RISC-V
 

...appears to have some established RISC players rattled:

[url=https://www.theregister.co.uk/2018/07/10/arm_riscv_website/]Up in arms! Arm kills off its anti-RISC-V smear site after own staff revolt[/url] • The Register

Many of engineers in ARM feel bad about that sad story.


That made me furious and ashamed of working for ARM. There's no excuse for such stupidity and behavior.

K210 processor
 

I was just browsing around looking at RISC-V implementations, and I found this interesting bit of hardware: The K210 processor

Some overall specs/features from product pages:
28nm process dual core RISC-V 64-bit IMAFDC @ 400Mhz (clockable up to 800Mhz)
8MiB on-chip "high speed" SRAM
FPU with SP & DP support
FFT accelerator
KPU (Neural network processor)
"8MB/16MB/128MB Flash" ??? *not sure exactly how much flash space the modules actually come with

From what i've gathered it consumes roughly 0.25W power at stock clock rates.

Its mainly aimed at image/video processing, AI applications, but might be a nice low power and low cost alternative to RPi for some things.

One of the [url=https://www.seeedstudio.com/Sipeed-MAix-BiT-for-RISC-V-AI-IoT-p-2872.html]dev boards @ $13[/url]

There is a Raspberry Pi hat being released (not yet available I think)
[url]https://forum.seeedstudio.com/viewtopic.php?f=115&t=31497&sid=242925f320b9c56ab5f65a3778f48af7[/url]

You can get a sort of [url=https://www.seeedstudio.com/Sipeed-MAIX-I-module-w-o-WiFi-1st-RISC-V-64-AI-Module-K210-inside-p-3210.html]barebones module board for $8 w/o wifi[/url] or [url=https://www.seeedstudio.com/Sipeed-MAIX-I-module-WiFi-version-1st-RISC-V-64-AI-Module-K210-inside-p-3206.html]$9 with wifi[/url]

I don't really understand FFT multiplication well enough to know if the "FFT accelerator" on this is useful for any GIMPS related work, but here's some specs from the docs:
[quote=https://s3.cn-north-1.amazonaws.com.cn/dl.kendryte.com/documents/kendryte_datasheet_20181011163248_en.pdf]
The FFT accelerator is a hardware implementation of the Fast Fourier Transform
(FFT).
• 64-point, 128-point, 256-point or 512-point length
• FFT and IFFT operation modes
• 32-bit or 64-bit input data width
• Supports pure-real, pure-imaginary or complex input data
• DMA transfer support
[/quote]


Overall it looks pretty impressive to me in terms of cost and power consumption, what do you folks think?


Some videos:
Crowdfunding Marketing video:
[YOUTUBE]BJTTjiWc0MQ[/YOUTUBE]

NES emulator and some demo micropython stuff:
[YOUTUBE]el6CB-h9Lo0[/YOUTUBE]

Playing Quake 1:
[YOUTUBE]poBBrIWt_HE[/YOUTUBE]

[QUOTE=hansl;522072]Overall it looks pretty impressive to me in terms of cost and power consumption, what do you folks think?[/QUOTE]
Agreed, it looks quite nice for the price!

One thing to note is that memory seems to be limited to the internal 8MB of SRAM.

I couldn't find any information about performance. They mention that the FPU is pipelined and nothing else.

[quote] "8MB/16MB/128MB Flash" ??? *not sure exactly how much flash space the modules actually come with[/quote]The board seems to come with 128Mb of flash that is 16MB (scroll down the product page and look at the picture of the board with list of components).

Here's an interesting talk about lowering energy requirements for highly parallelizable tasks on RISC-V. Again they are targeting AI applications(low precision) but still some neat points about energy efficiency which I think are relevant to Mersenne hunting:
[YOUTUBE]Hz-ay-XBc_M[/YOUTUBE]

Also, about a wekk ago I finally ended up ordering one of those K210 boards and a couple other bits to play around with, but I guess something in my cart was on backorder so I'm still waiting for it to ship...