AMD Volcanic Islands to Implement CPU Cores

AMD’s upcoming generation of graphics cards, dubbed Volcanic Islands, is to include CPU cores on-chip.

According to website TechPowerUp!, engineers at AMD have been working on implementing CPU cores in their new GPU design. The result would be a clear departure from previous VLIW4, VLIW5 and GCN, which were all evolutionary architectures instead of complte redesigns. Chinese website Chiphell has leaked what is claimed to be a block-diagram of what could be, by current nomenclature, the AMD HD8970. Codenamed Hawaii, the new Graphics Processing Unit is shown implementing two distinct processing modules: A Parallel Compute Module, which looks like a standard GPU, and a Serial Processing Module, which would take care of serial compute tasks as a CPU would.

The classic GPU part would be based on 16 SPUs for a total of 4096 shader processors, along with 256 TMUs (textue mapping units) and 64 ROPs (render output units). Those numbers represent exactly twice as much theoretical processing power as was present in Tahiti XT, the chip behind Radeon HD7970. It is quite unclear how much CPU power will be present on the chip, though, as the block-diagram is fuzzy and no detailed specifications have been posted. What we can observe are 8 modules that resemble AMD’s current high-performance x86 design, with a shared central floating point unit in each module and shared L3 cache, which would add up to a 16-cores CPU. The whole chip, including CPU and GPU parts, is connected to memory through a 512-bit wide bus.

It is worth noting that AMD also signed a licensing agreement with ARM last year, which could mean that the possibility of ARM cores on AMD chips is not to be excluded. We know from previous announcements from Nvidia that their upcoming all-new architecture will also feature on-die CPU, this time confirmed as an ARMv8 part. Given the timeframe, both Maxwell and AMD’s Volcanic Islands series of GPUs are expected to be built on TSMC’s 20nm bulk process. They should bring enhanced power efficiency, thanks in part to the new 20nm process, and ease of programmability through integrated CPU cores; although we can’t know for sure which one of the designs will prevail in each of the new features.

The next round of the exciting and seemingly never-ending graphics war is to unfold in the last weeks of this year, TSMC’s technology allowing.

– dostov