WHAT IS HBM2 MEMORY?

Putting a CPU and GPU into the same processor package is not a new idea. Intel and AMD have been doing it for years. In fact, they’ve fused the two into the same slice of silicon, not just put them into a shared package, but as yet, neither has created a CPU-GPU product with genuine highperformance credentials, not from a 3D rendering and gaming perspective at least.

Part of the reason for that comes down to power consumption and thermals. Combining two high-performance components concentrates power consumption and heat dissipation into a very small space. It also makes for very large chips that can be expensive to produce.

But all of that is more a problem that prevents competing at the very highest level. As the latest Xbox One X and Sony PlayStation 4 Pro consoles prove, it is possible to create a single-chip solution capable of highperformance gaming. Instead, as far as the PC is concerned, the problem is memory bandwidth. Thus far, the CPU-GPU fusion chips, or APUs as they are also known, have used a shared memory controller for both parts of the chip. Worse, they’ve used a standard CPU spec controller with much less raw throughput than the superwide controllers that are used by discrete graphics chips.

That kind of graphics controller, paired with multiple graphics memory chips, isn’t viable for a single-package product that also houses the CPU. The solution is High Bandwidth Memory, and, more specifically, the second generation version of the technology, known as HBM2. For starters, HBM2 uses layers of memory stacked on a single chip. That means all 4GB of graphics memory in the new IntelAMD package is housed within a single HBM2 chip. That happens to be good for overall package size.

HBM2 also has a superwide memory interface at a full 1,024 bits for each memory die. That compares with the dual 64-bit controllers of mainstream CPUs. That superwide interface allows HBM2 memory systems to offer big bandwidth at relatively low frequencies—perfect for keeping a cap on power consumption and heat dissipation. Just one HBM2 memory stack can achieve 256GB/s of raw throughput for the GPU alone. Existing APUs are limited to around 50GB/s, shared with both the CPU and GPU elements.