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

#201

Post by elteto » Fri, Dec 18, 2020, 9:11 PM UTC

adrianmonk wrote: post #183

elteto wrote: post #75

Earlier quoted context omitted.

AMD should see the writing on the wall and start designing ARM CPUs. I think that would be very exciting because they have the engineering expertise and talent necessary to compete against Apple's M1. I don't know of any other company that can do that (besides Intel, but you know...). Additionally, there is a huge hole open right now for ARM on everything else that is not an Apple product. Apple normalized ARM on con…

> AMD should see the writing on the wall and start designing ARM CPUs. This is a really tough strategy decision. Focus all energy on delivering a knock-out blow to Intel (in x86/64) or try to do other stuff too? For the next several years at least, there is still a whole lot of money to be made selling x86/64 chips. (I don't know if Microsoft has plans to move mainstream Windows users off x86/64, but even if they do,…

This is a very thoughtful comment, thank you.

I would love to be a fly on the wall on the corresponding AMD strategy meetings :)

I think to not focus on x86 right now would be a tremendous blunder on AMD’s part. And at the same time ignoring ARM for too long will give others (like Nvidia) time to setup shop and capture the market.

I wonder how much of the Zen architectural “backend” is amenable to be reused with a different ISA (I don’t even know if such thing would be possible). Perhaps they could have a very low-effort initiative to try something along those lines.

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

#202

Post by waynesonfire » Fri, Dec 18, 2020, 9:25 PM UTC

klelatti wrote: post #85

jeffbee wrote: post #81

Earlier quoted context omitted.

Right? "This platform does/does not run my software" is a huge checkbox when buying computers. The viewpoint of people who actually specify, buy, and operate servers at scale is severely underrepresented in these discussions.

So you're never going to consider moving to another architecture because your software (currently) doesn't run on it? So Amazon made a huge mistake with Graviton then? Last time I checked Amazon 'specify, buy, and operate servers at scale'.

I don't know if it's a mistake, the competition is welcome. I spun up some arm servers because they were lower cost but then ended up having to worry about software availability issues. It's non-trivial and I don't think it was worth the savings at this time and went back to x86.

Regardless, the claim that there is no value in Intel let alone x86 is a stretch.

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

#203

Post by FeepingCreature » Fri, Dec 18, 2020, 9:27 PM UTC

txdv wrote: post #17

32cores for 800$, feasible for a nice home server rig

Or (as https://news.ycombinator.com/user?id=thinkmassive pointed out) wait for SOPine ClusterBoard https://pine64.com/product/clusterboard-with-7-sopine-comput... and get 28 cores for $310. ($100 board, $30 per A64 compute module once it comes out, seven modules on a board.)

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

#204

Post by kllrnohj » Fri, Dec 18, 2020, 9:27 PM UTC

elteto wrote: post #75

cameron_b wrote: post #42

This line is the kicker for me: "Intel’s current Xeon offering simply isn’t competitive in any way or form at this moment in time." The rest of the paragraph twists the knife more, and declares there to be an open market on server hardware for AMD and Ampere to fight it out. With application stacks being what they are and optimizations for Xeon so deeply integrated, it's unlikely that folks running private datacenter…

AMD should see the writing on the wall and start designing ARM CPUs. I think that would be very exciting because they have the engineering expertise and talent necessary to compete against Apple's M1. I don't know of any other company that can do that (besides Intel, but you know...). Additionally, there is a huge hole open right now for ARM on everything else that is not an Apple product. Apple normalized ARM on con…

> AMD should see the writing on the wall and start designing ARM CPUs.

ARM's brand new 80-core system struggles to match AMD's year old 64-core system, while also being entirely unable to run some of the workloads either at all or in the 2S configuration. And with mostly comparable power draw and a much worse turbo story that ARM is trying to pretend is somehow a good thing.

Competition is great, but it's a bit premature to claim "the writing is on the wall" is it not? AMD has historically made ARM CPUs in the past, they're not rigidly stuck to x86 like Intel is, but it's also not like x86 is in immediate danger, either.

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

#205

Post by foobarian » Fri, Dec 18, 2020, 9:30 PM UTC

avianlyric wrote: post #200

foobarian wrote: post #195

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. They are already using TSMC so they are not held back by density.

> 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.

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

#206

Post by UncleOxidant » Fri, Dec 18, 2020, 9:31 PM UTC

loudmax wrote: post #118

ipodopt wrote: post #82

Earlier quoted context omitted.

The RISC-V ISA would be the better bet. a. They start investing now and they can have some leverage in the spec. b. The spec is minimal and open to expansion that it sets itself up for more domain specific processors. Which we will be seeing alot more of thanks to the slowing of transistors per dollar growth. c. There is a pretty big theme of moving toward more open and democratized standards in tech. d. They would b…

I'd love to see RISC-V or another open source ISA become the dominant design. But I'm not sure RISC-V is there yet. Companies have been pouring resources into optimizing ARM CPUs for years now, and they're now starting to get to the point where their performance competitive with x86. RISC-V is just getting started. I don't know if there's a RISC-V CPU that's comes anywhere close to the raw performance of higher end A…

> Companies have been pouring resources into optimizing ARM CPUs for years now

Yes, companies (Apple especially, but also Samsung and Qualcom) have been making optimized ARM CPUs, but those designs remain within those companies. Very similar optimizations could be made on a RISC-V implementation and AMD has proven that they have the ability to make design optimizations.

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

#207

Post by qchris » Fri, Dec 18, 2020, 9:31 PM UTC

walrus01 wrote: post #110

bluGill wrote: post #20

Earlier quoted context omitted.

There are several multi-board RPi&Co solutions. Pine64 has some interesting offerings in this space.

The problem with most of these is availability of a Linux operating system image that is as mature, and close to 'stock' debian as raspbian is. Usually it's some two year old version of Ubuntu that's been cobbled together by the vendor.

I believe there's been some work in the this space by the Armbian[1] community, which is aiming to create a unified base distro for a wide range of ARM single-board computers including Odroid, Pine64, etc.

[1] https://www.armbian.com/

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

#208

Post by coder543 » Fri, Dec 18, 2020, 9:34 PM UTC

dragontamer wrote: post #182

pinewurst wrote: post #87

Earlier quoted context omitted.

What they mean (IMHO) is that Intel pricing on Xeon Platinum SKUs is so disproportional to their comparative performance that they would be foolish to purchase.

Xeon Platinum supports 8-socket machines. 6x memory channels x 8-sockets == 48x DIMMs for your massive 6TB RAM servers. Platinum is crazy priced for a crazy reason: its a building block to truly massive computers that only IBM matches. Supercomputers don't even typically use Platinum: Platinum is for the most vertically-oriented systems. --------- Xeon Gold is your more typical 2-socket or 4-socket machines: same per…

> Xeon Platinum supports 8-socket machines. 6x memory channels x 8-sockets == 48x DIMMs for your massive 6TB RAM servers.

Both AMD Epyc Rome and Ampere Altra can support 8TB of RAM in dual socket machines, so that spec isn't particularly impressive anymore.

> Platinum is crazy priced for a crazy reason: its a building block to truly massive computers that only IBM matches.

Meh. A dual-socket Rome system with 128 cores (256 threads), 8TB of RAM, and 128 PCIe 4.0 lanes is an enormous machine. A full 8 socket Platinum system is going to cost an order of magnitude (or two) more, and the performance improvement is going to be less than 2x.

An 8-socket system is going to consume an enormous amount of power, cost an unbelievable amount of money, and the complex NUMA domain combined with Intel's interconnect isn't going to "just work" for almost any software, so you're likely going to put a lot of effort into making that system barely work.

Intel Platinum is priced crazy because Intel is building monolithic processors (which means low yields) and Intel likes to have a substantial profit margin. Together, those mean high prices.

At a certain point, you're better off investing in making your application work on an accelerator of some kind, in the form of a GPU or an FPGA or task-specific ASIC, rather than giving Intel all your money in the hopes of eeking out a marginal performance improvement. Alternatively, finding software that can scale to more than one machine.

Maybe IBM has truly large systems that are worth considering. Intel doesn't right now, unless you're using some software that demands Intel for contractual reasons no one can argue with.

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

#209

Post by kllrnohj » Fri, Dec 18, 2020, 9:37 PM UTC

a2tech wrote: post #13

corty wrote: post #11

Nice hardware, but if I order a hundred, how many will I get? Arm servers so far have been vaporware, only available in single-digit quantities after a very long wait. There is a new announcement every half year or so by one firm or the other. But we've never managed to get beyond getting a sample, and not for a lack of trying. Numbers just aren't available, models change extremely fast, producers go bancrupt. Maybe…

Like Amazon? You know you can get an ARM server there any day of the week for extremely reasonable prices? Edit: to clarify, I mean on AWS, not from the Amazon store.

Amazon's Graviton2 is in some of the Anandtech benchmark results. The availability & pricing is probably related to how far behind it is in those benchmarks, too...

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

#210

Post by dragontamer » Fri, Dec 18, 2020, 9:42 PM UTC

coder543 wrote: post #208

dragontamer wrote: post #182

Earlier quoted context omitted.

Xeon Platinum supports 8-socket machines. 6x memory channels x 8-sockets == 48x DIMMs for your massive 6TB RAM servers. Platinum is crazy priced for a crazy reason: its a building block to truly massive computers that only IBM matches. Supercomputers don't even typically use Platinum: Platinum is for the most vertically-oriented systems. --------- Xeon Gold is your more typical 2-socket or 4-socket machines: same per…

> Xeon Platinum supports 8-socket machines. 6x memory channels x 8-sockets == 48x DIMMs for your massive 6TB RAM servers. Both AMD Epyc Rome and Ampere Altra can support 8TB of RAM in dual socket machines, so that spec isn't particularly impressive anymore. > Platinum is crazy priced for a crazy reason: its a building block to truly massive computers that only IBM matches. Meh. A dual-socket Rome system with 128 core…

But not with 6-channels per socket (48-channels total across an 8-socket computer). That's a heck of a lot of bandwidth.

I was doing the specs from memory and conservatively: 128GB sticks across 48-channels.

https://www.supermicro.com/en/products/system/7U/7088/SYS-70...

This 8-way Supermicro system supports 192x DIMMs. So... 128GB sticks x 192 == 24TBs of RAM or so, maybe 48TBs.

> At a certain point, you're better off investing in making your application work on an accelerator of some kind, in the form of a GPU or an FPGA or task-specific ASIC, rather than giving Intel all your money in the hopes of eeking out a marginal performance improvement. Alternatively, finding software that can scale to more than one machine.

I mean, these 8-way systems are what? $500k ? Less than $1MM for sure. How many engineers do you need before a machine of that price is reasonable?

Really, not that many. If you can solve a problem with a team of 10 engineers + 48 TBs of RAM, that's probably cheaper than trying to solve the same problem with 30 engineers with a cheaper computer.

-----------

> An 8-socket system is going to consume an enormous amount of power, cost an unbelievable amount of money, and the complex NUMA domain combined with Intel's interconnect isn't going to "just work" for almost any software, so you're likely going to put a lot of effort into making that system barely work.

NUMA scales better than RDMA / Ethernet / Infiniband. A lot, lot, LOT better.

All communications are just memory-reads, and that's a real memory read. Not an "emulated memory read transmitted over Ethernet / RDMA pretending to be a memory read". You incur the NUMA penalty but that's almost certainly better than incurring a PCIe penalty.