's keynote presentation at ARM TechCon 2017. Greg leads the Future Silicon Technology group within ARM Research and is as qualified as anyone to speak about the industry's current challenges.

It's a bit of a scary time for the silicon industry. Progress is slowing down in many areas, and it's unclear to everyone what technologies and processes will successfully drive advances in the future.

Introducing The List

With the increasing popularity of UHD TVs in the market and the Xbox One X’s awesome backwards compatibility support, I’ve been working on a small side project over the past few months to make sense of the increasing complexity of console and handheld gaming. Today I’m launching a public spreadsheet which I’m calling The List. Consider it a console gaming “optimization” guide.

The List is linked under a new Reference section of the blog, which I will be expanding over time to include various tech-related reference information. The goal is to list the best version and means of playing almost every (good) console or handheld game ever, within reason. For example, I’m guessing not many people would realize that the best way of playing of some original Xbox games is through an Xbox One X connected to a 1440p FreeSync monitor.

I’m sure this will disappoint some people, but I am not including PC games. This is for a couple reasons. In general, 95+% of multi-platform games will of course run best on the PC (on Windows, specifically), so there isn’t much value to add in that regard. Also, optimizing software settings for PC games and all their potential hardware configurations can be maddeningly complicated if you truly want to play without compromises.

Console gaming is just complicated enough that I think a simple guide has merits, especially with the advent of the PlayStation 4 Pro, Xbox One X, and even Nintendo Switch. This guide was initially inspired by the confusion created by the various output display modes of some games as part of their support for the PlayStation 4 Pro.

Unsurprisingly, I am heavily leaning on the excellent analysis of the folks at Digital Foundry, who are the best at head-to-head comparisons of games. While graphical and audio comparisons are fairly straightforward, comparing ports on features and overall quality is much trickier and more subjective. I will try my best to be thorough and fair.

I’m launching The List with 100 games to start with. I’m not sure if more than a couple people will find this project of value, but it’s worth a shot. It would be awesome if anyone would like to contribute suggestions or help out, but I am by no means expecting anything. And if anyone has any specific games they would like to learn more about, please do message me on Twitter. Thanks.