4 ms·
> Intel's largest mistake was integrating graphics on their CPU. This cost them more than the entire cellphone CPU market is worth. Er… you mean like every si
by pcwalton 11y ago
> Intel's largest mistake was integrating graphics on their CPU. This cost them more than the entire cellphone CPU market is worth.
Er… you mean like every single mobile SoC?
- Retric 11y agoThese chips are 3D stacked with ram, but produced seperatly. Which means if one part is defective you can use a different part boosting yields. But, they also get much lower latency. Now recent Phone Chips have included on die GPU's but phones are so power limited it's not an issue.
- dragontamer 11y agoMobile SoCs use DDR3 or DDR4 RAM, just like Desktops and Laptops. HBM is only being used on AMD's R9 Fury Graphics Card, as far as I'm aware. A rather high-end solution that simply doesn't exist in the phone market right now.
- Retric 11y agoIt's not a question of ram type. Physical distance is a large chunk of desktop RAM latency.
- ajross 11y agoNo it isn't. Physical distance change line capacitance and limits clock rates (or requires higher voltages, same thing). DRAM latency has always been completely dominated by precharge time inside the array; the whole idea of synchronous DRAM was to take advantage of that to stream data across many clock cycles (that is, get a very "wide" interface to that can transfer a whole cache line at once on a few dozen wires) as it's happening.
- Retric 11y agoYou get ~1/2 the speed of light propagating though copper or ~150million meters per second. It's around two feet from CPU to edge of last chip if you trace the wires. So, round trip adds ~1nanosecond to latency. https://en.wikipedia.org/wiki/CAS_latency https://en.wikipedia.org/wiki/CAS_latency Shows latency's in the 6.3ns to 13 ns range for DDR3. Thous ~10% of latency is from distance. PS: I could get more accurate numbers, but with such a wide range of latency number that's pointless.
- dragontamer 11y agoDo you not know what HBM memory is? You're talking about "stacked memory" and stuff, so I thought you'd know. As I noted before: Mobile CPUs are using standard DDR3 or DDR4 RAM, on a motherboard drawn with copper wires to the CPU. The ONLY use of "3d stacked memory" that I know of is the HBM RAM of the AMD Fury X. http://www.legitreviews.com/wp-content/uploads/2015/05/HBM-DRAM.jpg http://www.legitreviews.com/wp-content/uploads/2015/05/HBM-D... I'm laying counter to your fact-claims. Only high-power GPUs have stacked RAM right now. Anyone who has opened up a phone can see the DDR3 or DDR4 RAM on the phone's motherboard, with the phone SOC designed to go "inefficiently all the way" to the DDR3 RAM over "long latency" copper wires. Read the spec sheet yourself if you don't believe me. http://system-on-a-chip.specout.com/l/1107/Qualcomm-Snapdragon-808-MSM8992 http://system-on-a-chip.specout.com/l/1107/Qualcomm-Snapdrag...
- Retric 11y agoYes, it's a similar idea. These terms get repeated SoC used to refer to a single chip CPU, and they keep reusing the term as other components get integrated. Anyway, I am referring to packaging like: http://cdn.arstechnica.net/wp-content/uploads/2013/04/ASIC_+_Memory_PoP_Schematic.jpg http://cdn.arstechnica.net/wp-content/uploads/2013/04/ASIC_+... Yes, HBM also does more or less the same thing. The issues are twofold heat which is vastly less of a problem in the mobile world so with HBM it's a stack of ram chips next to a processer, with Mobil the CPU is under what might only be one RAM chip and they call it SoC. http://arstechnica.com/gadgets/2013/04/the-pc-inside-your-phone-a-guide-to-the-system-on-a-chip/ http://arstechnica.com/gadgets/2013/04/the-pc-inside-your-ph... Note: CPU's are already 3d with many layers depending on how you count. Anyway after die size, heat, latency, and manufacturing costs are really the largest factors. Some PC ram chips run hot enough they really can't be stacked. You can slow them down and trade latency for total bandwidth. ed:Less of an issue with active cooling and a heat sink. However, they can call this HBM, because you can have more chips so even if each chip is slower overall you get higher speeds. PS: Don't forget HBM is designed for a GPU/CPGPU without a large cache. CPU's have a huge cash which let's them have very different approach aka lot's of cheap but slow RAM. Benchmarks show faster ream really does not add much for a modern CPU with most workloads. You could use this for a CPU but it's harder to swap in more ram.