DDR4 Memory Technology: What's in a Number?30 Sep, 2014 By: Alex Herrera
Herrera on Hardware: Capable memory is essential for efficient workstation operation. But do you need the newest generation of memory technology?
A chain is only as strong as its weakest link — it's an adage that applies in many contexts, especially computer hardware design. Of the four main components in a workstation, the central processing unit (CPU) and graphics processing unit (GPU) often get the lion's share of attention. But if the remaining two are ignored when configuring a new CAD machine, they can become key bottlenecks that can render moot all the dollars spent on premium CPU and GPU processors. Memory and storage subsystems don't always get the billing they should, but they can be just as critical in determining what users will ultimately see in their day-to-day work.
The fastest processors in the world — whether they're ranked by GHz or floating-point operations per second (FLOPS) or millions of instructions per second (MIPS) — won't accomplish much if they can't read data and write results quickly enough. That need to effectively "feed the beast" is why big and sophisticated on-chip memory caches are commonplace on both CPUs and GPUs — so that the processing elements that need data can be located closer to where the data resides. Still, caches can only get so big, and ultimately processors' accesses need to be served by main system memory. And often, the bigger and multiple datasets common in product design mean CAD workflows can see a higher percentage of accesses requiring system memory cycles. In the end, a high-performance memory subsystem almost always becomes a critical component in a high-performance computer tasked with running a modern, demanding CAD workflow.
Fourth-Generation DDR Technology
Virtually all digital computer processing is synchronized by a clock. Most digital components tend to operate off one of the two edges of a clock, either the rising edge or the falling edge (typically the former). But memories are different. In the effort to squeeze every possible bit per second out of memories, DDR synchronous, dynamic random access memory (SDRAM) drives data on both edges of the clock, a method that is spelled out in the name behind the acronym: double data rate.
A simplified, conceptual view of DDR clocking.
First introduced in 2007, the third generation of DDR technology (governed by JEDEC), appropriately referred to as DDR3, is the current workhorse for workstation, PC, and server memory subsystems. However, with Intel's recently introduced Xeon E5-2600 v3 processor (codenamed "Haswell-E") supporting DDR4, DDR3's successor is now poised to gradually inherit the memory throne for workstations.
Of course, the silicon industry isn't in the habit of introducing new standards and technology unless there's a payoff. First and foremost, DDR4's goal is to increase performance, with secondary goals of improving density and power efficiency. And that leads to the more salient question, particularly for demanding CAD professionals always looking to speed computation and rendering: How much performance gain are we talking about?
Marketing brochures tend to tout big, seemingly clear-cut numbers, but in reality, it's one of those questions that gets the all-too common answer: it depends. First off, any improvement in the capabilities of a memory subsystem will be manifested in shorter compute times only if the workloads are to some degree memory-bound. In the case of CAD, that will often but not always be true, and it will apply more frequently in the case of big design projects. Second, there are two basic metrics typically used to assess memory performance — bandwidth and latency — and in the case of the DDR technology, progress in one dimension doesn't necessarily mean progress in the other.