SiFive Announces The Performance P550 As The Fastest RISC-V Processor Yet

[brucehoult]

They obviously can't have bridged the gap if Cortex-A72 is more than 3 times faster than the U74 (Pi 4 is still 4.5 times faster than your GB5 result).

Pi 4 1.5 GHz Geekbench 5 multi-core results are around 500, a little less than twice Pi 3 results.

Nope, IPC is extremely important to overall performance.

IPC is just one factor. There are many equally-effective ways to design CPUs, trading off one factor against another.

So you found a U74 board somehow and overclocked it to 1.5GHz?!?

"found somehow"? lol.

Apparently you don't have any understanding of what overclocking is.

Different chips of the same type, even from the same batch, can run at different speeds. Manufacturers such as Intel test each chip to find the maximum speed it will run reliably at (at the expected extremes of voltage, temperature etc) and slap on a label. They sell the ones that tested at a higher speed for more money.

SiFive is not speed grading FU-740 chips. They don't have a rated speed. They are sure they will all run at 1.2 GHz at least, so they are shipping them running at that speed. Users doing their own speed grading is no different to the manufacturer doing it.

I haven't yet seen anyone report that theirs doesn't work at 1.5.

The upcoming BeagleV StarLight also uses the U74 core (a newer faster version of it). It's a commercial mass-production product. They are advertising a rated speed of 1.5 GHz for it.

I'm not sure what you're trying to prove here - despite the overclocking and 4 times the L2 cache in U74, a Pi 3B still beats U74 by 60% on the overall integer score.

GeekBench is a worthless benchmark here. It tests things that are irrelevant to U74 users, such as benchmarks using SIMD or crypto instructions. If you need those, you don't get a U74. Simple.

Worse, Pi 3B is one of the slowest A53 boards available. There is a faster B+ model and several 2GHz Cortex-A53 boards.

Perhaps you didn't notice I included my Odroid C2 in my benchmark.

On this particular benchmark U74 needs 11.3% fewer cycles than Cortex-A53 (AArch64 version) - that's neither a huge difference nor a surprise.

No one is claiming the U74 thrashes the A53. The claim is that it is competitive with it (and the A55).

SPEC and Geekbench are more representative for what people do with modern CPUs.

I completely agree about SPEC.

Geekbench might be relevant for casual computer users or people doing photo or video work. It's irrelevant for software developers.

Maybe you are happy with these awful U74 results, but I bet many people will be disappointed if they believe the marketing claims.

Which marketing claims are those? "The world's fastest RISC-V native development platform". That is a precisely correct claim, in 2021.

Yes, I'm happy. The HiFive Unmatched is a far more pleasant machine to sit at and do software development on than a Pi 3. In many respects it's better than a Pi 4 too. The CPU cores are slower than a Pi 4 but everything else is better.

Most normal people should wait for the BeagleV which will be almost as good at a much cheaper price -- or for machines late next year or early 2023 with the Vector, Bitmanip, Scalar Crypto, and CMO extensions. But for anyone helping get the software ecosystem ready for those coming machines the HiFive Unmatched is easily the best choice.

[coder]

Nope, IPC is extremely important to overall performance.

I agree with brucehoult , except that he didn't mention perf/W. When that's it's a priority, wide-and-slow is usually a good bet for better power-efficiency. However, in terms of perf/area (which correlates to perf/$), a faster-clocking, lower-IPC core is probably better.

As for the rest of your post, I don't see where it's advancing the discussion. At this point, it's probably best to sit and wait for more data.

BTW, if you have a vested interest in the matter, I think you should disclose it.

[PerformanceExpert]

I agree with brucehoult , except that he didn't mention perf/W. When that's it's a priority, wide-and-slow is usually a good bet for better power-efficiency. However, in terms of perf/area (which correlates to perf/$), a faster-clocking, lower-IPC core is probably better.

As for the rest of your post, I don't see where it's advancing the discussion. At this point, it's probably best to sit and wait for more data.

BTW, if you have a vested interest in the matter, I think you should disclose it.

You should ask brucehoult that since he is the one making claims without any evidence. I'm simply debunking those ridiculous claims.

I have a S21 Ultra that scores 1133 on Geekbench, while the fastest available SiFive board does 58 - a factor of 19.5x slower... If SiFive is only 4 years behind Arm like Bruce claims, one should be able to buy a RISC-V device that gets 1100+ in 4 years. How likely do you think that is true?

The most efficient CPUs clock around 3GHz and use very wide, high IPC pipelines to get the best perf/W and perf/area. x86 designs use deeper, lower IPC pipelines which clock at 5GHz but use several times more area and power. So there is no evidence that high clocks are good for perf/area. Either way, the performance gap between RISC-V and Arm is way to large to bridge by frequency - the vast majority will have to come from IPC gains.

Anyway, we shall see whether RISC-V achieves a score of 1100 before 2030 - at least that is a realistic possibility.

[coder]

You should ask brucehoult that since he is the one making claims without any evidence. I'm simply debunking those ridiculous claims.

I think you need to chill. You're on thin ice.

Post the data you've got and let people decide for themselves who's debunking whose ridiculous claims.

[PerformanceExpert]

I think you need to chill. You're on thin ice.

Post the data you've got and let people decide for themselves who's debunking whose ridiculous claims.

I already posted lots of data, but here is another one: a 2.6GHz Cortex-A53 scores 208/820. Despite the old process, this is one of the best A53 implementations, so once the fastest RISC-V silicon beats this one, it would be correct to claim to be faster than Cortex-A53. If that happens next year (which is possible), you can claim to be 7 years behind Arm at that point.

But right now there is no evidence of any RISC-V device that beats Cortex-A53.

[metalpenguin]

Most of the volume is in tiny microcontrollers which can't run Linux. I mean a Linux board so one verify the performance claims in the real world. I'd love to see Phoronix results for an U84 board.

The announced P550 doesn't even have a vector unit, so it'll be even longer before that gets added. Without it isn't even equivalent to Cortex-A72, so I think all the talk about catching up is way too optimistic.

You're right to take the announcements with a grain of salt. However you also seem to be implying that it's simply not believable that SiFive made such a performance leap from the U74 to the P550. They are completely different microarchitectures. Nobody is claiming that SiFive is coming out with cores in the next year or two that can compete with ARM's high-end like the A710 or the X2. I personally believe that's also not a big deal. There is growing industry demand for an alternative to ARM, and there are plenty of companies looking for mid-level cores as well.

Designing cores is hard and takes lots of work, so the P550 was the first generation of an out-of-order core effort. But I'm willing to bet that they will improve it quickly in the future generations.