You probably want to buy that DDR4 as soon as you can - memory prices (DDR3, DDR4) have been consistently going up - not down[1]. It's insane. The price-fixing fines the memory manufacturers paid were clearly not punitive enough, we need more anti-trust action in this area.
I recently made a build on a budget, I had to snipe online specials on DDR4; waiting for the prices to come down is not a winning strategy.
http://www.china.org.cn/business/2017-12/23/content_50157479...
https://www.theregister.co.uk/2017/12/21/china_memory_insour...
https://www.reuters.com/article/us-samsung-elec-chips-outloo...
Pretty blatant market manipulation here by the DRAM cartel - and China has enough downstream manufacturing at stake here that they're willing to go to the mat over it.
This commonly touted meme seems to ignore market realities:
https://www.dramexchange.com/WeeklyResearch/Post/2/4879.html
Data-centers buy TBs (not GB, but Terabytes) per machine. And then they buy ~20 machines per rack, and then fill an entire room full of racks.
When Data-centers decide its time to upgrade from old 1.5V DDR3 machines (or 1.35V "low voltage" LDDR3) to 1.2V DDR4 2666 MT/s like they have been doing in late 2017 through (estimated) 1H 2018, it is only natural to expect DDR4 prices to rise.
The power-savings from switching from 1.35V "low-voltage" DDR3 to 1.2V DDR4 are big enough that there are lots of upgrades incoming from the biggest purchasers of RAM in the world.
> waiting for the prices to come down is not a winning strategy.
2H 2018 seems to be the current market predictions. This also works in line with waiting for Zen+ (aka Ryzen 2). If DDR4 prices are still high at that point, I'll reconsider then.
I plan on recycling my old R9 290x until GPU prices come down... but waiting till August 2018 may be long enough for GPU prices to normalize as well.
So, which data-centers are upgrading from DDR2->DDR3 then? I ask because DDR3 prices are also trending upwards.
LPDDR3 has one slight advantage: an advanced sleep state which uses less than 1/10th the power than normal sleep. Laptops, tablets, and cell phones use this sleep state regularly to remain in "suspended mode" for extended periods of time.
DDR4 cannot do this. As such, LPDDR3 still has some demand in the mobile marketplace, despite fewer fab-labs making it. As such, I'd expect DDR3 prices to go up.
LPDDR4 has the advanced sleep state, but is unsupported by Intel / AMD processors.
But Ryzen's "split" L3 cache seems to be great for process-level parallelism (think compile times), and seems to scale to more cores for a cheaper price. They have an Achilles heel though: 128-bit wide SIMD means much slower AVX2 code, and no support for AVX512.
But for most general purpose code, Ryzen looks downright amazing. Even their slower AVX2 issues is mitigated by having way more cores than their competition. AMD sort of brute-forces a solution.
AMD's GPGPU solution looks inferior to NVidia, but for my purposes, I basically just need a working compiler. I don't plan on doing any deep learning stuff (so CuDNN is not an advantage in my case), but I'd like to play with high-performance SIMD / SIMT coding. So AMD's ROCm initiative looks pretty nice. The main benefit of ROCm is AMD's attempts to mainline it. They're not YET successful at integrating ROCm into the Linux mainline, but their repeated patch-attempts gives strong promise for the future of Linux / AMD compatibility.
The effort has borne real fruit too: Linux accepted a number of AMD drivers for 4.15 mainline.
NVidia's CUDA is definitely more mainstream though. I can get AWS instances with HUGE NVidia P100s to perform heavy-duty compute. There's absolutely no comparable AMD card in any of the major cloud providers (AWS / Google / Azure). I may end up upgrading to NVidia as a GPGPU solution for CUDA instead.
OpenCL unfortunately, doesn't seem like a good solution unless I buy a Xeon Phi (which is way more expensive than consumer stuff). AMD's ROCm / HCC or CUDA are the only things that I'm optimistic for in the near future.