News China's SpacemiT develops 64-core RISC-V datacenter CPU — 12nm chip allegedly performs like a 10-year old Xen or Opteron but with higher core count

[Admin]

SpacemiT's VitalStone V100 has 64-cores, but its single-thread performance is comparable to that of 2011 processors.

China's SpacemiT develops 64-core RISC-V datacenter CPU — 12nm chip allegedly performs like a 10-year old Xen or Opteron but with higher core count : Read more

 

[pug_s]

Why does Toms hardware complaining about the 'poor' single core performance in a 64 core cpu?

 

[das_stig]

TH are drinking the anti-China racist cool aid and should look at China's cpu development with neutral eyes, yes they can't compete with Intel/AMD/Arm at the moment, maybe they are a decade behind and to have something that can be used in real world tasks in such a short time frame is commendable. If western countries had put as much effort in, we probably would have had quantum computers coming in to Y2K.

See if I get censored by TH thought police !

 

[nookoool]

I thought all generalized Risc-V processor seem to be pretty weak anyways compare to anything modern.... also shouldn't it be compare to ARM servers instead? Thought at one point they also went for weak core with high core count for the server space.

 

[Dementoss]

Unfortunately, we have no idea about its pricing, nor pricing of servers based on the VitalStone V100 processors.

I don't imagine many people, reading this article, will be particularly bothered about the prices.

 

[bit_user]

Why does Toms hardware complaining about the 'poor' single core performance in a 64 core cpu?

But they're comparing it to the single-core performance of other multi-core server CPUs, so it's a relevant comparison.

If we know how fast their single-thread performance compares, then we can scale the performance of the compared CPU by the core-count ratio to get a decent sense of the multi-core performance and how that would compare to modern server CPUs.

 

[bit_user]

TH are drinking the anti-China racist cool aid and should look at China's cpu development with neutral eyes,

I don't see anything racist about the article. Seems pretty fact-based, to me.

yes they can't compete with Intel/AMD/Arm at the moment, maybe they are a decade behind and to have something that can be used in real world tasks in such a short time frame is commendable.

Whether or not it's commendable is a judgement that would require someone with deep experience in CPU design and a foot in the server world. The authors at this site don't have that kind of experience, so it's fine if they leave such notions for readers or others to decide.

The fact that you want them to be praising this CPU designer shows you're not unbiased.

I really don't see a problem with just reporting the facts as they are. That's called journalistic integrity. In fact, a good publication is careful not to mingle editorializing with reporting, which is actually something this site could be a little bit better at, although that's not necessarily what many of its readers want.

If western countries had put as much effort in, we probably would have had quantum computers coming in to Y2K.

Quantum computers are completely different. I think there's no equivalence.

 

[bit_user]

I thought all generalized Risc-V processor seem to be pretty weak anyways compare to anything modern....

The RISC-V IP cores that exist so far generally tend to be rather weak. Nothing about the ISA prevents someone from making one that's comparable in performance & efficiency to the latest x86 or ARM cores, though.

RISC-V's reputation for being weak stems from the fact that it first got a foothold in the low-power embedded market. It's a more recent development that people are using them to build server CPUs. There's a company called Ventanna, which is already on its second generation of RISC-V server CPUs (the first one was never publicly sold, making it more of a prototype). Here's some info about them, but it's pretty out-of-date. You can find more recent news, on the internet, pretty easily:

https://www.tomshardware.com/pc-com...nd-ventana-to-build-a-risc-v-cpu-gpu-platform
​

also shouldn't it be compare to ARM servers instead?

Would've been nice, but they're probably limited in what's been published about it. SPEC2006 is an old benchmark that's rarely used any more, because it's heavily influenced by the amount of cache in modern CPUs and not a good test of the interconnect and memory subsystems. So, finding SPEC2006 scores for modern ARM server CPUs isn't necessarily easy.

What I think is more likely: the author was aware that most readers of this site are more familiar with x86 server CPUs and opted just to focus on how it compares to those.

Thought at one point they also went for weak core with high core count for the server space.

AMD has a line of server CPUs which feature "compact" Zen cores. These have the same design and per-clock performance of full-sized Zen cores of the same generation, but don't clock as high and have half the per-core L3 cache.

Intel went further and started making server CPUs comprised entirely of their E-cores. Last year, they launched the first of these, which was their 144-core Sierra Forest. It has no hyper-threading and supports dual-CPU configurations. It caught up to AMD's Zen 4C EPYC, but was leapfrogged again by Zen 5 and Zen 5C EPYCs. The next one is called Clearwater Forest and should be a lot stronger.

 

[thestryker]

RISC-V has a long way to go, but I wouldn't be surprised if the developments out of China push the performance aspect forward. Now that RVA23 was ratified a few months ago there's at least a clear path for implementations. Outside of China I think it will take one of the big companies going all in for performant enough general purpose cores to happen. Qualcomm seems like the most likely candidate, but it's so capital intensive I'm not sure how likely it is for any publicly traded company to get it done in a quick time frame.

 

[pug_s]

But they're comparing it to the single-core performance of other multi-core server CPUs, so it's a relevant comparison.

If we know how fast their single-thread performance compares, then we can scale the performance of the compared CPU by the core-count ratio to get a decent sense of the multi-core performance and how that would compare to modern server CPUs.

You're comparing apples to oranges here. These are low powered efficient cores compared to Intel/AMD server cores which are much more power hungry. Plus the fact that you have 64 of them which is probably not possible with the power envelope with the Xeon's Opteron's at that time.

 

[bit_user]

RISC-V has a long way to go, but I wouldn't be surprised if the developments out of China push the performance aspect forward.

I expect China's main contribution will be towards helping flesh out Linux support for RISC-V, especially in server contexts that are still pretty limited. That's probably the main value of this CPU - as a development vehicle, not so much a deployment platform.

Outside of China I think it will take one of the big companies going all in for performant enough general purpose cores to happen.

I mentioned Ventana. Also, depending on your definition of "general purpose", SiFive already has some pretty good cores. Then, there's Tenstorrent's cores, but I'm not sure those are RVA23-compliant (I think I read where somebody claimed they weren't, although that was a while ago).

Edit: see details below.

 

[bit_user]

You're comparing apples to oranges here. These are low powered efficient cores compared to Intel/AMD server cores which are much more power hungry.

They're using a 12nm process node, so that puts them at an immediate disadvantage on both fronts. Without power figures, I really don't see how you can be so sure they're "low powered efficient cores".

Plus the fact that you have 64 of them which is probably not possible with the power envelope with the Xeon's Opteron's at that time.

They were made on 32nm, so no.

 

[thestryker]

I mentioned Ventana.

They're still not shipping and Imagination has pulled out of RISC-V development. While I don't think either of these things are a death sentence in any way it has put me in a "once they're shipping we'll see" mindset.

SiFive already has some pretty good cores.

The 870 is the RVA23 one and I haven't seen anything but paper specs yet. These are also supposed to be shipping this year and if both companies deliver it ought to be relatively exciting.

 

[bit_user]

depending on your definition of "general purpose", SiFive already has some pretty good cores.

SiFive claims their P870-D cores are scalable and server-grade:

• Six-wide, out-of-order core
• RVA23
• CHI bridge
• RAS Functionality
• Scalable to 256 cores

More:

https://www.sifive.com/cores/performance-p870d
​

Then, there's Tenstorrent's cores, but I'm not sure those are RVA23-compliant (I think I read where somebody claimed they weren't, although that was a while ago).

Okay, it seems that at least by now, this is not correct. Here's what Tenstorrent says about their TT-Ascalon cores:

• 8-wide decode
• RVA23 Compliant
• > 18 SPECint2006 per GHz
• area & power-efficient
• clusters incorporate up to 8 cores
• CHI-E coherent and AXI5-LITE non-coherent interfaces
• Support for Trusted Execution Environment
• RVV 1.0 compliant vector unit
• Dual 256b-wide vector units

They characterize them as suitable for:

• High Performance Compute
• Cloud Servers
• Data Centers
• Wireless Communications
• Automotive ADAS/AV
• Client Computing devices

I got most of that from their PDF, which you can find here:

https://tenstorrent.com/ip/tt-ascalon
​

Their SPECint 2006 score of > 18 per GHz is more than twice that of SpacemiT's X100's 9 per GHz claim.