Alibaba Crafts A 16-Core RISC-V Chip @ 2.5GHz

[Guest]

Anyone willing to analyze this two sentences and write down on how many levels this is wrong (or clueless)?

I'm just trying to shit on every non-copyleft license on every occasion I get.

[tuxd3v]

phoronix maybe you could run coremark on some modern x86 processors as a comparison, in multi-threaded and single-threaded modes, maybe also at a locked clock frequency to more accurately measure coremark/mhz. I would assume coremark is small enough to run entirely in the L1 (or L2?) cache, due to not using much memory since it's a embedded benchmark.
That sounds like it would make an interesting comparison -- seeing how far risc-v or arm is behind ryzen 3000 or intel 9th gen in this admittedly limited fashion.

It would be nice to have Coremark tests on x86 cpus.

I am impressed with C-Sky risc-v implementation..
But no details how they achieved it..

For comparison,
Intel I7-4700HQ has a coremark/Mhz of 7.29( with 1 thread running in 1 core.. )

[tuxd3v]

According to CoreMark score table, an Intel Core i5-8500 with 6 threads running at 3 GHz is scoring 57.21 CoreMarks/MHz. That would be 9.53 CoreMarks/MHz/core.

But does you fixed the frequency, or it is boosting up?
You should fix the frequency at a fixed value, also it would be nice for you, to run it in a core that is less used..

Check for example:

cat /proc/interrupts

And launch a thread on a core with less usage..

[coder]

CoreMark/MHz seems means score just divided by MHz of a single core ... look at their scores list, x86 usually scores much more than else:

https://www.eembc.org/coremark/scores.php

The only way I can make sense of those scores is if it's an all-core score. This would explain why the Tilera chips top the list (hint: sort by CoreMark/MHz).

If we calculate it like this, then this Raspberry Pi 4 would score 13 CM/Mhz there:

https://openbenchmarking.org/result/...AS-RASPBERRY18

Well, the 8-core NXP LS2088A Cortex-A72 @ 1.8 GHz scored 36.10. Scaling that down to 4 cores @ 1.5 GHz would give 15.04, which is roughly in line with your number. Then, divide by 4 and we get 3.76, which matches up pretty well to the Xeon E-2288G's per HT score of 4.83.

[audir8]

There's a reason why companies have a no-GPL policy sometimes, even when the GPL would basically not affect them or help them if done right. A GPL ISA does make sense, you could better control it, and if someone wanted to add SIMD, or non-derivative instructions, they could do it without worrying about the GPL. Not sure what it will take for this impression of the GPL to be fixed, but maybe getting along with other open source licenses for non-derivative works should be made a priority for maintainers of GPL projects.

I certainly expect better RISC-V dev boards to keep coming out, but for anything a mainstream consumer would use, SiFive (or a company like Amazon/Broadcom like previously mentioned) will need need a vastly different set of priorities than just competing with licensed ARM. I hope someone will get there eventually with a RISC-V/SIMD+GPU SoC/PIB as a Celeron/Pentium/A-series competitor though I don't expect it to be an easy market at all because of Intel/AMD.

[coder]

ISA isn't like a compiler. A better analogy would be a proprietary compiler for a language with a free license.

A better analogy would be replicating an API. All you're doing is implementing a functional specification. Before the Sun/Java ruling, this was clearly not something one could protect via copyright, and that's what open source licenses hang on. Otherwise, you'd need to secure the ISA with patents and create an open source license for gaining access to those.

In any case, I guess the only way we'll know is if an open source CPU design actually takes off. I could see that maybe for embedded use, but I still think it wouldn't be in Alibaba's interests to use a fully open source core, for their purposes. They're doing this not just to protect themselves from US IP, but also for competitive reasons.

[coder]

I'm just trying to shit on every non-copyleft license on every occasion I get.

IMO, it's more useful to look at how GPL actually works, and then figure out how that can map onto something like an ISA.

As I said, I think the most robust way would've been for RISC-V to hold some patents needed to implement its ISA (which is presumably how ARM and Intel protect their ISAs). Then, you could restrict licensing of those patents to someone who abides by your terms.

[dungeon]

The only way I can make sense of those scores is if it's an all-core score. This would explain why the Tilera chips top the list (hint: sort by CoreMark/MHz).

Well If you fllter just certified scores, 95% of scores were running just single core bench

So i am kind of not sure if this CM/MHz scores are even valid as multithreaded

Highest score there for certified single thread is 5.82 , blah, blah. non-certified goes up to 9.61

Whatever, it seems it is more validated to run as single

[bridgman]

The only way I can make sense of those scores is if it's an all-core score. This would explain why the Tilera chips top the list (hint: sort by CoreMark/MHz).

Well, the 8-core NXP LS2088A Cortex-A72 @ 1.8 GHz scored 36.10. Scaling that down to 4 cores @ 1.5 GHz would give 15.04, which is roughly in line with your number. Then, divide by 4 and we get 3.76, which matches up pretty well to the Xeon E-2288G's per HT score of 4.83.

I just built and ran CoreMark on my Kaveri box with GPU active and the benchmark workload seemed to be single core only.

The score was 13218; if the CPU was running at max boost (3.4 GHz) that gives 3.89 CM/MHz, but could be higher depending on what the actual clock freq was.

EDIT - looks like default is single core but you can build for N cores if you want (following is from README.md):

# Parallel Execution
Use `XCFLAGS=-DMULTITHREAD=N` where N is number of threads to run in parallel. Several implementations
are available to execute in multiple contexts, or you can implement your own in `core_portme.c`.

~~~
% make XCFLAGS="-DMULTITHREAD=4 -DUSE_PTHREAD"
~~~

Above will compile the benchmark for execution on 4 cores, using POSIX Threads API.

[dungeon]

I just built and ran CoreMark on my Kaveri box with GPU active and the benchmark workload seemed to be single core only.

The score was 13218; assuming the CPU was running at max boost (3.4 GHz) that gives 3.89 CM/MHz.

It could become multithreaded if you build it (in case of 4 core Kaveri) with 'make XCFLAGS="-DMULTITHREAD=4"'

edit: seem i am reading too fast, while posting too slow