558 karma · joined April 25, 2014
The comparison to ADHD made in the article is an interesting one: although for sure many ADHD kids do age out of the diagnosis, I know many ADHD adults who still have it significantly impact their lives. What do we do with addiction if you've made it past your brain developmental point and you're still addicted?
I feel like x86 offerings often fall down so flat in this area. Even when theoretical memory bandwidth is provided, there are often other significant limitations that prevent its use. I've run into this personally with AMD Zen cores, where it is pretty easy to saturate the infinite fabric bandwidth.
Do you know why they would outfit it with LPDDR5 instead of regular DDR5? I don't have a good feel for the difference here. At first blush I would assume low power DDR would run slower not faster, but that doesn't seem to be the case anywhere I see it deployed.
Along the lines of another comment on this post, part of the problem is the GPU compute model is a lot more abstract that what is presented for the CPU.
That abstraction is really helpful for being able to simply write parallel code. But it also hides the tremendous differences in performance possible...
According to the steam software survey as of Feb24, there's still almost 8% of the market that thus doesn't have access to AVX2. Which for many practical purposes means falling back to SSE for everything unless you want the complexity of dynamically dispatching functions according to instruction set and the complexities of it (https://news.ycombinator.com/item?id=24577069). Even for significant performance, I usually do not.
The reason that whole rant is relevant is that the trick in the article (which is simple and short) doesn't even have a cost in SSE code! For SSE you have to do a load and shuffle anyways since there is no lane broadcast instruction. (The set1_ps intrinsic hides this from you, since it will compile instead to a broadcast if you have set the compiler suitably). Since you're already shuffling, there's no downside to this trick, only reduced register pressure.
Comparing ancient empires to modern living standards is nonsensical for the purposes of determining if plunder was the basis for empire. And indeed, we know plenty of cases where plunder made wealthy and then ravished entire nations such as the Spanish conquest of Middle America.
His fundamental point is accurate though - there is not a fixed pot of wealth in the world, and the increased wealth of our modern age is due to technology and efficiency, not simply taking things from others.
Of course, the first thing they're known for is letting you buy comically huge volumes of everyday items.
But one of the reasons that they actually have a cult following is that their corporate buyers are actually incredibly good at finding interesting items and offering them at a good price.
Browsing Costco essentially becomes a curated tour of what's new in the land of food, gadgets, etc. All backed by an extremely consumer friendly set of policies allowing returns etc, which minimizes your risk.
Thus at sufficiently small scale it does break down, but the fractal nature over a very broad range of scales is clear; I feel like the Wikipedia article is tripped into pedantry at that point :)
The coastline paradox is the classic example of a fractal; the dimension of a fractal is actually greater than that of the space it lives within!
Rant out of the way, obviously 10x faster is still an order of magnitude faster! I've continued to be curious if there's a way to properly batch geometry in the QT backend to remove all the state-changes that the QT painter is doing that kill performance; a powerfully GPU-accelerated backend for widget based apps would be incredibly useful, although given the dominance of web/electron apps these days not a useful feature for a commercial company to build.
The most relevant line is this: "to most other languages you are simply passing the address of a buffer". That's true! And so for the love of all that's holy, don't transfer ownership. Simply pass a (non-owning, defined-lifetime) pointer to your data to the relevant dependency.
Have asynchronous issues? Need to extend lifetime sometimes? Well, yeah, unique_ptr is not going to help you there. It's not magic.
Care to share?
For example, here's a quick scratchpatch I used for calculating memory bandwidth / CPU transfer rates: https://instacalc.com/56842
Things like branches, loop termination not being regular, alignment, aliasing, function calls, etc all can and do stop your otherwise totally amenable code from vectorizing.
In the terminology of the article, I'm still at Phase 2. By the time I've worked out the dataflow and constraints needed to make a loop parallelize, I've more or less converted it into the form needed for intrinsics already, so why not write them and be sure.
More broadly, the biggest impact I've seen using MSVC (which most of my code is written in), is the function inliner makes absolutely terrible decisions, particularly with lambdas, which is really unfortunate because there's no way to force-inline the lambda. Combine that with the above - the presence of the lambda will prevent vectorization, even though if it was inlined it would then be trivially vectorizable. :(
To prove it true, you just need to show a human who can run that fast. To provide it fast, you have to accumulate incontrovertible evidence that it's simply beyond us...
The essence of a bicycle is not a motorcycle with a human motor. However, that is a very good analogy for what a bicycle is.
Consistent with your experience I lost weight but it was almost all lean mass -fat was reduced only slightly.
I switched to a ketogenic diet after that and watched the fat melt away. IF clearly works for some people but it sure didn't for me.
But this is true of almost any popular language! By the metric people are applying to C++ here, there is no one on earth who "knows Java", which is patently ridiculous.
A better metric would be "Can write code in that language for some domain".