The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster • Hacker News Board

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#241

Post by klelatti » Sat, Dec 19, 2020, 1:03 AM UTC

kllrnohj wrote: post #239

klelatti wrote: post #219

Earlier quoted context omitted.

It's not Arm's system it's Ampere's. Arm is claiming nothing about this cpu.

This system uses ARM's Neoverse-N1 CPU cores. Ampere put it in a socket, but it's still "ARM's" CPU core. This isn't a custom design like Apple's CPUs. And per the article it sounds like it's ARM's fault this can't exceed 3.3 ghz. That seems to be a Neoverse-N1 limitation: "Fundamentally, the Altra’s handling of frequency and power in such a manner is simply a by-product of the Neoverse-N1 cores not being able to clo…

System != Core for a whole range of reasons as you know - not least price. I actually think that Ampere should get the credit for producing a system this competitive (even with an off the shelf core).

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#242

Post by avianlyric » Sat, Dec 19, 2020, 1:17 AM UTC

foobarian wrote: post #205

avianlyric wrote: post #200

Earlier quoted context omitted.

> I'm also curious if there is something inherent to ARM that would stop them from making similar optimizations to the AMD chips. Yeah ARM isn’t x86. Apple made big gains by leveraging the difference in instruction length and complexity between x86 and ARM. Something AMD can’t do, and they’ve said as much.

I still don't have a clear understanding. Apple also made big gains by integrating performance sensitive stuff on the package. AMD could do the same. I wonder if they are constrained by some hardware-level interoperability requirements across the motherboard, like with chipsets or DMA controllers or whatnot made by different companies. Apple clearly has a nice advantage here.

My understanding from reading around (I can't find sources right now) is that a big chunk of Apple's performance comes from how wide their instruction decode hardware, and long their instruction pipeline is.

ARM instructions are fixed length so it's very easy to stack up decoders next to each other. You don't need complex logic to figure out where instruction boundaries are, which you do with x86.

This added to a super long instruction pipeline means that the M1 has the ability to reorder a significant (something silly like 2x longer than any AMD or Intel CPU) number of instructions to increase CPU utilisation and thus performance.

AMD can't do this on an x86 CPU because the complexity involved is too high, and think someone has found comments from them basically saying so.

Yeah Apple did make gains from a better integrated package, some of that would be tricky to pull off else where without additional software support. Some of the other integrations would be tricky to pull of because it would reduce the amount of customisation OEM could do. Remember one of the big things Apple integrated was RAM, and making it unified memory for all the co-processors. That could be tricky to do, unless we move into a world where RAM is always integrated into CPUs.

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#243

Post by redshirt2 » Sat, Dec 19, 2020, 2:32 AM UTC

klelatti wrote: post #154

random5634 wrote: post #116

Earlier quoted context omitted.

AMD is doing fine. "The writing on the wall". Zen3 is going to very competitive, their interconnect on chip is great. Sure, add a line for ARM. What's interesting to me, I feel like a lot of this started when intel wouldn't do the chip for the iphone. All of a sudden you had this leading edge mfg pushing arm hard which drove volumes through the roof, and all the copycats (android etc) copied the chip selection, scree…

I don't think that Intel producing the iPhone chip was ever likely to happen - sure Jobs had to ask Intel as a courtesy for a partner but it was never realistic. - Arm has dominated low power / mobile segment from well before the iPhone due to its superior power consumption; - The non Arm IP (eg Imagination GPU) was also important and it was unlikely that Intel would have been integrated that on their SoCs. I had an…

Apple was also one of the first investors in Arm and one of the reasons the original company Acorn existed.

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#244

Post by lukeh » Sat, Dec 19, 2020, 4:22 AM UTC

avianlyric wrote: post #199

spockz wrote: post #189

Earlier quoted context omitted.

The reason why the M1 performs so well, next to its insanely wide out of order capabilities, is that Apple placed Specialized processors really close and gave everything a fast connection. Then because they control a large part of The OS and applications they rewrote the underlying macOS libraries and the applications to make use of these specific co processors. They were able to achieve this because they are vertica…

I have to disagree with this, I really don’t think that Apple is a competitor here. The M1 chip is undoubtably a remarkable piece of engineering. But it quite clear that Apple are never going to sell it standalone. It also seems extremely unlikely that Apple are going to get back into the server game again. Apple sell high markup, low volume (relatively) products. Data centre severs are the opposite, and Apple have s…

What about Linux servers for their own datacenters?

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#245

Post by blasdel » Sat, Dec 19, 2020, 6:30 AM UTC

coder543 wrote: post #224

dragontamer wrote: post #221

Earlier quoted context omitted.

> If the application is relying heavily on the NUMA interconnect to transfer tons of data, it's not going to be making efficient use of the processor cores or the memory bandwidth. It's a total all around bust. You're just wasting money at that point. If the interconnect is your bottleneck, you spend money on the interconnect to make it faster. Basic engineering: you attack the bottleneck. You don't start talking abo…

Preface: If you have actually encountered applications that must be run on 8-socket systems because those are literally the only fit for the application... I would love to hear about those experiences. With the advent of Epyc Rome, most use cases for these 8-socket systems vanished instantly. It would be fascinating to hear about use cases that still exist. Your experiences are obviously different than mine. If you n…

Don't forget that Rome also has a wildly nonuniform interconnect between the core complexes, and the system integrator gets much less control over it than Intel's UPI links. When you really need to end up with a very large single system image at the application layer, the bigger architecture works out to be much cheaper than 256gb DIMMs or HPC networking.

8-socket CLX nets you 1.75x the cores, and 3x as many memory channels vs. a 2-socket Rome system. It also scales to a single system image with 32 sockets if you use a fabric to connect smaller nodes:

* 4-socket nodes: https://www.hpe.com/us/en/servers/superdome.html

* 2-socket nodes: https://atos.net/en/solutions/enterprise-servers/bullsequana...

That's 48tb of DRAM with all 128gb DIMMs, or 12tb+128tb when using 512gb Optane PDIMMs.

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#246

Post by sitkack » Sat, Dec 19, 2020, 6:41 AM UTC

tlamponi wrote: post #48

So how are the cores interconnected? As ARM server designs love to throw out tens to hundreds of cores which sounds good on paper, but are not really useful to lots of workloads as they have a shit interconnect or lots of NUMA nodes. Interconnect is key, e.g., see the improvements between zen2 and zen3, especially for the `4 An anecdote: I got some hands on a box with a Cavium ThunderX2 96 logical core processor abou…

> So how are the cores interconnected?

No one answered that and from the core to core ping-pong test the second processor their is some weird, possibly performance crushing issues.

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#247

Post by spockz » Sat, Dec 19, 2020, 8:03 AM UTC

avianlyric wrote: post #199

spockz wrote: post #189

Earlier quoted context omitted.

The reason why the M1 performs so well, next to its insanely wide out of order capabilities, is that Apple placed Specialized processors really close and gave everything a fast connection. Then because they control a large part of The OS and applications they rewrote the underlying macOS libraries and the applications to make use of these specific co processors. They were able to achieve this because they are vertica…

I have to disagree with this, I really don’t think that Apple is a competitor here. The M1 chip is undoubtably a remarkable piece of engineering. But it quite clear that Apple are never going to sell it standalone. It also seems extremely unlikely that Apple are going to get back into the server game again. Apple sell high markup, low volume (relatively) products. Data centre severs are the opposite, and Apple have s…

I agree that it is unlikely that Apple will sell the M1 chip separately exactly for the reasons I mentioned in my post.

However, they do not have to sell the chip individually to be a competitor to AMD. If Apple’s are sufficiently attractive it will eat the market share of consumer devices from the likes of Dell/HP/etc. When they start selling less devices they have less need for AMD or Intel CPUs.

This leaves only the datacenter market where Apple will not compete with Intel and AMD.

Indeed, there is lots of room for a (new?) player to pull of the same trick as Apple for general purpose cpus. I just think they are not as well positioned because of the lack of vertical integration. So AMD and Intel indirectly have to worry about what Apple is doing.

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#248

Post by timthorn » Sat, Dec 19, 2020, 8:24 AM UTC

redshirt2 wrote: post #243

klelatti wrote: post #154

Earlier quoted context omitted.

I don't think that Intel producing the iPhone chip was ever likely to happen - sure Jobs had to ask Intel as a courtesy for a partner but it was never realistic. - Arm has dominated low power / mobile segment from well before the iPhone due to its superior power consumption; - The non Arm IP (eg Imagination GPU) was also important and it was unlikely that Intel would have been integrated that on their SoCs. I had an…

Apple was also one of the first investors in Arm and one of the reasons the original company Acorn existed.

Acorn existed fully independently of Apple. Check out the history of the BBC Micro and its forebears.

Apple came on the scene later, when Acorn had developed the ARM processor. You're right that they were an original investor in the spinout company.

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#249

Post by lazylizard » Sat, Dec 19, 2020, 9:33 AM UTC

er.

so how much for 1u chassis, 512gb ram, 2 x 1tb nvme, 4 x 10gbe nic, dual psu? what can i run on this? proxmox? kubernetes?

or 4u chassis, 128gb ram, 2 x 128gb nvme, 4 x 10gbe nic, 100 x 16tb sata ? what can i run on this? centos? Ubuntu? zfs? gluster?

Re: The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

#250

Post by st1ck » Sat, Dec 19, 2020, 12:10 PM UTC

aljarry wrote: post #36

iagovar wrote: post #14

Earlier quoted context omitted.

If I can't get a machine for local development I wouldn't care personally. I like the idea of many many threads as I work a lot with data, but i need something like a PC to try things on. Also, don't really like vendor lock-in.

I'm doing some development on AWS currently on g4dn instance with T4 accelerator. With pings lower than 50ms and remote VS Code it gets really efficient. The biggest issue of course is additional mental burden on being remote + you can't simply print image in ssh console... I thought it would be horrible, but after a brief adjustment period it turned out to be okay-ish for my current work.

But you can view images in terminal emulator. Not every one, but kitty, alacritty, st, iTerm2 for MacOS, etc. should support it. This works for me (need to install ueberzug, e.g. via pip):

    { declare -p c=([action]=add [identifier]=0 [x]=0 [y]=0 [width]=80 [height]=80 [path]="mypic.png") ; sleep 5;} | ueberzug layer --parser bash