It's easy to dismiss a basic article like this, but it's basically a discovery that every Junior engineer will make, and it's useful to talk about those too!
558 karma · joined April 25, 2014
It's easy to dismiss a basic article like this, but it's basically a discovery that every Junior engineer will make, and it's useful to talk about those too!
At a first glace through, it seems to be written at a good blend between concept and implementation followthrough, something that notoriously is not always there for CG papers :) And it's also refreshing to read something that is not neuro-AI-generation of this or that for a change!
It's Risk vs Reward. Anyone who built something great took huge risks, for a huge return.
There is no return here. At best, you've sunk a ton of capital into a low-profit business that is propped up only by government subsidy.
If you instead went on to produce millions of pizzas for people and 30,000lb of flour, that 300lb you used to learn looks like a pretty reasonable investment.
Unrolling trivial loops to remove loop counter overhead hasn't been productive for quite a whole now but unfortunately it's still the default for many compilers.
Scalar code should be unrolled by the compiler to the SIMD word width to expose potential parallelism. But other than that, correctly predicted branches are free, and so is loop instruction overhead on modern wide-dispatch processors. For example, even running a maximally efficient AVX512 kernel on a zen5 machine that dispatches 4 EUs and some load/stores and calculates 2048 bits in the vector units every cycle, you still have a ton of dispatch capacity to handle the loop overhead in the scalar units.
The cost of unrolling is decreased code density and reduced effectiveness of the instruction / uOp cache. I wish Clang in particular would stop unrolling the dang vector loops.
I've never had a single corruption problem in that time, with probably hundreds of thousands of emails. Take that for what its worth.
The only complaint I've ever had is when they redid their UI a year or two ago it got unbearably slow - which improved over the next few iterations until its now fine again.
According to the steam hardware survey for Feb 2025, AVX512 support is already up to around ~16%. That may not sound like a lot, but considering that it's basically all coming from Zen4&5 based CPUs..
Once Intel's next design also supports it, it's probably worth targeting for consumers in relevant markets.
The overall throttling is dynamic and reactive based on heat and power draw - this is unavoidable and in fact desirable (the alternative is to simply run slower all the time, not to somehow be immune to physics and run faster all the time)
The fully investigation in this article is really interesting, but the TL;DR is: Light up enough of the core, and frequencies will have to drop to maintain power envelope. The transition period is done very smartly, but it still exists - but as opposed to the old intel avx512 cores that got endless (deserved?) bad press for their transition behavior, this is more or less seamless.
I found the part about member access operators interesting.
I've written C++ for 25 years, and although I've used "pointer"-to-member(function/fields) ability many times (indeed, this is how you bind signals to your member slots statically in QT!), I had no idea that it could both be null, and that -1 was used for a 'null' 'pointer' in this instance.
Good example about how large the language is, with so many weird dark darks. The number of air quotes needed above is pretty illuminating on its own.
"Beyond this, people make mistakes. Brilliance represents an upper bound on the quality of your reasoning, but there is no lower bound. The most brilliant scientist in the world can take really dumb stances. Indeed, the success that often goes with brilliance can encourage a blind stubbornness. Not always—some top scientists are admirably skeptical of their own ideas—but sometimes. And if you want to be stubborn, again, there’s no lower bound on how wrong you can be. The best driver in the world can still decide to turn the steering wheel and crash into a tree."
It is one of those profound realizations that seems so obviously true it's irrelevant. But then ask if we evaluate the decisions and statements from smart people this way. Generally the answer is no.
While the brilliant person will have higher quality reasoning on average due to the stretching of the distribution... any individual belief or statement they come up with is being drawn from a distribution that still includes boneheadely wrong.
It is not:
https://fdc.nal.usda.gov/food-details/746771/nutrients
In fact an orange is a mixture of glucose, fructose, and sucrose (which itself is glucose and fructose). An orange actually looks surprisingly like table sugar in terms of the sugar types, although you are correct that the pulp / fiber / etc do mean that the response of our body is likely different.
Double integer register count, three-operand instructions, improved predication support including conditional load/stores.. If I remember correctly, they actually implement it by having the scalar instructions use the EVEX prefix introduced with avx512.
I would really love to seem them settle on APX + AVX10/512 as a next-generation x86 baseline.
In particular, I'd really like AMD and Intel to get on the same page in terms of avx10 / avx512 support.
Many people correctly note that avx512 support is not super relevant today, but this can be laid heavily at the feet of Intel's process troubles and a terrible decision to use ISA for market segmentation purposes.
Zen4/zen5 show that it is possible to implement wide vector units on a cpu in an extremely beneficial way - even if you're running reduced width execution units under the hood, the improved frontend performance is really useful - and also actually saves power, as the decoders and schedulers account for a fair chunk of power consumption these days.
Where the balance slants all the way towards gpus again is the tensor units using reduced precision...
For a Zen 5 core, that means 16 double precision FMAs per cycle using AVX 512, so 80gflop per core at 5ghz, or twice that using fp32
The editable fields within a markdown explainer is really intuitive.
What I'd personally like to see is a better data exploration tab or similar - basically some place that makes it easier to view the cells in use in the doc, and edit them. basically a spreadsheet tab ;)
GPU compute performance is both technically interesting, and matters to much more than simply gaming!
I think a lot of folks get mixed up with GPU and/or SIMD architecture, where you execute both sides of the branch out of necessity: Some of the lanes need to go one way and some the other, so you have to do both.