CPU Node Names, Intel Packaging, Alder/Meteor Lake Expectations, Q&A - The Full Nerd ep. 185
video description
Date: 2022-03-15
Related videos
Comments and reviews: 10
Verne
Pure processing-wise, I don't think consumers need more than about eight cores. Go much past that point, and unless you're slogging through transcodes or video editing or doing something that's worthwhile to multithread and you just have a bunch of idle cores.
I have an 2020 iMac 5K with a core-i9 (10910) 10 cores/20 threads and unless I'm doing something which is designed to multithread, most are just sitting idle sucking up power.
Really heavy multiprocessors are simply designed that way because manufacturers have trouble making each core _faster, _ and to increase the processing power they simply carbon copy the same cores and depend on software wizardry to make up for their engineering deficiencies.
There are (currently) three ways to make your cores faster: higher chip densities, higher clock speeds (which exponentially increase energy use and also increase cooling requirements, or wider cores which along with long instruction queues, massive reorder buffers, and massively redundant arithmetic units allow the out-of-order execution unit to execute instructions in parallel.
Apple's A14 and M1 Firestorm cores have a 690 instruction execution queue, a massive reorder buffer, and an out-of-order instruction scheduler which can execute up to eight instructions _simultaneously. _ That's how they get such fantastic single core speeds, and why four of the 3. 2ghz Firestorm cores can run rings around their Intel counterparts. It's also why they feel so much snappier than any Intel computer, because _all_ processes can benefit from a fast single core processor vs. those which can be easily multithreaded (which are seldom UI facing and are things like transcoding where work can be parceled out to multiple cores by segmenting the video source between keyframes.
The problem is due to decoder complexity, variable instruction architectures like x86 can only be a maximum of three or four wide for a subset of instructions because the decoder has to guess where each new instruction begins, and if it guesses wrong all subsequent work is irrelevant. To get a wider CPU core, you _need_ to run a fixed instruction length architecture.
Intel has all but exhausted pushing clocks up - the 10900K ran so hot they had to maintain 14nm and shave the top of the chip to produce sufficient surface area to cool the chip - so their only hope for higher speed seems to be to push process shrink.
I suppose they could try the big/little core gambit, but I'm not at all sure that the little cores can run the whole instruction set and cache coherency and dispatching can be a b-tch.
Apple has a hardware dispatcher - does Intel have something similar?
reply
Pure processing-wise, I don't think consumers need more than about eight cores. Go much past that point, and unless you're slogging through transcodes or video editing or doing something that's worthwhile to multithread and you just have a bunch of idle cores.
I have an 2020 iMac 5K with a core-i9 (10910) 10 cores/20 threads and unless I'm doing something which is designed to multithread, most are just sitting idle sucking up power.
Really heavy multiprocessors are simply designed that way because manufacturers have trouble making each core _faster, _ and to increase the processing power they simply carbon copy the same cores and depend on software wizardry to make up for their engineering deficiencies.
There are (currently) three ways to make your cores faster: higher chip densities, higher clock speeds (which exponentially increase energy use and also increase cooling requirements, or wider cores which along with long instruction queues, massive reorder buffers, and massively redundant arithmetic units allow the out-of-order execution unit to execute instructions in parallel.
Apple's A14 and M1 Firestorm cores have a 690 instruction execution queue, a massive reorder buffer, and an out-of-order instruction scheduler which can execute up to eight instructions _simultaneously. _ That's how they get such fantastic single core speeds, and why four of the 3. 2ghz Firestorm cores can run rings around their Intel counterparts. It's also why they feel so much snappier than any Intel computer, because _all_ processes can benefit from a fast single core processor vs. those which can be easily multithreaded (which are seldom UI facing and are things like transcoding where work can be parceled out to multiple cores by segmenting the video source between keyframes.
The problem is due to decoder complexity, variable instruction architectures like x86 can only be a maximum of three or four wide for a subset of instructions because the decoder has to guess where each new instruction begins, and if it guesses wrong all subsequent work is irrelevant. To get a wider CPU core, you _need_ to run a fixed instruction length architecture.
Intel has all but exhausted pushing clocks up - the 10900K ran so hot they had to maintain 14nm and shave the top of the chip to produce sufficient surface area to cool the chip - so their only hope for higher speed seems to be to push process shrink.
I suppose they could try the big/little core gambit, but I'm not at all sure that the little cores can run the whole instruction set and cache coherency and dispatching can be a b-tch.
Apple has a hardware dispatcher - does Intel have something similar?
reply
Bert
-5: 00 If we Dutch look at the belly size of the average middle aged American, we know to avoid US food. We destroy potato tops/leaves, because they contain the toxin solanine.
It is better, we stick to our 2 most popular national dishes: Stew Sauerkraut with Smoked Sausage and Nassi Goreng with spicy Satay Sauce, that is why we are the largest and one of the healthiest people in the world: ): )
-30: 00 Industry standards are great, it allows me to build a computer from parts and it allows me in 2021 to use a 2003 Pentium 4 HT with IDE HDDs as backup server for my 2019 Ryzen desktop with nvme-SSD using 1 Gbps Ethernet: )
-48: 00 That problem with some additional instructions in a big core is a non-issue. The first time an unknown instruction is executed, it will generate an interrupt and the software knows from that moment on, that that process can be run on the big cores only and it will flag the process accordingly. The software will probably implement at least two separate queues for processes ready to run, one for the big cores only and one for all cores. Dependent on the CPU load you could add preferences for big or small cores. You could add a facility at process creation to flag a process as big core only for some critical real time processes. The fun will come, if we start using more specialized cores in one chip or if we add management CPU cores to graphics cards. RISC-V already allows to design specialized CPUs for certain App domains. NVIDIA tries to buy ARM for a reason, I think related to the delegation of some more general processing to a CPU core on the video card. It would be a logic next step after direct IO.
reply
-5: 00 If we Dutch look at the belly size of the average middle aged American, we know to avoid US food. We destroy potato tops/leaves, because they contain the toxin solanine.
It is better, we stick to our 2 most popular national dishes: Stew Sauerkraut with Smoked Sausage and Nassi Goreng with spicy Satay Sauce, that is why we are the largest and one of the healthiest people in the world: ): )
-30: 00 Industry standards are great, it allows me to build a computer from parts and it allows me in 2021 to use a 2003 Pentium 4 HT with IDE HDDs as backup server for my 2019 Ryzen desktop with nvme-SSD using 1 Gbps Ethernet: )
-48: 00 That problem with some additional instructions in a big core is a non-issue. The first time an unknown instruction is executed, it will generate an interrupt and the software knows from that moment on, that that process can be run on the big cores only and it will flag the process accordingly. The software will probably implement at least two separate queues for processes ready to run, one for the big cores only and one for all cores. Dependent on the CPU load you could add preferences for big or small cores. You could add a facility at process creation to flag a process as big core only for some critical real time processes. The fun will come, if we start using more specialized cores in one chip or if we add management CPU cores to graphics cards. RISC-V already allows to design specialized CPUs for certain App domains. NVIDIA tries to buy ARM for a reason, I think related to the delegation of some more general processing to a CPU core on the video card. It would be a logic next step after direct IO.
reply
Eternalduoae
Ugh, i know it's a pedantic comment but i hate the way people use -optics- to describe the way things look instead of talking about the -visual-.
Optics are what you look through, not how the thing you're looking at appears.
If something looks bad, your eyes aren't the problem. It's not the fault of a telescope if the meteor is crashing down on you.
Yes, i know this comes from vacuous political discourse but i expect better from the MT/s guy. and Mr. I refuse to say -bigLITTLE-; .
Traditionally, you'd have said -such and such is a bad look-. That's correct. But saying -such and such is bad optics- is mixing metaphors and implies that YOU cannot see what the reality is for the subject in question.
reply
Ugh, i know it's a pedantic comment but i hate the way people use -optics- to describe the way things look instead of talking about the -visual-.
Optics are what you look through, not how the thing you're looking at appears.
If something looks bad, your eyes aren't the problem. It's not the fault of a telescope if the meteor is crashing down on you.
Yes, i know this comes from vacuous political discourse but i expect better from the MT/s guy. and Mr. I refuse to say -bigLITTLE-; .
Traditionally, you'd have said -such and such is a bad look-. That's correct. But saying -such and such is bad optics- is mixing metaphors and implies that YOU cannot see what the reality is for the subject in question.
reply
Declineto
Windows has the MMCSS which already discriminates between different kinds of processes and manipulates priorities in the scheduler accordingly. In particular games and audio can't run without stuttering without the MMCSS running. Meanwhile the Windows kernel has had big little support since Windows 8. So, it's not farfetched that Windows could work with big little Intel processors just as well as Android and IOS by just extending the functionality of services already in Windows. But more likely, Microsoft will end up increasing the number of services manipulating the scheduler's process priority queue behind its back.
reply
Windows has the MMCSS which already discriminates between different kinds of processes and manipulates priorities in the scheduler accordingly. In particular games and audio can't run without stuttering without the MMCSS running. Meanwhile the Windows kernel has had big little support since Windows 8. So, it's not farfetched that Windows could work with big little Intel processors just as well as Android and IOS by just extending the functionality of services already in Windows. But more likely, Microsoft will end up increasing the number of services manipulating the scheduler's process priority queue behind its back.
reply
excited
They are really sending their shill out to push this. The numbers are not measuring the density, but the minimum feature SIZE. You can draw 2 thick lines close together, or you can draw 2 thin lines slightly further apart in the same space. The line spacing is not what determines the performance of a transistor though, the line thickness does much more. Parking 2 cars closer together doesn't make the cars faster.
reply
They are really sending their shill out to push this. The numbers are not measuring the density, but the minimum feature SIZE. You can draw 2 thick lines close together, or you can draw 2 thin lines slightly further apart in the same space. The line spacing is not what determines the performance of a transistor though, the line thickness does much more. Parking 2 cars closer together doesn't make the cars faster.
reply
excited
It wouldn't make sense to move past 16 cores for mainstream processors when most people are still on 8 and the workloads are not there to take advantage of more cores. It made sense to move past 4 because people needed more performance for the workloads. For those that need more cores there are consumer products with up to 64 cores. Ian has become Intel shill number 1.
reply
It wouldn't make sense to move past 16 cores for mainstream processors when most people are still on 8 and the workloads are not there to take advantage of more cores. It made sense to move past 4 because people needed more performance for the workloads. For those that need more cores there are consumer products with up to 64 cores. Ian has become Intel shill number 1.
reply
Jason
My opinion looking at all the advanced feature that were being developed within the last few windows builds under windows 10 what is now being seen in win 11 is likely going to be a universal platform.
reply
My opinion looking at all the advanced feature that were being developed within the last few windows builds under windows 10 what is now being seen in win 11 is likely going to be a universal platform.
reply
theexmann
What this conversation didn't mention is that AMD redefined traditional CPU market segments based on workload/price points with the introduction of Ryzen. Threadripper is a perfect example of that.
reply
What this conversation didn't mention is that AMD redefined traditional CPU market segments based on workload/price points with the introduction of Ryzen. Threadripper is a perfect example of that.
reply
David
Transistor density metrics won't get standardized until shopping around for a process node gets commoditized. While the market is 100 companies in the world, they'll stay proprietary definitions.
reply
Transistor density metrics won't get standardized until shopping around for a process node gets commoditized. While the market is 100 companies in the world, they'll stay proprietary definitions.
reply
Jelipe
Gordon I-m going to try and predict the future. the internet totally turns against intel and reviewers using ddr5 with alder lake reviews.
Thanks for another great episode
reply
Gordon I-m going to try and predict the future. the internet totally turns against intel and reviewers using ddr5 with alder lake reviews.
Thanks for another great episode
reply
Add a review, comment
Other channel videos
