Data center hardware buyers should evaluate two emerging server memory types poised to arrive in future high-performance...
High Bandwidth Memory (HBM) is a high-performance interface intended to support data throughputs for memory devices that push beyond fast memory in conventional forms.
Hybrid Memory Cube (HMC) technology brings memory performance beyond conventional high bandwidth memory designs, such as double data rate third and fourth generation (DDR3 and DDR4), but the two approaches differ technologically and in how they improve server memory performance.
High Bandwidth Memory in future servers
The idea of HBM is straightforward: bring the memory devices closer together and closer to the CPU or GPU. Current computer designs generally segregate memory and computing chips by placing memory modules in groups of slots located on the motherboard. This type of server memory places practical limits on clock frequencies, holding back the amount of data that can move in each clock cycle.
Take a deep dive into server memory types
Now that you've familiarized yourself with these two specific server memory types, take a step back and learn what affects memory performance in the evolving IT space.
The HBM methodology stacks memory chips in a matrix, then assembles the processor and memory stack together to essentially form one component that goes onto a server motherboard.
HBM stacks are not physically integrated with the CPU and GPU, instead using an interposer. However, HBM is indistinguishable from integrated, on-chip (on-processor) memory approaches, according to assertions by proponents such as semiconductor maker AMD.
So how big a deal is HBM? An HBM module can far exceed the bandwidth of conventional memory while operating at a lower frequency and with less power. For example, a typical graphics DDR5 package uses a 32-bit bus running up to 1,750 MHz at 1.5 volts for a bandwidth of up to 28 GB per second. An HBM package uses a 1,024 bit bus running at only up to 500 MHz and 1.3 V to achieve bandwidths over 100 GB, according to AMD research. And HBM offers the versatility to work with CPU or emerging GPU-enabled servers. Future servers may see some permutation of HBM serving both CPUs and GPUs.
How does Hybrid Memory Cube affect future servers?
Conventional server memory types, such as DIMM, use a parallel interface on the server's motherboard to connect individual chips. Hybrid Memory Cube, on the other hand, works by stacking memory chips into vertical assemblies to create 3-D arrays with serial memory access. Those arrays add a logic layer to manage the memory, and the server manufacturer can install each assembly close to the processor. This near-memory or short-reach design is more common and offers higher performance than an alternative far-memory architecture that targets power efficiency.
Hybrid Memory Cubes can be chained together in links of up to eight packages. HMCs offer 15 times the bandwidth, 70% less energy use and a 90% smaller physical footprint than DDR3 memory devices, according to the Hybrid Memory Cube Consortium, a supplier group that promotes the technological standard. For example, Micron's 2 GB and 4 GB HMC product technologies tout bandwidths of 120 GB and 160 GB, respectively. HMC products are available now, and devices like Intel's Xeon Phi graphics co-processor use HMC technology which boasts memory throughput about 50 percent higher than GDDR5 memory devices.
HMC competes with HBM, and the two technologies are not compatible.
Bookmark the guide to server memory choices
Go back to the basics of selecting server memory