Hi, folks. Does anyone know what DDR chip is placed on Unmatched board? I tried to set up DDR speed in u-boot up to 1000 MHz and my board started working incorrectly and Linux curshed. In a chip specification, it is said that the board is capable to work up to 1200 MHz (2400 MT/s). It looked like a DDR chip is not capable to work at this speed(
I don’t know what document you are looking at. You didn’t specify. Both the Unmatched Product Brief and the Unmatched Datasheet on our web site
say 1866 MT/s.
I don’t know what u-boot change you tried. You didn’t specify. You can’t change DDR speed in u-boot unless you have an NDA with Cadence and a lot of specialized knowledge and equipment for doing DRAM training.
2400 MT/s you can see at 1.12 DDR memory subsystem in this document https://sifive-china.oss-cn-zhangjiakou.aliyuncs.com/HiFIve%20Unmatched/fu740-c000-manual-v1p2.pdf. So it seems that this chip can work at 2400 MT/s.
DDR speed you can change in u-boot dts clock-frequency = <933333324>; (u-boot/fu740-c000-u-boot.dtsi at master · u-boot/u-boot · GitHub, line 80). This is ddrclock and it is 1/2 of MT/s.
2400 MT/s looks like the theoretical max for the DDR controller that we licensed, but the actual max frequency will depend on the SoC and board and DDR memory. The production Unmatched board runs at 1866 MT/s.
You can’t change the DDR speed by just changing a DTS entry. You need to also change the data used to program the DDR controller, and we only have data for 1866 MT/s in u-boot.
Keep in mind that SiFive (excepting OpenFive) is not a hardware company. We are an IP Core business. Making SoCs and boards is not a core part of our business, and we are not experts in this area. I would expect better SoCs and boards from actual hardware companies. Unfortunately the market isn’t there yet, so SiFive is producing SoCs and boards until our customers do.
The U74-MC (everything from the cores to the L1 and L2 cache) seem to be very good, but yeah the FU-740 SoC seems to let it down. The Allwinner D1 seems much better performing as an SoC, but then is of course let down by having just one single-issue core.
It’s going to be very very interesting to see how the Xuantie 910 board (apparently the SoC is called “ICE”) I should have in mid November will perform. The OoO cores have very good performance numbers in the Hot Chips presentation from a year ago – will they hold up in reality?
I hope someone’s working on a U84 board for 2022.
Ok, I got you. Many thanks)
CPU have serious memory speed problems with sequential access of large memory blocks: Memory access is too slow. RAM seems capable to run on much higher speed.
Good to know. As the CPU clock frequency is so easy to change this was something I really wondered