2012 MBP Retina — Memory Bandwidth

2012-06-21 updated 2012-07-30 • SEND FEEDBACK •
Related: bandwidth, laptop, MacBook, MacBook Pro, Macs, memory, memory bandwidth

Memory bandwidth correlates slightly with performance in memory-intensive compute jobs. It is an interesting data point, but little more.

The chipset in the MacBook Pro (June 2012) has outstanding memory bandwidth, substantially better than 3 modules in the Mac Pro Westmere 6-core. This keeps the CPU cores humming efficiently.

Suprisingly, the MacBook Pro 15" model cannot achieve the same memory bandwidth as the MacBook Pro Retina, even though both were 2.7 GHz models with the same chipset. Perhaps Apple’s soldered-on memory in the Retina model is somehow more effcient.

This was so suprising that I obtained a 2nd test kit from OWC to rule out a particular pair of modules; I obtained the same results within 1%.

Memory bandwidth via memcpy(), MB/sec, as tested by MemoryTester

Adam W, ASIC design engineer writes:

You seemed to wonder why the memory on the MBP Retina would be so much faster than the memory on the regular MBP or any other machine. You were correct when stating the memory soldered on would be more efficient. When designing a motherboard the manufacturer must adhere to a defined standard for the DIMM that is selected. This includes wire lengths, and capacitance, and results in generally slightly less than ideal settings to make it compatible with a vast array of different manufacturers. By soldering the memory directly on to the board Apple is able to define the exact memory access scheme, as they will only ever have to work with the chips that they selected, and don’t have to meet any standards for the DIMM connectors, etc.

This allows the fine tuning of latency to the absurd levels. In fact, with the 22nm Intel there is an option via low level register programming, to reduce the latency of the DDR memory down to 0 clocks. While this is not possible in reality, it gives Apple the room it needs to fine tune the memory and provide the best throughput possible. The values you were reporting were almost 100% utilization of the FSB of the CPU (16,384 MB/s being 100%). You were reporting 99.90% utilization, which is UNHEARD of, but not impossible.

If nothing else, you must tip your hat to Apple for this, that is an incredible job of matching memory to memory controller, and setting the CPU to properly utilize the memory. You will NEVER see this on a computer that supports DIMMs, it is not possible. So, maybe it is worth the soldered on memory, to get the performance they are reporting. I suspect they didn’t go any higher than 16GB because the capacitive loading of the memory bus would have prevented them from achieving such magnificent throughput. I bet you will see this on later models, but they need their flagship computer to be FAST on the maiden launch. I don’t think there is any doubt that it is FAST.

[I am an ASIC design engineer, and I am very familiar with CPUs, memory, memory controllers, cache, FSBs, etc. I have also worked on NAND/NOR flash, so I can answer technical details about that, too. I just don’t have good software experience, so I can guess on software questions, but not be sure.]

DIGLLOYD: what a helpful explanation and it makes sense!