Just posting some recent news that I found really interesting. I thought it would be nice to share it with my BG brothers here. Please take a look if you haven't heard of these already
Title: PCI Express 4.0
Motherboards outfitted with PCI Express 4.0 will be capable of powering all but the most high-spec graphics cards purely through the motherboard. The motherboard will still need to be supplied with the appropriate power, naturally, but not having to wire in a graphics card should significantly improve air flow. I can almost hear the fans of neat wiring trembling with excitement.
Of the current crop of graphics cards, all should be amply powered by the 300W+ baseline power delivery of PCIe 4.0, while the ceiling is expected to be as high as 500W. Even the mighty GeForce GTX Titan X Pascal only sucks on an estimated 250W TDP. Consider that PCIe 3.0 is 75W and you can see this is a considerable achievement for PCI-SIG.
Aside from this we’re looking at a doubling of bandwidth. At x1 bandwidth it’s running at 1.969 GB/s, up from 984.6 MB/s on 3.0, while PCIe 4.0 x16 runs at 31.508 GB/s compared to 15.754 GB/s on PCIe 3.0 x16.
PCI-SIG is reportedly “"just 'days away' from releasing version 0.7 of the PCIe Express 4.0 specification to members”, according to Tom’s Hardware. The p'an is for version 1.0 of PCIe 4.0 to be available by the tail end of 2016.
Naturally enough with the spec already in some manufacturer’s hands we’re already beginning to see the first products creep out with PCI Express 4.0 support, including a recent ethernet adapter. Where we're really interested in seeing PCIe 4.0 support is of course graphics cards though. Both AMD and Nvidia are members of PCI-SIG so we could see support role out for Nvidia Volta or AMD Navi.
It's not just graphics cards where we can expect to see tangible benefits either. Many people are increasingly using PCIe slots to power high-speed storage like m.2. The additional bandwidth on offer could offer lightning fast storage the likes of which we haven't seen before.
Title: HBM3 Memory
Hot Chips is taking place this week, a conference dedicated to high performance microprocessors. It kicked off yesterday and runs through until tomorrow. There’s already been plenty of juicy news regarding graphics memory, including the first tease for third-generation High Bandwidth Memory, or HBM3. It will supposedly offer twice the bandwidth of HBM2 (which has yet to hit the market for gaming graphics cards), and yet will be cheaper to produce. But the best bit? It’ll allow up to 64GB VRAM in the coming generations of graphics cards.
Up until now our experience with HBM has limited to the first generation seen on AMD’s Radeon Fury graphics cards. At 128GB/s per package it was quite a leap up in terms of bandwidth, while its 3D stacking mean for far smaller form factor graphics cards than GDDR5. HBM2 is set to double the bandwidth to 256GB/s and double the capacity when it presumably arrives alongside AMD’s Vega GPUs, but it’s HBM3 which now sits atop the memory throne. This will offer speeds of 512GB/s per memory stack, along with up to 64GB HBM3 capacity.
The natural evolution of memory is that it gets faster and in greater capacities, so that’s business as usual for HBM3, but where things get interesting is the price. HBM2 is prohibitively expensive, which is probably why we didn’t see it in either Polaris or Pascal, but according to SK Hynix manufacturing of HBM3 will in fact be cheaper. Presumably the manufacturing processes and route to market have been ironed out during HBM and HBM2, driving costs down.
As for when we can expect to see HBM3 - don’t hold your breath just yet. The technology is still in its formative stages and it’s likely we’ll see a generation of HBM2 graphics cards first. As for what use we'll have have for 64GB video memory, well, I'm sure we'll find something. We always do.
Title: DDR5 RAM
HBM3 wasn’t the only exciting memory-related goodness to come out of Hot Chips; we also got some of the first info on DDR5 system memory. The new RAM specification will succeed DDR4 in 2020 if development goes to plan, with the big changes being a nice big bump up to a maximum RAM limit of 128GB, along with a doubling of bandwidth.
As it stands the chief aim of DDR5 DRAM is to offer twice bandwidth at 1.1V, making for a huge leap in memory efficiency. In short this means greater performance while consuming less power, particularly useful for mobile devices such as gaming laptops. Clock speeds will vary from 3200MHz all the way up to 6400MHz once the technology matures. Exact specifications for DDR5 are expected to be released later this year.
Obviously the main thing now is for the likes of Micron and SK Hynix to convince you that DDR5 is something you actually need. DDR4 has practically zero gaming benefits over DDR3, but Mike Howard, director of DRAM at IHS, claimed the"incumbent always has tremendous inertia. The DDR incumbent will keep rolling by the virtue that it's there. A lot of the DDR5 development gets back to the shortcomings for DDR4.”
He goes on to say that gamers will need vaster system memory for VR and other complex computing tasks. Current VR doesn’t seem to have any noticeable problems with bog-standard DDR3, so we’ll have to wait and see on that particular claim. There’s also the small matter of competing standards, with likes of RRAM (resistive RAM), MRAM (magnetoresistive RAM) and PCM (phase change memory) all looking to oust DDR as the de facto standard memory technologies.
The first we’ll see of the physical DDR5 units is in 2018, before production begins in 2019 for the first server RAM modules to roll out in 2020. Expect commercial DDR5 memory to roll out shortly after, probably at an extortionate price which drops incredibly quickly.