Um, that article indicates they used "threads" not cores for the xeon side: "Xeon processors from Intel. In the Intel corner, we have the Grantley platform (Broadwell) using two 10-core Xeon processors with Hyper-Threading enabled (40 cores) and Purley (Skylake) using two 12-core Xeon processors with Hyper-Threading (48 cores)."
No, but threads do not provide the same throughput as an additional real core. 16 threads != 16 cores in terms of compute power. That the ARM still beat it in terms of power consumption though suggests this could get interesting, but its a garbage benchmark for lack of controls or evident understanding of the hardware in question.
I think calling it a $2,000 chip is a mistake, because clearly the $2,000 version is a poor value compared to the others. For only 2 fewer cores, you can get it for $600 less. For 8 fewer cores, you can get it for $1100 less.
The 2452 seems like the best value by far. You get 40 cores compared to 16c/32t for AMD EPYC 7301 (which costs about $800, too), and I'm guessing the performance should be at least as good overall, if not like +50% better.
As for Intel, no contest. Intel's 16 core Xeons start at around $2,000, so perf/$ should be at least 2-3x in Qualcomm's favor.
Err. I meant the 2432 version is the best value. I wish chip companies would stop using such confusing codenames (likely on purpose). Anyways, I'd say this is a pretty good start from Qualcomm. If they stick to using the cutting edge process from Samsung or TSMC, rather than wait like 2 years, as their Arm server competitors were doing before, I think they can become a decent competitor in the server chip market. Also, the next-generation should be built using TSMC's 7nm process, so Qualcomm may have an even more competitive chip then. Plus, they would have gotten a chance to learn what the market is actually looking for and better optimize the chip for their customers' wishes.
Also, for anyone wondering why I compared the cores this way, I was kind of basing it on Cloudflare's review of it. Google "Cloudflare Arm takes wing" to check out their review, as Anandtech doesn't allow links here.
When you consider your average hyper-v / vmware cloud deployment, it might very well be that the 48core version is best choice because it aligns better with the license-pricing.
Hardly. So in the cloudflare review it only wins in gzip compression and openSSL across the board. It wins a few cases in brotli. Other than that, it gets beaten pretty handily. My guess is that the 60MB cache is what's doing the trick here. The bdw and sky systems are hampered by being 2P, and lesser total cache which likely helps (the reviewer states this about brotli). And they didn't test it against the elephant in the room, EPYC. For $2k, you'll get an EPYC 1P with 24 cores, 2 more memory channels, and 4X PCIe lanes. And unlike Intel, AMD is motivated to compete on price.
The thing going for Falkor is the low power consumption.
I'd say the 46 core chip may be an even better value proposal than the 40 core version. The reason is that you need to buy a lot more than the CPUs to build a server, especially the memory is going to be a significant cost factor in these price regions. So the 400$ higher CPU price for 15% more cores may be a lot less than 15% of the total price.
The smaller chip would still be good for evaluation or situations where the throughput is limited by something else than the CPU.
I agree. But it really depends how many servers you are building overall. If you are building dozens, use the 40 core chip and just built one or two more to make up for it. At a savings of over $1000 for each server this would likely save a ton of money.
One thing that to think about the number of cores is not a good factor on performance. Difference of core and CPU architecture should be taken in account and not just the core count. For example a Single Xeon core has much more performance than AMD and ARM core.
A lot depends on applications running on system - for example if Application use AVX-512 on Xeons - than using only AVX-256 especially in AMD implementation which is basically two 128 units combine. Also ARM is RISC and by designed it takes many ARM instruction to implement a single AVX instruction.
The IPC difference between high-end Arm cores and Xeon is fairly small nowadays, and there are Arm cores with higher IPC.
Also it's incorrect to claim RISC requires more instructions. On typical code AArch64 would use fewer instructions. On highly vectorized code having wider vectors will help of course, but that's not the type of application being targeted here.
I have almost 7 years in OS development and understand Microprocessors a lot - basically todays high processors instructions take use many RISC level instructions to do one of higher level functions.
I would say the closest Intel Server chip in comparison to this is the 16 Core Atom C3958
You're kidding right? Try building a few large applications for both x64 and AArch64 and compare codesize and static instruction counts. You will be educated.
And that slow Atom? That has difficulty keeping up with mobile phone cores - it's slower than the 2-way Cortex-A73 and runs at a 30% lower frequency. And it has just 16 cores while it would need well over 100 to get similar performance...
Considering the 40 Core model is less than half the price of the 48 core model, wtf is right. Perhaps dual CPU motherboards will surface. And if you were building multiple servers just build one or two additional servers to make up for the 8 core deficiency and save a ton of money overall.
These are Enterprise chips,where $1k isn't considered a lot of money. Take a look at Xenon pricing for reference. Or, how much do you think 768GBs of ram costs?
Depends if you are paying a per core or per cpu license. If your cost is $42k/year per cpu, paying an extra grand for an 20% more performance doesn't matter. I'm guessing there is aggressive silicon selection to keep the same power draw.
Qualcomm clearly put some effort into the SPEC comparison numbers and did nothing outright wrong, as far as I can see. However, the method involves at least one pretty rough step: simply halving the numbers for 2-socket systems. But we know CPUs never scale perfectly from 1 to 2 sockets. There's also the price comparison, where they choose the more expensive dual socket CPUs. The 24 core single socket Epyc might be a far better comparison.
That's not the actual big issue. How many people can code, adapt code and/or debug code written for it. How does JS or implementations of it scale with the cores, how does it handle databases, how about storage IO and a lot stuff that might be more important than core count.
If these companies trying to push ARM into the server market want to succeed they need the software to run well, and that means getting machines to developers. Given the prevalence of open source software in the server world, that means hardware that open source developers can afford to buy as a second machine.
Now the 40 core version isn't a bad price, it is still way too expensive for a secondary machine (and that is assuming two things - that the motherboard is reasonable, and that there is a desktop machine size motherboard available).
Until someone comes out with an ARM64 chip / motherboard combo that offers decent USB/network/SATA or m.2/standard memory sockets with UEFI with reasonable performance at a price that people will pick up as a secondary machine to develop and test on it will be a struggle for these companies to get all the software the enterprise customers rely on to perform well.
Further to this, look at the list of software in the chart - MongoDB, Hadoop, Spark, etc - and now consider what happens when you run into trouble and try and file a bug report with any of those groups. How much attention is your bug going to receive when the developers don't have hardware to reproduce it on?
I'm hoping these types of many-core ARM systems might make for good Erlang boxes, or enable even more cost-effective 'cloud function' services in the future.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
37 Comments
Back to Article
LemmingOverlord - Friday, November 10, 2017 - link
@Ian, is Johan testing it?Ian Cutress - Friday, November 10, 2017 - link
He hasn't got one yet. We should be near the top of the list when they send them out to press though.IGTrading - Friday, November 10, 2017 - link
HotHardware already got some tests on their own platform, CloudFlare.The results are impressive considering that Intel's CPU use exactly 200% of the power used by Qualcomm's new chip to achieve less performance.
For almost all benches, Qualcomm wins while using half the power and having a lower processor price.
Intel doesn't look good at all in CloudFlare.
IGTrading - Friday, November 10, 2017 - link
Link : https://hothardware.com/news/qualcomm-centriq-2400...cekim - Friday, November 10, 2017 - link
Um, that article indicates they used "threads" not cores for the xeon side:"Xeon processors from Intel. In the Intel corner, we have the Grantley platform (Broadwell) using two 10-core Xeon processors with Hyper-Threading enabled (40 cores) and Purley (Skylake) using two 12-core Xeon processors with Hyper-Threading (48 cores)."
Garbage test - eagerly awaiting some real data.
Notmyusualid - Friday, November 10, 2017 - link
Yep, ddriver 2.0.Wilco1 - Friday, November 10, 2017 - link
Wait - are you saying that switching hyperthreading off will improve the Xeon throughput?cekim - Tuesday, November 14, 2017 - link
No, but threads do not provide the same throughput as an additional real core. 16 threads != 16 cores in terms of compute power. That the ARM still beat it in terms of power consumption though suggests this could get interesting, but its a garbage benchmark for lack of controls or evident understanding of the hardware in question.Krysto - Friday, November 10, 2017 - link
I think calling it a $2,000 chip is a mistake, because clearly the $2,000 version is a poor value compared to the others. For only 2 fewer cores, you can get it for $600 less. For 8 fewer cores, you can get it for $1100 less.The 2452 seems like the best value by far. You get 40 cores compared to 16c/32t for AMD EPYC 7301 (which costs about $800, too), and I'm guessing the performance should be at least as good overall, if not like +50% better.
As for Intel, no contest. Intel's 16 core Xeons start at around $2,000, so perf/$ should be at least 2-3x in Qualcomm's favor.
Krysto - Friday, November 10, 2017 - link
Err. I meant the 2432 version is the best value. I wish chip companies would stop using such confusing codenames (likely on purpose). Anyways, I'd say this is a pretty good start from Qualcomm. If they stick to using the cutting edge process from Samsung or TSMC, rather than wait like 2 years, as their Arm server competitors were doing before, I think they can become a decent competitor in the server chip market. Also, the next-generation should be built using TSMC's 7nm process, so Qualcomm may have an even more competitive chip then. Plus, they would have gotten a chance to learn what the market is actually looking for and better optimize the chip for their customers' wishes.Also, for anyone wondering why I compared the cores this way, I was kind of basing it on Cloudflare's review of it. Google "Cloudflare Arm takes wing" to check out their review, as Anandtech doesn't allow links here.
mmrezaie - Friday, November 10, 2017 - link
Thanks for the mentioning of the article. It was a great read.zepi - Friday, November 10, 2017 - link
When you consider your average hyper-v / vmware cloud deployment, it might very well be that the 48core version is best choice because it aligns better with the license-pricing.deltaFx2 - Saturday, November 11, 2017 - link
Neither MS nor VMWare have an ARM virtualization solution so this discussion might be moot at least for now. Perhaps sometime in the future?xakor - Friday, November 10, 2017 - link
How is either a 16c Xeon or AMD not gonna totally mop the floor with this? Note that I'm not trying to be arrogant likely just very ignorant.psychobriggsy - Friday, November 10, 2017 - link
Read the cloudflare review of the chip and then maybe consider positing the opposite of your current question.deltaFx2 - Saturday, November 11, 2017 - link
Hardly. So in the cloudflare review it only wins in gzip compression and openSSL across the board. It wins a few cases in brotli. Other than that, it gets beaten pretty handily. My guess is that the 60MB cache is what's doing the trick here. The bdw and sky systems are hampered by being 2P, and lesser total cache which likely helps (the reviewer states this about brotli). And they didn't test it against the elephant in the room, EPYC. For $2k, you'll get an EPYC 1P with 24 cores, 2 more memory channels, and 4X PCIe lanes. And unlike Intel, AMD is motivated to compete on price.The thing going for Falkor is the low power consumption.
MrSpadge - Friday, November 10, 2017 - link
I'd say the 46 core chip may be an even better value proposal than the 40 core version. The reason is that you need to buy a lot more than the CPUs to build a server, especially the memory is going to be a significant cost factor in these price regions. So the 400$ higher CPU price for 15% more cores may be a lot less than 15% of the total price.The smaller chip would still be good for evaluation or situations where the throughput is limited by something else than the CPU.
Samus - Saturday, November 11, 2017 - link
I agree. But it really depends how many servers you are building overall. If you are building dozens, use the 40 core chip and just built one or two more to make up for it. At a savings of over $1000 for each server this would likely save a ton of money.HStewart - Friday, November 10, 2017 - link
One thing that to think about the number of cores is not a good factor on performance. Difference of core and CPU architecture should be taken in account and not just the core count. For example a Single Xeon core has much more performance than AMD and ARM core.A lot depends on applications running on system - for example if Application use AVX-512 on Xeons - than using only AVX-256 especially in AMD implementation which is basically two 128 units combine. Also ARM is RISC and by designed it takes many ARM instruction to implement a single AVX instruction.
Wilco1 - Friday, November 10, 2017 - link
The IPC difference between high-end Arm cores and Xeon is fairly small nowadays, and there are Arm cores with higher IPC.Also it's incorrect to claim RISC requires more instructions. On typical code AArch64 would use fewer instructions. On highly vectorized code having wider vectors will help of course, but that's not the type of application being targeted here.
HStewart - Friday, November 10, 2017 - link
I have almost 7 years in OS development and understand Microprocessors a lot - basically todays high processors instructions take use many RISC level instructions to do one of higher level functions.I would say the closest Intel Server chip in comparison to this is the 16 Core Atom C3958
https://ark.intel.com/products/97927/Intel-Atom-Pr...
it is a relatively new processor for servers and runs at 31W is designed compete with ARM servers.
HStewart - Friday, November 10, 2017 - link
One more thing, the following link shows the 22 core Xeon beats the Qualcomm 40 corehttps://www.heise.de/newsticker/meldung/Server-Pro...
Wilco1 - Friday, November 10, 2017 - link
Actually the Centric was still a little faster at half the power and a quarter of the cost... So which is the better one in your opinion???Wilco1 - Friday, November 10, 2017 - link
You're kidding right? Try building a few large applications for both x64 and AArch64 and compare codesize and static instruction counts. You will be educated.And that slow Atom? That has difficulty keeping up with mobile phone cores - it's slower than the 2-way Cortex-A73 and runs at a 30% lower frequency. And it has just 16 cores while it would need well over 100 to get similar performance...
Samus - Saturday, November 11, 2017 - link
Considering the 40 Core model is less than half the price of the 48 core model, wtf is right. Perhaps dual CPU motherboards will surface. And if you were building multiple servers just build one or two additional servers to make up for the 8 core deficiency and save a ton of money overall.The pricing is truely insane.
Rocket321 - Sunday, November 12, 2017 - link
These are Enterprise chips,where $1k isn't considered a lot of money. Take a look at Xenon pricing for reference. Or, how much do you think 768GBs of ram costs?andychow - Tuesday, November 14, 2017 - link
Depends if you are paying a per core or per cpu license. If your cost is $42k/year per cpu, paying an extra grand for an 20% more performance doesn't matter. I'm guessing there is aggressive silicon selection to keep the same power draw.MrSpadge - Friday, November 10, 2017 - link
Qualcomm clearly put some effort into the SPEC comparison numbers and did nothing outright wrong, as far as I can see. However, the method involves at least one pretty rough step: simply halving the numbers for 2-socket systems. But we know CPUs never scale perfectly from 1 to 2 sockets. There's also the price comparison, where they choose the more expensive dual socket CPUs. The 24 core single socket Epyc might be a far better comparison.Source for some of that:
https://www.heise.de/newsticker/meldung/Server-Pro...
edzieba - Friday, November 10, 2017 - link
The biggest question is the same one as all the past put-a-bunch-of-ARM-cores-on-a-die chips: does anyone actually WANT them?LauRoman - Friday, November 10, 2017 - link
That's not the actual big issue. How many people can code, adapt code and/or debug code written for it. How does JS or implementations of it scale with the cores, how does it handle databases, how about storage IO and a lot stuff that might be more important than core count.mdvle - Friday, November 10, 2017 - link
Exactly.If these companies trying to push ARM into the server market want to succeed they need the software to run well, and that means getting machines to developers. Given the prevalence of open source software in the server world, that means hardware that open source developers can afford to buy as a second machine.
Now the 40 core version isn't a bad price, it is still way too expensive for a secondary machine (and that is assuming two things - that the motherboard is reasonable, and that there is a desktop machine size motherboard available).
Until someone comes out with an ARM64 chip / motherboard combo that offers decent USB/network/SATA or m.2/standard memory sockets with UEFI with reasonable performance at a price that people will pick up as a secondary machine to develop and test on it will be a struggle for these companies to get all the software the enterprise customers rely on to perform well.
mdvle - Friday, November 10, 2017 - link
Further to this, look at the list of software in the chart - MongoDB, Hadoop, Spark, etc - and now consider what happens when you run into trouble and try and file a bug report with any of those groups. How much attention is your bug going to receive when the developers don't have hardware to reproduce it on?cekim - Friday, November 10, 2017 - link
bug response:"Couldn't open a 10G DB on my Galaxy S8 testbed - can you try this patch and report back?"
Glad to see competition the CPU market. This market has a rough history and a lot of hype (and benchmark fudging), so we'll see....
Wilco1 - Friday, November 10, 2017 - link
Who says developers can't buy low cost hardware? Eg.https://shop.softiron.com/product/overdrive-1000/
mdvle - Friday, November 10, 2017 - link
Better, but $600 for a box with no video capabilities isn't going to appeal to a lot of people looking for a secondary machine.peevee - Friday, November 10, 2017 - link
Does it support multi-CPU configurations, and if it does, how many CPUs? And what platform is required?keta - Friday, November 10, 2017 - link
I'm hoping these types of many-core ARM systems might make for good Erlang boxes, or enable even more cost-effective 'cloud function' services in the future.