4 ms·

> ARM on the other hand seems to thrive on the other end of the spectrum--low power consumption and inexpensive Tell that to the Fujitsu people building the A6

> ARM on the other hand seems to thrive on the other end of the spectrum--low power consumption and inexpensive Tell that to the Fujitsu people building the A64fx chip. (Which by the way turns out uses HBM for system memory. I was dreaming of AMD doing this!!) Or even just to Marvell/Cavium :) I mean, it does thrive in low-power applications, but it's by no means tied to them. A lot of companies going for ARM in the datacenter now: Marvell, Ampere, Huawei, NXP, Mellanox… and Amazon! > I don't know how well the POWER line has progressed since say 2004ish Well, IBM has been building big, powerful and relatively expensive chips. NXP has been slowly sort of getting off the POWER ISA and embracing ARM more. > Wouldn't most solutions require that you cram the IO subsystems into the silicon SoC package? Huh? The whole point of "SoC" is cramming everything onto the one package. It's the most efficient way to do things, and everything from the Raspberry Pi's Broadcrap chip to the biggest AMD EPYC is a SoC. Most desktop systems aren't 100% True™ SoCs as they require a chipset/PCH, but that chipset is very minimal these days.

cartoonworld 7y ago
>Tell that to the Fujitsu people building the A64fx chip. (Which by the way turns out uses HBM for system memory. I was dreaming of AMD doing this!!) Or even just to Marvell/Cavium :) I hear ya, but aren't those efforts kind of like very high-performance controllers for other equipment? Surely the Cavium stuff isn't processing frames directly on the host cpu, instead they leverage the ISA to add support for specialized subsystems, (At requisite cost) right? Plus the advantages that the datacenters are chasing down with this architecture are density related, which is to say a lot of little things doing little jobs. How much of the datacenter push is replacing general purpose, hot cores, with specialized lower power devices? Who's running The Cloud on arm yet? How many ARM cores are we gonna cram into a laptop? (I would like many, please.) Wrt the Fujitsu effort, it seems like an interesting development, but I don't envision that platform hitting desktops. Its possible, but where are the units? You can already buy a Talos II today. >The whole point of "SoC" is cramming everything onto the one package. It's the most efficient way to do things, and everything from the Raspberry Pi's Broadcrap chip to the biggest AMD EPYC is a SoC. Yes of course. But this bloats the power and heat budget of the SoC by more than a little. The biggest AMD EPYC SoC (I really disagree with this generalization of SoC but I understand what you mean) is a huge piece of silicon and not the same kind of ARM ISA lego game. Growing an arm die with specially designed HPC subsystems to that size, also including 128 PCIE lanes, and 10GbE, probably means some really expensive ISA purchases or in-house development, if looking at Fujitsu gives us any clues. AMD and Intel are already experts at managing the power and heat budgets. Reproducing these efforts looks like another steep engineering cost to my eyes. One additional mention, and of coursethank you for your time: Seems you might be interested by some other ARM ISA datacenter efforts, check out the RockPro64 Arm cluster that Pine Microsystems has created to host their online services. Ctrl+F "cluster" in the following link. I don't know where to find more detailed information yet. https://www.pine64.org/2019/11/05/brave-heart-edition-pinephones/ https://www.pine64.org/2019/11/05/brave-heart-edition-pineph...

cartoonworld 7y ago

>Tell that to the Fujitsu people building the A64fx chip. (Which by the way turns out uses HBM for system memory. I was dreaming of AMD doing this!!) Or even just to Marvell/Cavium :) I hear ya, but aren't those efforts kind of like very high-performance controllers for other equipment? Surely the Cavium stuff isn't processing frames directly on the host cpu, instead they leverage the ISA to add support for specialized subsystems, (At requisite cost) right? Plus the advantages that the datacenters are chasing down with this architecture are density related, which is to say a lot of little things doing little jobs. How much of the datacenter push is replacing general purpose, hot cores, with specialized lower power devices? Who's running The Cloud on arm yet? How many ARM cores are we gonna cram into a laptop? (I would like many, please.) Wrt the Fujitsu effort, it seems like an interesting development, but I don't envision that platform hitting desktops. Its possible, but where are the units? You can already buy a Talos II today. >The whole point of "SoC" is cramming everything onto the one package. It's the most efficient way to do things, and everything from the Raspberry Pi's Broadcrap chip to the biggest AMD EPYC is a SoC. Yes of course. But this bloats the power and heat budget of the SoC by more than a little. The biggest AMD EPYC SoC (I really disagree with this generalization of SoC but I understand what you mean) is a huge piece of silicon and not the same kind of ARM ISA lego game. Growing an arm die with specially designed HPC subsystems to that size, also including 128 PCIE lanes, and 10GbE, probably means some really expensive ISA purchases or in-house development, if looking at Fujitsu gives us any clues. AMD and Intel are already experts at managing the power and heat budgets. Reproducing these efforts looks like another steep engineering cost to my eyes. One additional mention, and of coursethank you for your time: Seems you might be interested by some other ARM ISA datacenter efforts, check out the RockPro64 Arm cluster that Pine Microsystems has created to host their online services. Ctrl+F "cluster" in the following link. I don't know where to find more detailed information yet. https://www.pine64.org/2019/11/05/brave-heart-edition-pinephones/ https://www.pine64.org/2019/11/05/brave-heart-edition-pineph...