HNHacker News
TopNewBestAskShowJobs

tarlinian

213 karma · joined November 27, 2016

submissionscomments
tarlinian··on Ask HN: What are the least competitive consumer and enterprise markets?
What does this mean...you still need some sort of analog capacitor to collect photons and convert them to electrons.
tarlinian··on TSMC plans to halt chip supplies to Huawei in 2 months
HiSilicon does not own a fab. They're just like Qualcomm, a fabless design house.
tarlinian··on Linux founder tells Intel to stop inventing magic instructions and fix problems
This is absolutely not true. They are available for every publicly available SKU.

https://www.intel.com/content/www/us/en/products/docs/proces...

tarlinian··on The x86 Advanced Matrix Extension
There is more recent work. This blog post by Travis Downs is the most detailed analysis of transition behavior I've seen: https://travisdowns.github.io/blog/2020/01/17/avxfreq1.html

For general guidelines on when to use AVX-512, this (older post) remains best guide I've seen: https://lemire.me/blog/2018/09/07/avx-512-when-and-how-to-us...

tarlinian··on The x86 Advanced Matrix Extension
So in some sense very similar to SKX behavior? The first iteration of the instruction implementation requires judicious use of instructions, while later implementations (this is something to be upset about...those "later implementations" should have been available quite some time ago).

This is also ignoring the fact that none of these penalties come into play if you use the AVX512 instructions with 256-bit or 128-bit vectors. (This still has significant benefits due to the much nicer set of shuffles, dedicated mask registers, etc.)

tarlinian··on Ask HN: What are some good resources to learn how electricity works?
It also is infinitely more clear and simple to understand when you correctly realize that "F equals M times A" instead. I assume this was just an accidental transposition and you're not actually a crackpot. :)
tarlinian··on Apple, ARM, and Intel
Unlike Intel, Micron has practically outsourced most of their manufacturing to Singapore, Taiwan and Japan. Their only production fabs in the US are in Utah and Virginia both of which are mostly focused on niche or older products. More importantly, they only produce memory. Logic and memory processes are very different.
tarlinian··on Apple, ARM, and Intel
There are many other aerospace companies that are in principle capable of designing high performance airplanes in the US. There are literally no other companies in the US with modern semiconductor manufacturing expertise (in logic, not memory, GF doesn't count). The level of capital expenditures and technological expertise and IP required to manufacture leading edge semiconductors is staggering. It's unlikely you could even start a new company that could manufacture leading edge semiconductors for the entire $23 billion that is in that bill. (They'd also probably get sued into oblivion by existing manufacturers with basically zero patent protection.) Just look at what happened to Globalfoundries. It was given effectively free capital for close to a decade, a not totally instate starting point of starting from nothing AMD's old fabs, and in the end gave up.

The "diagnosis" of Intel is that it threw away the chance of higher volume for lower margin which gave TSMC a bunch of money to catch up to them. (I think that's a very hindsight is 20/20 kind of view too...I don't think anyone thought that the iPhone would become larger than the entire PC market when it came out...and at those volumes Apple would eventually realize anyway that it wouldn't make sense to keep paying Intel to make chips for them.) I don't see how a start up is going to capture that volume without some sort of pseudo-captive market to grow with (which is basically how the various companies in China are progressing...).

tarlinian··on Apple, ARM, and Intel
IMO your last statement is kind of ridiculous. Intel is literally the only company that does high volume wafer manufacturing in the US at anything approaching leading edge geometries. Are there any other American companies the government could even realistically support? (The only other US company that has any modern wafer fabs in the US is Micron and the only really have R&D and legacy long-life products being manufactured in the US.) Maybe if this program occurred 20 years ago you might have had another option. (And you would have basically looked like an idiot as for the next 10 years Intel pretty much made all the right decisions in manufacturing while everyone else in the US went down dead ends.)
tarlinian··on Arm is offering early-stage startups free access to its chip designs
This is the opposite of what you want for a having nice software tools.
tarlinian··on Why a Data-Security Expert Fears U.S. Voting Will Be Hacked
"Went missing" makes this sound way more nefarious than it is. It just means that the ballot wasn't returned (e.g., the person didn't vote. I vote by mail and there are elections where my ballot would have gone missing because I didn't have the time to make myself well informed enough about the candidates to place an informed vote.) I don't exactly see what's so implausible about this.
tarlinian··on A note on mask registers (AVX-512)
I have no particular reason at the moment to care about any of this, but your posts are always very interesting reading.

I was actually surprised to see the zeroing idioms not being recognized. I sort of assumed those were cheap to implement. I suppose they may not actually have much value since you would never actually want to use a zero mask register as an actual mask?

tarlinian··on MDS: Microarchitectural Data Sampling side-channel vulnerabilities in Intel CPUs
Some additional pages by Intel describing mitigation techniques for non-HT domains (including the new overload of the VERW instruction): https://software.intel.com/security-software-guidance/insigh...

Details of which steppings of which processors are affected by which CVEs: https://software.intel.com/security-software-guidance/insigh...

tarlinian··on Memory Makers Plan to Cut NAND Flash Production
The depreciation on the equipment used to make NAND would not come close to supporting the prices you describe. In a market with a small number of players, public disclosures of capacity adjustments are sufficient to "collude". But 10 TB for $100 would require a something like a 10x reduction in pricing. None of the memory manufacturers have margins close to that good.
tarlinian··on Analyzing Core I9-9900K performance with Spectre and Meltdown mitigations
I agree, I remember reading this article and thinking something was odd...how could you implement page table isolation in hardware? The Spectre improvement could plausibly be more similar to software mitigation...
tarlinian··on How AI Training Scales
I think you're confused about the benefits/drawbacks of increasing batch size. The training progress per optimizer step increases with batch size (otherwise the smallest batch would be always optimal). Increasing the batch size improves the quality of a single optimizer step but does not scale linearly. (i.e., larger batch training requires more samples to converge than small batch training even though fewer steps are required). Depending on the problem scale the computational efficiency of larger batches makes this tradeoff worth it because even though you need to process more samples to converge, it will take less wall clock time due to improved efficiency.
tarlinian··on IBM’s Old Playbook
> They're a waste of engineering talent and have contributed nothing to the world.

That's arguably true now...but to claim that IBM hasn't contributed anything to the world is just hyperbole. I mean, they invented the relational database and hard drive, and practically every modern computing concept is just a reimplementation of what existed on a mainframe.

tarlinian··on FoundationDB Summit Program Announced
The program includes talks about other layers that Apple is developing. Does anyone know if they are planning on open sourcing any of those layers in the future?
tarlinian··on Intel’s Plans for 3DXP DIMMs Emerge
It's not a news story...mostly analysis/predictions. (If you're talking about generic announcements of XPoint DIMMs that was in the news at the end of may: https://www.anandtech.com/show/12828/intel-launches-optane-d...)
tarlinian··on NetBSD 8.0 released
Lazy FPU state restoration almost never makes sense on modern CPUs (regardless of security issues) because the cost of an interrupt for state restoration is so high and because "FPU" registers are used for a lot more than just FP calculations these days.
tarlinian··on China, India grapple with the consequences too many men
Generally speaking I was under the impression that Islam and Christianity in particular are significantly better represented in southern India, especially if you consider political influence.
tarlinian··on FCC Authorizes SpaceX to Provide Broadband Satellite Services
National sovereignty over airspace does not extend into orbit. Otherwise nearly every earth science satellite in polar orbit would be in violation of international law.
tarlinian··on Intel Publishes Spectre and Meltdown Hardware Plans: Fixed Gear Later This Year
I'm not exactly sure what's more "nebulous" about the timing for consumer chips. This timing is specified for both the consumer and server chips in the same sentence in the press release: "These changes will begin with our next-generation Intel® Xeon® Scalable processors (code-named Cascade Lake) as well as 8th Generation Intel® Core™ processors expected to ship in the second half of 2018." (I guess the nebulous thing is exactly which 8th generation Cores will have the update.)
tarlinian··on Why has CPU frequency ceased to grow? (2014)
Transistor switching frequency has almost nothing to do with processor clock speeds which are almost entirely limited by wire RC delay. You will get increased drive currents by switching to higher mobility materials but the performance improvement over strained silicon isn't that large.
tarlinian··on Intel hit with 32 lawsuits over security flaws
Are you really implying that the fact Apple doesn't even disclose which processors were vulnerable to this a positive?
tarlinian··on Intel hit with 32 lawsuits over security flaws
Do you know how many CPUs Apple makes? It's certainly an order of magnitude larger than all the unaffected vendors combined...
tarlinian··on Intel hit with 32 lawsuits over security flaws
The fact that modern processors speculatively load memory has been known for a long time (and it's not like the fact that permission checking only occurs during retirement is documented...the reason no one considered it a problem was that it was thought of as an implementation detail and literally something you wouldn't document because in the future you might arbitrarily change it). This isn't exactly novel information...and putting two and two together only seemed obvious after the fact.
tarlinian··on Intel hit with 32 lawsuits over security flaws
Really, the "exact same feature"? None of the processors in that report even support OoOE.
tarlinian··on Microsoft Cognitive Toolkit 2.0
The CNTK C++ API is much more straightforward to use than TF(especially on Windows) and is much better suited to deploying serialized models into production as a part of existing C++ applications. The provided examples require way more boilerplate than an equivalent CNTK application.
tarlinian··on Catching Up with Intel's Python and AI Work at PyCon [audio]
It's still CPython but many of the scientific software stack pieces (NumPy/SciPy/SKLearn) have been re-written to call Intel optimized libraries (MKL/DAAL mainly).
← PreviousPage 2 of 3Next →