sarinaide :
AMD wants socket stability at least until after the Steamroller arch, maybe revised chipsets and faster DIMMS and PCI-e will be bolted on but the socket 942 will still carry over at least until Steamroller.
-Fran- :
- Hate on AM3+
Well, my personal "hate" (dislike to be more precise) for AM3+ is that the only new thing that it actually brought to the table was SLI support. And that's about it. The socket itself is not "new" at all. and the interconnects are basically the same as previous sockets from the s939 era. That's a good 8 years old now I think and AMD hasn't improved almost nothing in the transport subsystem. Hyper Transport is too damn old and on each CPU jump, they've added just a few extra MB/s worth of bandwidth.
There is nothing wrong with AM3+. The only thing it does not support is an on-CPU northbridge/PCIe controller. That is arguably really only important for small form factor/laptop/SoC type of situations where reducing chip count is critically important for packaging reasons and because it might use a watt or two less power. AMD also makes the FM1/FM2 line of sockets with northbridge/PCIe on board anyway for those that want it. Having an off-die northbridge/PCIe controller is not a bad thing since it gives greater flexibility in your choice of CPU vs. I/O capabilities. If you want more than 20 lanes of PCIe on an Intel platform, you have to step waaaay up in CPU and board price to LGA2011. Either that or spend a bunch of dough on an LGA1155 board with PCIe switches. Getting a midrange AM3+ chip with a 990FX board gives you 42 lanes of PCIe without having to spend $300+ on the CPU for the privilege.
The other things people ding about AM3+ are memory bandwidth, that it uses HyperTransport, that it is based on a socket which debuted in 2006 (AM2), and that the 800/900 series chipsets that ship with the platform are old.
- The only one with any real merit is the last one. AMD could easily update the chipset to something with PCIe 3.0 support and probably will if they continue to use AM3+ for much longer.
- Bandwidth: AM3+ has more bandwidth than LGA1155 as LGA1155 only officially supports two channels of DDR3-1600 vs. AM3+ supporting two channels of DDR3-1866. LGA2011 has more bandwidth with four channels of DDR3-1600 but LGA2011 is really a server platform. Also recent tests by Tom's show that there is very little difference in performance between two and four channels of memory in LGA2011. Also, G34 is LGA2011's real competitor and G34 has more bandwidth with four channels of DDR3-1866.
- HyperTransport: has way more than enough bandwidth to feed the off-die NB/SB on a desktop platform. AMD uses HyperTransport turned up just a fuzz faster (2.6 -> 3.2 GHz DDR) to connect eight-die quad Opteron setups for crying out loud, and many of those connections use split links. Intel also ripped off HyperTransport for its QuickPath interconnect. HT bandwidth/latency has really only been a problem at one short period in time when AMD didn't get HT3 ready in time for the 8-way Barcelona quad-core Opterons. An 8-socket server platform is a far cry from the desktop. And even then once AMD upgraded to HyperTransport 3 with the Shanghais, they were fine again and continue to be fine using HyperTransport 3 with the C32/G34 Opterons.
- "Old socket." Who cares exactly when a socket was conceived when it still works? Most people tend to be frustrated by sockets changing too quickly and giving them no upgrade path at all.
m32 :
^5nm tech? LOL. How many years is that going to take (if it can happen)?
At least a decade if it can happen at all. Quantum tunneling and leakage is supposed to make things below 10-14 nm not viable. But then again way back in the day UV litho was said to be too difficult, then EUV and immersion was said to be too difficult/expensive, so we'll just have to see.
mayankleoboy1 :
HSA = CPU+GPU+misc PU's. So if your code is not scaling to more than 2 CPU cores, how will it scale to ~2000 GPU's smaller cores ?
My impression is that the GPU cores in HSA will largely replace the FPU rather than the CPU. Heavy FP stuff can be made to run on a GPU much easier than CPU stuff can be made to run on a GPU. My hunch is due to the FPU now suddenly being separate from the cores in Bulldozer/Piledriver (makes it easy to in essence point the FP scheduler at the GPU and let 'er rip) and the big successes of GPGPU have all been with heavy FP tasks.