Depends if it finally adopts SVE or not. AMD64 is now officially more energy efficient than ARM thanks to AVX-512. SVE is the direct competitor to AVX-512 but AFAIK it's not available on consumer ARM yet.
>"AVX-512 based argsort algorithms using O(1) space for 32-bit and 64-bit data types. It returns the indices that would sort an array. These are up to 6x faster when compared to a scalar solution that uses std::sort. This new feature is leveraged by NumPy in np.argsort in its latest release v1.25."
>"AVX-512 based quick select algorithm for 16-bit, 32-bit and 64-bit data types. These are equivalent to std::nthelement but performs a lot faster. The performance depends on the ratio K/N (where K is the index of the element the array is partitioned around and N is the array size). For smaller values of K, it is up to 6x faster for 64-bit data, up to 15x faster for 32-bit data and up to 7x faster for 16-bit data. The performance gets better as K gets larger. "
>"AVX-512 based partial sort algorithm for 16-bit, 32-bit and 64-bit data types. These are equivalent to std::partial_sort. As with quick select, its performance depends on the ratio K/N (where K is the size of partial sorted array and N is the array size) and it tends to perform a an order of magnitude faster for larger values. It is about 1.05x faster for tiny partial array sort and up to 20x faster for slightly larger partial arrays. "
>"AVX-512 sort for _Float16 data type using AVX-512 FP16 ISA. In NumPy, these are nearly 3x faster than AVX-512 based sort that emulates float16 data type."
The point is ARM is now officially no longer more energy efficient than AMD64 in performance per watt in software that uses AVX-512. Python using it is SERIOUS BUSINESS because retard devs no longer have to do any assembly programming to take advantage of it.
cool python programs using numpy sort finish 0.2ms earlier
compare the battery life while running actual software forget microbenchmarks
3 weeks ago
pixDAIZ
>"For smaller values of K, it is up to 6x faster for 64-bit data" >"up to 15x faster for 32-bit data" >"up to 7x faster for 16-bit data." >"The performance gets better as K gets larger. "
Intel did a 360 quickscope noscope regarding AVX-512. Why do you think that is?
do you understand that a program will do about 500 billion operations and sorting a vector is only 1 of them? you can make something that takes 1 milisecond a billion times faster its not going to be noticeable if you only use it once
its not going to make your battery life go from 3 hours to 12 much less with no software using it because its barely available in any cpu so no ones going to make use of it for a decade
hell most current software doesnt even use avx2 and its over a decade old
Didn't all the processor engineers leave Apple? They go where the money is and Apple determined their efforts to buy out talent didn't work. It made some content creators happy but t hey were already buying Apple. It didn't attract serious engineers or gay gamers and it never will.
the lead designer made a startup (nuvia) and took a bunch of apple engineers with him then sold it to qualcom for a billion
qualcom still has nothing to show for it
3 weeks ago
pixDAIZ
see
I don't know what to tell you senpai. AVX-512 in the past would either cause power consumption to skyrocket or CPU frequencies to plummet to the point where AVX2 would sometimes be faster. No longer the case every since AMD fixed it on Zen 4. Because of that AVX-512 is an actual happening now. Enough for intel to splice it back into their estrogen cores.
>"AVX-512 based argsort algorithms using O(1) space for 32-bit and 64-bit data types. It returns the indices that would sort an array. These are up to 6x faster when compared to a scalar solution that uses std::sort. This new feature is leveraged by NumPy in np.argsort in its latest release v1.25."
>"AVX-512 based quick select algorithm for 16-bit, 32-bit and 64-bit data types. These are equivalent to std::nthelement but performs a lot faster. The performance depends on the ratio K/N (where K is the index of the element the array is partitioned around and N is the array size). For smaller values of K, it is up to 6x faster for 64-bit data, up to 15x faster for 32-bit data and up to 7x faster for 16-bit data. The performance gets better as K gets larger. "
>"AVX-512 based partial sort algorithm for 16-bit, 32-bit and 64-bit data types. These are equivalent to std::partial_sort. As with quick select, its performance depends on the ratio K/N (where K is the size of partial sorted array and N is the array size) and it tends to perform a an order of magnitude faster for larger values. It is about 1.05x faster for tiny partial array sort and up to 20x faster for slightly larger partial arrays. "
>"AVX-512 sort for _Float16 data type using AVX-512 FP16 ISA. In NumPy, these are nearly 3x faster than AVX-512 based sort that emulates float16 data type."
> apple releases m1 > ZOMG GUYS! INTEL IS BANKRUPT AND FINISHED! INDUSTRY WILL BE MOVING TO ARM! WOOP WOOP > performance is woeful, hardly any software was ready at time of launch > apple releases m2 > ZOMG GUYS! THIS TIME INTEL IS BANKRUPT AND FINISHED! ARM IS THE FUTURE GUYS! > m2 sells worse than m1, apple loses marketshare. halts production of m2 for over a month because nobody was buying it > "will it be good"
lmao.
Depends if it finally adopts SVE or not. AMD64 is now officially more energy efficient than ARM thanks to AVX-512. SVE is the direct competitor to AVX-512 but AFAIK it's not available on consumer ARM yet.
>halts production of m2 for over a month because nobody was buying it
Does this mean food isn't good? Because production of that got halted frequently at the time.
Of course, you're leaving out the real reason why, for obvious reasons.
the american woodcock is being displaced by rampant housing developments and all your midget brain can think about is apples latest processor? bro get a grip on reality, creating that thing probably killed off 5 species
Everything but pro res is part of the rumored AMD Zen5 SoCs. With the Xilinx IP AMD now has the best hardware video encoding.
Pro res can be added with no effort.
AMD has the AI accelerators now, they have scalable GPU, they can have more cores.
AMD can scale better for Desktop systems with proper dedicated GPUs and multiple compute dies, how many Apple wants.
They could literally build a 128 or 192 core Apple workstations with 8 GPU accelerators.
AMD has solved all of the scaling issues already in every direction, while Apple can only glue together 2 chips.
AMD's 4nm chips are also more energy efficient.
And I doubt that AMD would charge more than their whole silicon development department costs.
AMD silicon also seems to be smaller and more space efficient than Apple so they would also safe costs there.
The only way AMD will be involved at all is if AMD suddenly starts making ARM chips for Apple (which isn't that unlikely a scenario).
Apple doesn't make ISA changes lightly. They stuck through PowerPC for a full decade despite the absolute clusterfuck that it became once the relationship with Motorola soured.
If they go back to x86, it won't be until the mid-2030s at the earliest.
ISA is overvalued. Apple made it pretty universal and could go back or have them in parallel, like keep the low end one and use AMD custom for higher end SoCs. Their whole ecosystem supports both instruction sets.
>Their whole ecosystem supports both instruction sets.
Only out of necessity. Apple notoriously has a "fuck you" attitude toward backwards compatibility.
As has been pointed out, Apple's whole software ecosystem is now very architecture-neutral. Changing ISAs isn't a complex, multi-year affair like it was in the 1990s/early 2000s.
Didn't all the processor engineers leave Apple? They go where the money is and Apple determined their efforts to buy out talent didn't work. It made some content creators happy but t hey were already buying Apple. It didn't attract serious engineers or gay gamers and it never will.
>We're supposed to hear about it on Tuesday.
>is it going to be good??
holy shit
Depends if it finally adopts SVE or not. AMD64 is now officially more energy efficient than ARM thanks to AVX-512. SVE is the direct competitor to AVX-512 but AFAIK it's not available on consumer ARM yet.
cool what software is compiled to use avx512?
>blender's renderer
>a couple of crypto libraries
>ps3 emulator
maybe in 10 years its actually useful to have a cpu with it
see
The point is ARM is now officially no longer more energy efficient than AMD64 in performance per watt in software that uses AVX-512. Python using it is SERIOUS BUSINESS because retard devs no longer have to do any assembly programming to take advantage of it.
cool python programs using numpy sort finish 0.2ms earlier
compare the battery life while running actual software forget microbenchmarks
>"For smaller values of K, it is up to 6x faster for 64-bit data"
>"up to 15x faster for 32-bit data"
>"up to 7x faster for 16-bit data."
>"The performance gets better as K gets larger. "
Intel did a 360 quickscope noscope regarding AVX-512. Why do you think that is?
https://www.tomshardware.com/news/intels-new-avx10-brings-avx-512-capabilities-to-e-cores
do you understand that a program will do about 500 billion operations and sorting a vector is only 1 of them? you can make something that takes 1 milisecond a billion times faster its not going to be noticeable if you only use it once
its not going to make your battery life go from 3 hours to 12 much less with no software using it because its barely available in any cpu so no ones going to make use of it for a decade
hell most current software doesnt even use avx2 and its over a decade old
the lead designer made a startup (nuvia) and took a bunch of apple engineers with him then sold it to qualcom for a billion
qualcom still has nothing to show for it
see
I don't know what to tell you senpai. AVX-512 in the past would either cause power consumption to skyrocket or CPU frequencies to plummet to the point where AVX2 would sometimes be faster. No longer the case every since AMD fixed it on Zen 4. Because of that AVX-512 is an actual happening now. Enough for intel to splice it back into their estrogen cores.
>ARM is now officially no longer more energy efficient than AMD64 in performance per watt
ftfy
apple arm was barely better than zen3+ on a worse node
>"AVX-512 based argsort algorithms using O(1) space for 32-bit and 64-bit data types. It returns the indices that would sort an array. These are up to 6x faster when compared to a scalar solution that uses std::sort. This new feature is leveraged by NumPy in np.argsort in its latest release v1.25."
>"AVX-512 based quick select algorithm for 16-bit, 32-bit and 64-bit data types. These are equivalent to std::nthelement but performs a lot faster. The performance depends on the ratio K/N (where K is the index of the element the array is partitioned around and N is the array size). For smaller values of K, it is up to 6x faster for 64-bit data, up to 15x faster for 32-bit data and up to 7x faster for 16-bit data. The performance gets better as K gets larger. "
>"AVX-512 based partial sort algorithm for 16-bit, 32-bit and 64-bit data types. These are equivalent to std::partial_sort. As with quick select, its performance depends on the ratio K/N (where K is the size of partial sorted array and N is the array size) and it tends to perform a an order of magnitude faster for larger values. It is about 1.05x faster for tiny partial array sort and up to 20x faster for slightly larger partial arrays. "
>"AVX-512 sort for _Float16 data type using AVX-512 FP16 ISA. In NumPy, these are nearly 3x faster than AVX-512 based sort that emulates float16 data type."
https://www.phoronix.com/news/Intel-x86-simd-sort-2.0
> apple releases m1
> ZOMG GUYS! INTEL IS BANKRUPT AND FINISHED! INDUSTRY WILL BE MOVING TO ARM! WOOP WOOP
> performance is woeful, hardly any software was ready at time of launch
> apple releases m2
> ZOMG GUYS! THIS TIME INTEL IS BANKRUPT AND FINISHED! ARM IS THE FUTURE GUYS!
> m2 sells worse than m1, apple loses marketshare. halts production of m2 for over a month because nobody was buying it
> "will it be good"
lmao.
impressive.
>halts production of m2 for over a month because nobody was buying it
Does this mean food isn't good? Because production of that got halted frequently at the time.
Of course, you're leaving out the real reason why, for obvious reasons.
Will the GPU finally get ray tracing support?
It depends, has TSMC made a new node for it?
It's apparently using TSMC's 3nm process.
Then it will be about as much better as 3nm is better than 4nm
the american woodcock is being displaced by rampant housing developments and all your midget brain can think about is apples latest processor? bro get a grip on reality, creating that thing probably killed off 5 species
the woodcock is fine
that's a cover up the bird is NOT FINE I asked them they said we're going fucking extinct bye
If the bird didn't walk like they were Michael Jackson then maybe those feathery fucks could escape from shit chasing it.
>no cuda
DOA
apple september events are about the iphone
macbooks are always shown 1-2 months after that
also rumors point to no m3 until spring 2024
M4 will be AMD custom. They can't compete.
AMD now has has everything that Apple made to transition to their own silicon.
>go from intel to arm
>go from arm to amd
Apple is retarded, but I don't know if they're that retarded.
Bolting Zen 4 cores onto the rest of the M2 in place of ARM would be an improvement at this point. Apple's architecture is a shit.
Everything but pro res is part of the rumored AMD Zen5 SoCs. With the Xilinx IP AMD now has the best hardware video encoding.
Pro res can be added with no effort.
AMD has the AI accelerators now, they have scalable GPU, they can have more cores.
AMD can scale better for Desktop systems with proper dedicated GPUs and multiple compute dies, how many Apple wants.
They could literally build a 128 or 192 core Apple workstations with 8 GPU accelerators.
AMD has solved all of the scaling issues already in every direction, while Apple can only glue together 2 chips.
AMD's 4nm chips are also more energy efficient.
And I doubt that AMD would charge more than their whole silicon development department costs.
AMD silicon also seems to be smaller and more space efficient than Apple so they would also safe costs there.
The only way AMD will be involved at all is if AMD suddenly starts making ARM chips for Apple (which isn't that unlikely a scenario).
Apple doesn't make ISA changes lightly. They stuck through PowerPC for a full decade despite the absolute clusterfuck that it became once the relationship with Motorola soured.
If they go back to x86, it won't be until the mid-2030s at the earliest.
ISA is overvalued. Apple made it pretty universal and could go back or have them in parallel, like keep the low end one and use AMD custom for higher end SoCs. Their whole ecosystem supports both instruction sets.
>Their whole ecosystem supports both instruction sets.
Only out of necessity. Apple notoriously has a "fuck you" attitude toward backwards compatibility.
eh all operating systems are multiplatform nowadays
As has been pointed out, Apple's whole software ecosystem is now very architecture-neutral. Changing ISAs isn't a complex, multi-year affair like it was in the 1990s/early 2000s.
Didn't all the processor engineers leave Apple? They go where the money is and Apple determined their efforts to buy out talent didn't work. It made some content creators happy but t hey were already buying Apple. It didn't attract serious engineers or gay gamers and it never will.
Wake me up when we are at m10. My m1 macbook pro is still comfy.
No we're not. It was rumored to be announced in October but the new rumor is that it was delayed until next year