I’ll start with the pure GPU side of things. As some of you may know, after the GeForce 600 series of graphics cards, Nvidia will begin production and sales of the GeForce 700 series, named Maxwell. The company’s been tight-lipped about Maxwell so far, so according to the new roadmap, it is slated for a 2014 launch instead of the 2013 schedule, so maybe only AMD’s Radeon HD8000 series will get launched in Q3; Maxwell will employ a unified virtual memory address space (like AMD APU in Sony’s PS4).
The second point puzzles me, because using standard system DDR RAM would prove much slower than using GDDR memory found on GPUs. Nvidia’s CUDA technology already uses a unified address space, however this seems to suggest that both compute and graphics cores will get to use RAM, which doesn't seem to have much of an advantage on a PC. Hold on to this thought for a bit though.
Nvidia also mentioned Maxwell’s successor, named Volta, that will feature stacked DRAM on the GPU package itself, sort of like CPUs have an L1, L2 and L3 cache. Intel will implement something similar in the upcoming Haswell chips (only some notebook CPUs will get this graphics DRAM). The point of this is extremely high memory bandwidth and very low latency for certain data sets. Nvidia is aiming for memory bandwidth hitting 1TB/s, which is over 3 times as large as the AMD Radeon HD 7970 GHz Edition and the GeForce GTX Titan (both have a peak theoretical bandwidth of 288 GB/s).
Again, I’m not sure if this is the full story, as you can only get so much memory to sit near the GPU. They’re stacking the DRAM, so maybe they’ll be able to make it dense, but I think that they’ll have to keep additional GDDR memory on standby, which may see lower rates. This is speculation, but, Volta probably won’t see the light of day till 2016.
Let’s move on to the other significant announcement: Tegra. Between CES and MWC there was a lot of talk about Tegra 4, but Nvidia thinks it’s the right time to tell us about Tegra 5 and 6, codenamed Logan and Parker. With Tegra 4, the main features were an integrated LTE modem and real-time HDR processing using the SoC’s compute performance. With Logan, the main feature is that the GPU cores get unified, like they are in their desktop cards.
In Tegra 2, 3 and 4, although Nvidia kept going on about how many GPU cores the SoC has, truth was that these cores were split into two sub-types: vertex and shader cores. So the Tegra 4 SoC has 48 shader cores and 24 vector cores, 72 in total.
The GPUs that make their way into desktop cards have unified vector and shader cores, which Nvidia calls CUDA cores. Logan will be the first TegraSoC to contain a GPU with CUDA cores. Logan’s GPU will be a derivative of the Kepler GPU architecture (found in the GeForce GT/GTX 600 series cards and Titan) and support the full OpenCL 4.3 standard (as opposed to Tegra 4’s cut down OpenGL ES 2.0).
Parker succeeds Logan in 2015, and brings with it Nvidia’s own Denver CPU, most likely based on an ARM architecture. Nvidia currently uses an ARM Cortex design, but this will be like Apple’s Swift core used in their A6 SoC (they’ll design it themselves). Denver will have a Maxwell based GPU, and the manufacturing process is likely to include sub-22nm 3D FinFET transistors (Intel currently uses its 22nm 3D Tri-Gate transistor design for the 3rd Generation Core processors). Denver will likely be 64-bit capable.
Going back to Maxwell’s unified virtual address space. Parker should enable the Denver CPU to access the Maxwell GPU’s on-die DRAM, and enable huge memory bandwidth for both the CPU and GPU. This is what AMD aims to do.
Nvidia also introduced their mini-cloud-computing environment, which they call a Visual Computing Appliance (VCI). It's based on their remote GPU-powered computing GRID concept, and allows small businesses to have a personalised GPU driven compute server. Unfortunately, this article’s already grown very long, so I digress.