News for hackers. Stuff that matters.
Re:Intel Tock-Ticks Chipsets Back to 22nm by tinus_hn on Friday September 28, 2018 @10:30AM (#18093646)
It doesn’t matter what AMD is building. The only thing that matters for Intel is how many products they ship because that’s where the money is made.
Re:Intel Tock-Ticks Chipsets Back to 22nm by ben_w on Friday September 28, 2018 @10:35AM (#18093695)
If Intel’s competitors are building better products than Intel, then Intel ships less products or has to charge less for their products. It matters what AMD produces.
Re:Intel Tock-Ticks Chipsets Back to 22nm by tinus_hn on Friday September 28, 2018 @06:47PM (#18097679)
In a free market that’s true. But it isn’t true. Guess what that means!
Re:Intel Tock-Ticks Chipsets Back to 22nm by rbanffy on Friday September 28, 2018 @10:45AM (#18093777)
What's better for one person may not be better for all others.
Re:Intel Tock-Ticks Chipsets Back to 22nm by ben_w on Friday September 28, 2018 @12:23PM (#18094627)
In principle I agree with you.

Unfortunately, your response doesn’t really deal with the point I made: not only does it apply equally to both companies, but also I cannot think of any economically significant use case where Intel will beat AMD while costing more ($ or watts) per gflop.

Re:Intel Tock-Ticks Chipsets Back to 22nm by loeg on Friday September 28, 2018 @01:34PM (#18095250)
Absolute single thread perf and/or memory latency still matter for many applications. AMD definitely wins on a $/gflop basis, especially at high core counts. But some applications are happy to trade absolute gflops for low latency and pay Intel out the nose for it.
Re:Intel Tock-Ticks Chipsets Back to 22nm by rbanffy on Friday September 28, 2018 @05:18PM (#18097143)
AMD wins on $/GFLOP for some applications, but not others. Their SIMD pipelines are half as wide as Intel's (even though they don't need to throttle as much as Intel when using them). They also got a kick out of the speculative execution attacks that hit Intel more (but that matters to cloud providers most, businesses less and doesn't matter at all for HPC users). You also need to consider what your compilers do and using the right optimizations for the specific architecture could make a huge difference.

It's a bit like x86 vs SPARC vs POWER vs ARM. Each has workloads that run better on them for the same amount of upfront and operational costs.

Re:Intel Tock-Ticks Chipsets Back to 22nm by loeg on Friday September 28, 2018 @11:30PM (#18098569)
Yeah, my statement was mostly aimed at high core count applications. It's definitely true that Intel remains more competitive in the low-core market segments. Extra boards and chasses add their own costs, though.

AMD is less than 50% the price at large core count models, relative to comparable to Intel, so even with half the SIMD pipeline they can still win on $/gflops there.

It isn't really the same as considering entirely different ISAs. The switching cost Intel -> AMD is far cheaper than switching Intel -> PPC. Sparc and ARM are not competitive at all.

Re:Intel Tock-Ticks Chipsets Back to 22nm by geezerjay on Friday September 28, 2018 @02:36PM (#18095837)
> But some applications are happy to trade absolute gflops for low latency and pay Intel out the nose for it.

Can you provide one example of where Intel products beat AMD's in latency in a way that prospective customers are willing to pay a premium for that performance delta?

Re:Intel Tock-Ticks Chipsets Back to 22nm by loeg on Friday September 28, 2018 @04:36PM (#18096826)
My employer, for example, probably can't live with the single thread perf regression that the combination of reduced IPC and steeper remote memory latency causes. We would love to pay less for CPUs, but they aren't the dominant cost in our platform (TBs of flash costs $$$$$) so the savings is marginal. (I don't really want to name my employer here, though.)
Re:Intel Tock-Ticks Chipsets Back to 22nm by rbanffy on Friday September 28, 2018 @05:27PM (#18097197)
Another case is memory. If your working dataset is, say, 24 TB, you want a machine that can have that much memory and you most likely have bigger problems than how much each of your 8 CPUs costs.

With that kind of problem, you may want to abandon x86 altogether.

Re:Intel Tock-Ticks Chipsets Back to 22nm by mmt on Friday September 28, 2018 @11:10PM (#18098510)
> With that kind of problem, you may want to abandon x86 altogether

At this end of the spectrum, that means abandoning commodity hardware, with everything it implies, not just a change in binary architecture, as it might for ARM vs. x86 in end user devices.

Recently, I looked up the max memory on IBM's system z, and it was only 32TB, but it looks like their Power Systems E980 can take 64TB. It appears that, like with their mainframe line, going that route means significant vendor lock-in, including proprietary memory modules.

It's certainly interesting that the current generation of high-end x86 CPUs took a step backwards from the last generation, in that that their max is 12TB and just catching back up to 24TB isn't expected to happen until next year.

Re:Intel Tock-Ticks Chipsets Back to 22nm by loeg on Saturday September 29, 2018 @01:52AM (#18098917)
Fortunately, we do not have that problem :-).
Re:Intel Tock-Ticks Chipsets Back to 22nm by rbanffy on Saturday September 29, 2018 @09:19AM (#18100163)
I'd be extremely happy to have that kind of problem.

For some reason, while I was in college, I got a complimentary subscription to a magazine called Supercomputing Review. It was one of the most interesting reads of that time.