Avx512 is no use for gaming anyway, so i don't give a hoot. Benches are just epeen, nothing else. The only thing AVX512 is any good for are serious apps. It seems the only reason to OC is for benches, as a high performance CPU (58/900x, ADL 127/900k) is gonna be fine for most gaming needs anyway.
Name 1 program/game not for professional use that uses AVX512, bet there is not many, so disabling AVX512 on ADL is going to have very minimal impact on it's performance on everyday apps anyway.
It will probably be some time before games start to utilize AVX-512, but there are certainly games which use AVX2, so don't think SIMD isn't useful for games. But it's not something that directly benefit FPS though, so if a game relies heavily on SIMD, it usually means that's a minimum requirement or you have to sacrifice game features if you don't have it.
Even beyond that, several compilers will avoid AVX-512 in favor of AVX-256 because of the issue with downclocking, which impacts all workloads occurring at the time that the CPU clocks down.
Will they? The big three all have offered support for some time.
And compilers don't have a will of their own to decide which ISA to use, that's
specified by the developer.
The downclocking argument is 100% BS and you know it. Even if the core runs a few hundred MHz lower, it will still churn through more data, so this is just nonsense.
The usefulness of AVX-512 completely depends on the workload and how well the task is vectorized. Even for situations that can use AVX-512, going with 256 or 128 might actually net you better performance depending on the application.
Most vectorized data which benefits from SIMD are greater than 512 bits (64 bytes). 512 bits is
tiny.
In fact, using a vector size of 512 bit is
genius, as it perfectly matches the cache line size of current x86 implementations.
Intel are such spoilsports. AMD need to come up with a CPU that does support them, performs better and at a cheaper price than the equivalent Alder Lake to give them a good kicking over it.
I wish AMD would come up with a CPU with AVX-512 support which kicks ass, like 4x FMAs and better energy efficiency. It would be a serious powerhouse.
What I am implying introducing another instruction set like crutch to the X86 won't make the end code smaller and more efficient for us desktop users.
Actually, optimized AVX core is smaller, more cache efficient, not to mention it eliminates a lot of branching, looping, register shuffling and load/stores. If the computational density is high enough, it offers orders of magnitude higher performance. But not all code is that computationally dense, and much of the kernel code is probably not.
If some code monkeys will start to use this instruction set just for fashion as a hipster trend, then bad things usually happen.
First of all, SIMD is used to some degree in many applications. I'm pretty sure you use it every day. Video playback, compression, web browsing (both compression and encryption), video editing, photo editing, etc. all use AVX/AVX2 or SSE. Without it, many of these things would be dreadfully slow. When popular applications start to get good AVX-512 support, you will not want to be left behind.
This time, if it would trigger the instruction and feed too many data through the long instruction pipe, it overheats as it has much longer peak execution and heats up more.
What on earth makes you come up with a claim like that? Stop embarrassing yourself.
Most AVX operations are a few clock cycles, and the work done is equivalent to filling up the pipeline many times.
Remember the early Intel burn test program introducing AVX to them, made them literately furmark class showing temps you never ever see in your daily usage. If programs would trigger it like on gaming, it would be rad and FPS would tank due to single core frequency decrease for the main render thread.
That's not how throttling works at all. This is utter nonsense.
That's PS3 example is just a rare exception. He tries to hammer nails using his shoes by using AVX512, just because it is more lazy and for what, few people using Xeons now?
FYI, AVX-512 is supported by Ice Lake, Tiger Lake, Rocket Lake, Cascade Lake-X and Skylake-X, so not just Xeons.
I think most (if not all) of you have missed the biggest advantage of AVX-512. It's not just AVX2 with double vector size, it's vastly more flexible and have a better instruction encoding scheme. It's much more than just simple fp add/sub/mul/div operations, it actually will allow previously unseen efficiency when implementing dense algorithms, for e.g. encoding, encryption, compression, etc. with an efficiency coming relatively close to ASICs.
, how can it be good. Massive sarcasm and jest I agree with your points.
)
