Earlier this month, SK Hynix updated its product databook for the first quarter of 2017.
Since November, SK Hynix lists 3D-NAND V4 with 72 cell layers and 256 Gbit in its catalog. This memory type was originally planned for Q2 2017, and according to to the latest product catalog, this schedule has not changed. In addition, 512 Gbit type is planned but for a fourth quarter.
FLASH will have a priority over DRAM
Global demand for NAND flash sees a great potential growth for memory manufacturers. The demand is still higher than supply, which unfortunately affects the prices of all memory types on the market. For SK Hynix over 72% of total sales were generated by DRAM, whereas NAND Flash was only responsible for 25% of sales. This proportion will probably not change very soon for Hynix, as recently announced factory construction will not be finished till 2019.
High Bandwidth Memory 2
HBM2 memory has been in Hynix databook for quite some time. However what is hard to explain is the data (oh the irony) it shows. So let’s look at the last three quarters:
Q3 2016 SK Hynix Graphics databook: lists two 4-stack 4GB HBM2 models (2.0 GHz, 1.6 GHz) available Q3 2016.
Q4 2016 SK Hynix Graphics databook: lists one 4-stack 4GB HBM2 2.0 GHz memory type, availability to be determined.
Q1 2017 SK Hynix Graphics databook: lists one 4-stack 4GB HBM2 1.6 GHz, available in first quarter.
So as you can see even Hynix plans can change quite often. The current plan is to release 1.6 GHz HBM2 stack with 4GB memory, and since we already know Vega has two stacks, this gives us 8GB and 409,6 GB/s bandwidth. That’s obviously lower than Fiji, in fact, it’s even lower than Pascal GP102 (480 GB/s), so it’s unclear if AMD is really going to use Hynix memory, or like NVIDIA use Samsung instead.
In the end, it will depend on one thing: how many modules can Vega really have. The chip showcased earlier this month by AMD only had two modules, but nothing should stop AMD from making a 4-module chip, even if it was for professional market only. The only difference would be a bigger interposer.
However, it seems unlikely that Vega would have 8-stack modules, as it was already discovered that the chip showcased by AMD had 8GB ‘memory cache’. If that chip was running 8-stack HBM2, it would be shown as 16GB instead. In fact, even Hynix databook has no information about such modules, so unless Micron or Samsung suddenly announce 8-Hi HBM2 availability, the educated guess for Vega would be 8GB HBM2 1.6 GHz with 410 GB/s bandwidth in total.
What’s also interesting is the lack of GDDR5X in Hynix portfolio. NVIDIA Pascal GPUs are already available with GDDR5X, which are manufactured by Micron. However, it seems unclear whether AMD is also interested in using GDDR5X from Micron for its new Radeon 500 series.
Well, I guess it would be an interesting marketing buzzword if AMD was to refresh Polaris for mid-range 500 series with GDDR5X, otherwise I see no point since Polaris is not really bottlenecked by memory bandwidth.