Agreed. if its like 2-5 posts maybe, but a multi-page thread that bouncing around in time frame? Nah. Just wrote a blog and tweat a highlight from it and link to the blog post.
When I click on the link, I get taken to a page on twitter.com that shows me a post, in normal linear order. It reads to me like ordinary English text, in the order you ordinarily read or write it in. The text is not interrupted, except by the little "replies / retweets / likes / share" bar that appears after every tweet.
For me it is the same, very easy to follow. But then I remembered you can hit the "star" icon in the top to change your feed to chronological - which I had already done. Maybe that is why it is easier for us?
This is a link to the thread, so it should present the material in ordinary "threaded" order, no matter how you view it (there shouldn't be any buttons to change the order).
let me rephrase your word, but negatively.
The text is interrupted by the space-wasting "replies / retweets / likes / share" bar that appears after every tweet.
Not to mention that on the desktop, the text is literally only quarter of the screen. Add that to the very frequent interrption bars, you are in for a very long scroll while reading
> Not to mention that on the desktop, the text is literally only quarter of the screen.
A quarter of the width? Well, naturally! The optimal line length for readability is somewhere in the 50-100 character range. This is one of those things that is actually backed by empirical research (lots of research... experiments in readability are cheap and easy to conduct, and people have been doing these experiments for a long time). The width of lines of text in tweets is on the shorter end of that range.
Hacker News, for example, is an example of a site with poor readability. The text is as wide as the window. To address this deficiency, I open Hacker News in a narrower window. It's still harder to read.
Why would I assume that the research is wrong? We’re talking about a century’s worth of readability research, here. Cheap, easy-to-replicate studies which have been done over and over again over the decades. Backed by simple-to-explain theories that explain why. Seems like a completely unreasonable assumption, to be frank.
Because it's highly that even if it's correct in aggregate, there will be outliers that the opposite is true, and they'll benefit from different formatting.
That is, even if it's right, it might still be wrong for them. So the question being asked is, why is it made so hard to change?
The context of the thread here wasn't about people with different visual abilities or anything complaining about the fact they're being left behind by Twitter's UX team, it was literally someone asking "Why do you think this is unreadable", some whining that "The text only uses a quarter of the screen", which was responded to with "A century of readability research has shown that's pretty good idea, actually!", which is a literal fact. This is very easy to understand and it isn't remotely computer-centric -- imagine a book where the "margins" didn't exist in any form and words just filled the page from edge to edge. It's not some random thing these basic margin rules also came to computers, and it's not some UX fad; you only need to ask what the difference is between a 1998 CRT and a 2021 4k screen to see why.
Like, you're talking about "what's right for them" as if that's supposed to pull my heartstrings, but the grandparent even literally said "I am going to re-phrase your question negatively", it's not like they're going to magically be convinced or actually argue about any readability metrics in any kind good faith, it's literally just a reflex. I wouldn't waste my time defending it on some egalitarian principles or whatever.
Ah, well I typically don’t have super wide windows, because it has a negative impact on readability. The difference between 1215px and higher widths isn’t really significant anyway… above a certain line length, readability has already reached its minimum and doesn’t decrease further as the lines get longer.
Isn’t that what the HN comment section is like, though? Or maybe I’m just used to reading things like design documents at work (with tons of comments), or standards documents (with footnotes that take up more space than the page body).
HN threads are more likely to have multiple sentence paragraphs in a single comment and replies are nested N levels deep. The UI between each reply is minimal (aka not tuned for “engagement” KPIs) and is almost too subtle.
Also, replies to tweets within one of these multiple-tweet posts are hard to even get to, and I genuinely cannot figure out why Twitter used that design. It can’t help engagement.
What a wonderful non-answer. “I think the answer to your question is obvious, so I will not answer it, and just give a sarcastic reply.” But, I’ll still answer your question about what could make it worse…
It could have super-wide text, like Hacker News. That would make it worse. The line length is a fairly comfortable length, unlike Hacker News’s dizzying long lines.
It could have unusual scrolling behavior, like those Apple product announcements. That would make it worse.
You’ve mentioned the biggest thing that makes websites worse—interspaced adverts and popups, which are downright ubiquitous now. Nearly every site I visit has those.
Apologies - there was no sarcasm intended.
I am dismissive of the twitter format for this content as I don’t like that something that seems suited to a blog post is split into dozens of short snippets with reactions, likes etc beneath. It is rare I make it though them as the format seems designed to be skimmed and skipped and I vastly prefer the HN format with more text, longer lines and a limited amount of cruft. I primarily use HN on mobile.
In case you were wondering, the highest-end computer they're comparing it to is a $6000 W6900X in a 12-core Mac Pro. It's just about the best GPU you can equip a Mac Pro with, but also admittedly a ways off from the fastest consumer hardware.
It will be interesting to see what these tests look like when Apple updates the Mac Pro with Apple Silicon. The rumors have the GPU having 2x and 4x more than the top end M1 Max SOC.
All I'll say is that it would be genuinely unprecedented if they could make that work. The M1 is impressive, but it's family of CPUs were a technical inevitability. ARM has notorious scaling issues, so I'm greatly interested to see if they bruteforce it or go a more tactical route.
The GPU designs in Apple chips has nothing to do with ARM, so even if we accepted that "ARM has notorious scaling issues" (a confusing claim, but even if we took the most generous route and assumed it meant "vendors using ARM have scaling issues" -- also not true) it is orthogonal.
The W6900X/6900XT is basically tied with the 3090 as the fastest consumer GPU. A $10,000 A100 is faster but I don't think it even has any monitor outputs.
As far as I can tell, the "12 core Mac Pro" is a Intel
W-3235. Thats a 2019 processor that isn't even particularly impressive by Intel's current standards - the current Xeon W line goes up to 38 cores, has faster memory, PCIe 4 vs 3, etc.
I don’t have an adblocker on my phone, so maybe I’m to blame but holy crap, the number of ads on that page are the opposite of “normal human consumption.”
I'm not knowledgeable on mining, but I thought I read that ETH performance was largely dependent upon memory bandwidth. If that is true, it should be possible to create performant ETH code with a 400 GByte/sec memory system. Of course, maybe I completely misunderstood something. Please correct me.
$0.70/day is still substantially more than the electricity cost if you need a mac for work but aren't using it 24x7. $200/yr is a good top-up to the beer money fund.
a) ML Compute doesnt (today) use the neural engine. It uses Accelerate.framework's BNNS module and Metal on GPU - meaning its CPU and GPU only.
b) The Apple Neural Engine (ANE) is a private inference accelerator which has no programmable API. Its only accessible via CoreML ML Program or ML Model format, and the CoreML runtime May, or May Not™ decide to run your model on the ANE depending on system load, other tasks, battery / power, and the phase of the moon.
That is on a extremely small network (from a 1998 paper) running on an extremely small dataset of small images (28px by 28px grayscale images) and they are comparing the ms per step which is going to vary dramatically. Not only that they do not bother to run the m1max and simply say it, "should run twice as fast as the M1 Pro"
In other words that benchmark is completely useless. They should run a standard network from the past decade at least (say VGG16) on useful image sizes and should give the 10 epoch training time if they want anything stable that may approximate hobbyist workloads.
It has an API, it is called CoreML. The compiler will use whatever resources will run your code the best, which on a phone may consider power consumption as well as performance. I am doubtful as programmers we know more about the hardware performance tradeoffs than the designers of the hardware.
"Core ML optimizes on-device performance by leveraging the CPU, GPU, and Neural Engine while minimizing its memory footprint and power consumption."
Yes, CoreML may or may not run on the Neural engine, depending upon what you are trying to do and the system conditions. It will run your code on whatever works best.
This benchmark makes no sense. I was looking for some reference points and found forum topic[1] with few PC benchmark results and they are `strange`. Only RTX 3090 benchmark result is around 12k in raster, while RTX 3080's result is 13k. And there are 2060's that run like 1050Ti(both giving result around 5k). And 1660 Super that outputs 10k with 3070 and 2070 giving similar results. There's something seriously wrong with this benchmark.
Performance doesn't scale with memory bandwidth, SM count or any other spec that is known to me. RTX 3090 and 3080 are not just similar cards from the same generation. They are the same GA102 die with some configuration differences : 3090 does have 24gb VRAM while 3080 has only 10. On 3090 82 SMs of initial 84 are active and on 3080 it's 68 SMs. And they both perform only marginally better than 1660 Super, while usually the difference between this cards is in 3x range. It's likely just poorly written software.
“It’s likely just poorly written software” sounds completely glib to me, it sounds like an intuitive leap rather than a reasoned-out conclusion.
Again, it sounds like an unusual workload. I don’t see why performance must scale with some spec in any particular straightforward way—I think we all remembered seeing workloads that would get faster once you added more CPU cores, then slower once you kept adding CPU cores—because back in the day, most of us didn’t have experience with systems that had lots of CPUs.
GPUs are optimized to suit common workloads and, quite predictably, aren’t optimized to suit uncommon workloads. Modern graphics uses a mix of CPU and GPU work. The back-and-forth, especially returning results from the GPU to the CPU, has historically caused all sorts performance problems.
We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive.
> GPUs are optimized to suit common workloads and, quite predictably, aren’t optimized to suit uncommon workloads.
That doesn't explain why there are such drastic performance differences between GPUs on same uarch. It shouldn't perform better on clearly worse GPUs.
> We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive.
I'm not saying that it `must` be poorly written software. I'm just saying that it's `likely`, given niche nature of this software. If performance is affected by some unusual factor they have to clearly indicate that.
Could the transfer speed over PCIE to the video card be the limiting factor then?. It would make sense why the M1 GPU bench very high as well as it likely has 200GB/s read to the shared memory compared to the ~32GB/s given by a PCIE4.0 16x slot.
Mac Pro with w6900x uses PCIe gen 3, which is 2x worse than gen 4. And given other results in that thread, I think it's unlikely to be the bottleneck. If it was, people with ryzen CPUs and 30 series cards would have much higher performance than they have on practice.
That is true, perhaps its a difference in drivers and operating systems?. it hard to tell if the benchmark is bad without having the same setup to test against because anything could impact it really.
The thread specifically says that they bet on fast transfer on and off of the gpu becoming common, and then the exact opposite of that happened. When you care about a performance characteristic that no one else cares about, you're going to get very weird performance results and it's not a given that a gpu better for normal workloads will be better for you.
Transfer on and off the GPU is limited mostly by PCIe and the latest Mac Pro that performs almost like M1 Max in rasterization has only PCIe3.0, while PCIe gen 4 with double the bandwidth is available on many PCs. If you look closely at the benchmarks performance doesn't correlate with PCIe bandwidth, and even if it did, it doesn't explain nearly 3x difference between 3090 and W6900X.
IIUC on the M1, transfer happens using unified memory. This should mean the bandwidth is not limited by pcie but rather by the memory bandwidth, which was quite high if I remember correctly. I
Memory bandwidth in w6900x is roughly 500gb/s. In 3090 it's more than 900gb/s. Yet, 3090 performs like 5500M(which is laughable). In M1 Max bandwidth is 400gb/s(but there is 200gb/s version). I'm comparing only raster performance which should be heavily GPU dependent.
The difference is that with the w6900x, the memory bandwidth is not used to transfer data between the dGPU and main memory. That happens over PCIe instead.
Anyone know what GPU "codegen" it's using on PC? Tech specs say GPU needs Direct3D 12.0 capable card, but I'm not convinced that's for GPU compute, but maybe it is a compute shader?... I'd assume OpenCL for cross-platform (but then if they're specialising Metal for MacOS maybe they don't care about that, and are using Direct3D.
Because I was confused with these benchmark results I tried to run it myself. And learned that current benchmark version and benchmark in affinity forum topic are not comparable. But still there results for that older benchmark version are quite strange. And for the reference: my 1080Ti outputs 13k, but there is huge variance run to run.
The forum features people running multiple GPU's using onboard graphics as a secondary GPU and improving performance drastically, yet a few cases where the onboard GPU really hurt the benchmark. Some other people are seeing modest improvement, others are doubling their scores with this multi-gpu feature.
Yeah... I'd take this benchmark with a grain of salt.
Affinity photo is a photoshop alternative. It is supposed to be quite good, and is exactly the type of workload I would expect to see running on a MacBook Pro. It’s only indicative of the performance of that one workload, but that is still quite useful IMO.
I'm curious, and this is a side question, but on the topic of energy use, is it more likely that maxing out a GPU or maxing out a CPU will have the biggest power draw, and therefore affect energy bills more or run battery lower?
For some examples, in my laptop the CPU is officially a 45W part though it can be configured to draw up to 70W. On the other hand, my GPU can be set to draw 115W.
In my desktop, it's a 65W (AMD) CPU and 230W GPU.
For an AMD consumer desktop, you wouldn't expect much higher than 125W power consumption from a CPU.
But the new Alder Lake Core i9 advertises power draw as high as 241W.
So you can end up with a more demanding CPU than GPU if you try.
Yup, and you won't get those specific numbers, so it may be helpful to use numbers that are available in other systems. Additionally, if you look at a chip diagram, you can see how much more of the transistor budget is used for the GPU cores than the CPU.
85 comments
[ 3.1 ms ] story [ 173 ms ] threadWhen I click on the link, I get taken to a page on twitter.com that shows me a post, in normal linear order. It reads to me like ordinary English text, in the order you ordinarily read or write it in. The text is not interrupted, except by the little "replies / retweets / likes / share" bar that appears after every tweet.
Not to mention that on the desktop, the text is literally only quarter of the screen. Add that to the very frequent interrption bars, you are in for a very long scroll while reading
A quarter of the width? Well, naturally! The optimal line length for readability is somewhere in the 50-100 character range. This is one of those things that is actually backed by empirical research (lots of research... experiments in readability are cheap and easy to conduct, and people have been doing these experiments for a long time). The width of lines of text in tweets is on the shorter end of that range.
Hacker News, for example, is an example of a site with poor readability. The text is as wide as the window. To address this deficiency, I open Hacker News in a narrower window. It's still harder to read.
Not with CSS, their markup is impenetrable. You have to have another site scrape and reformat the content. Shameful.
Because it's highly that even if it's correct in aggregate, there will be outliers that the opposite is true, and they'll benefit from different formatting.
That is, even if it's right, it might still be wrong for them. So the question being asked is, why is it made so hard to change?
Like, you're talking about "what's right for them" as if that's supposed to pull my heartstrings, but the grandparent even literally said "I am going to re-phrase your question negatively", it's not like they're going to magically be convinced or actually argue about any readability metrics in any kind good faith, it's literally just a reflex. I wouldn't waste my time defending it on some egalitarian principles or whatever.
Only on a small screen. An HN `div.comment` has a `max-width: 1215px;` rule.
Also, replies to tweets within one of these multiple-tweet posts are hard to even get to, and I genuinely cannot figure out why Twitter used that design. It can’t help engagement.
Short of interspaced adverts or popovers, what could be done to make it worse?
It could have super-wide text, like Hacker News. That would make it worse. The line length is a fairly comfortable length, unlike Hacker News’s dizzying long lines.
It could have unusual scrolling behavior, like those Apple product announcements. That would make it worse.
You’ve mentioned the biggest thing that makes websites worse—interspaced adverts and popups, which are downright ubiquitous now. Nearly every site I visit has those.
Honestly, it seems very easy to read for me.
https://threadreaderapp.com/thread/1451859111843356676.html
b) I just checked with everything disabled, I got 2 small ads over the whole thread, what did you get?
they cost money on iPhone.
And (to the sibling reply) people shouldn't use a browser without ad blockers in 2021 :-)
So no, not cost effective. Perhaps if better software is written specially for the hardware / Metal?
a) ML Compute doesnt (today) use the neural engine. It uses Accelerate.framework's BNNS module and Metal on GPU - meaning its CPU and GPU only.
b) The Apple Neural Engine (ANE) is a private inference accelerator which has no programmable API. Its only accessible via CoreML ML Program or ML Model format, and the CoreML runtime May, or May Not™ decide to run your model on the ANE depending on system load, other tasks, battery / power, and the phase of the moon.
See my post on SO: https://stackoverflow.com/questions/58437789/coreml-mlmodelc...
https://github.com/apple/tensorflow_macos/issues/25
https://forums.macrumors.com/threads/apple-silicon-deep-lear...
It is expected that the M1 Max should have similar performance to a RTX-2080 or Titan X for that task at least.
In other words that benchmark is completely useless. They should run a standard network from the past decade at least (say VGG16) on useful image sizes and should give the 10 epoch training time if they want anything stable that may approximate hobbyist workloads.
"Core ML optimizes on-device performance by leveraging the CPU, GPU, and Neural Engine while minimizing its memory footprint and power consumption."
https://developer.apple.com/documentation/coreml
Yes, CoreML may or may not run on the Neural engine, depending upon what you are trying to do and the system conditions. It will run your code on whatever works best.
[1] https://forum.affinity.serif.com/index.php?/topic/124022-ben...
Again, it sounds like an unusual workload. I don’t see why performance must scale with some spec in any particular straightforward way—I think we all remembered seeing workloads that would get faster once you added more CPU cores, then slower once you kept adding CPU cores—because back in the day, most of us didn’t have experience with systems that had lots of CPUs.
GPUs are optimized to suit common workloads and, quite predictably, aren’t optimized to suit uncommon workloads. Modern graphics uses a mix of CPU and GPU work. The back-and-forth, especially returning results from the GPU to the CPU, has historically caused all sorts performance problems.
We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive.
That doesn't explain why there are such drastic performance differences between GPUs on same uarch. It shouldn't perform better on clearly worse GPUs.
> We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive.
I'm not saying that it `must` be poorly written software. I'm just saying that it's `likely`, given niche nature of this software. If performance is affected by some unusual factor they have to clearly indicate that.
Yeah... I'd take this benchmark with a grain of salt.
https://www.hardwaretimes.com/intel-alder-lake-mobile-flagsh...
It's a mobile CPU with TDP of 35-45 watt. M1 has TDP of 30 watt.
Single Core:
M1 Max: 1785
12900HK: 1851
Multi Core:
M1 Max: 12753
12900HK: 13256
In my desktop, it's a 65W (AMD) CPU and 230W GPU.
For an AMD consumer desktop, you wouldn't expect much higher than 125W power consumption from a CPU.
But the new Alder Lake Core i9 advertises power draw as high as 241W.
So you can end up with a more demanding CPU than GPU if you try.
https://images.anandtech.com/doci/17019/M1MAX.jpg
From https://www.anandtech.com/show/17024/apple-m1-max-performanc...