Ask HN: What are your predictions on the server CPU market for 2019? | Hacker News Reader

Ask HN: What are your predictions on the server CPU market for 2019?

32 pointsjakobov19 comments

Will EPYC be able to compete with Xeon in TCO? Will Intel need to crash their prices to fight off AMD?

What will happen when AMD hits 10nm in its servers, is it game over for Intel?

Management of Intel is accustomed to have 60% profits and need a mind change and a price drop. I foresee that AMD will get a lot of market share: 30-50% and I use this wide range since besides price/performance there is also the factor of security. Looking at Spectre et al. AMD has a safer product. Last but not least, a single EPYC CPU is also an I/O monster: 128 PCI lanes for AMD vs 40 PCI lanes for Intel.

And the former CEO of Intel said their goal was to keep AMD to 15-20% of the server market. I agree. I think they will be able to capture more than 20%... 30-50% seems reasonable.

> keep AMD to 15-20% of the server market

For the position that Intel is in (x86 cpu dominance) this is awkward thing to say from a FTC point of view.

Its not like they require OEMs to bundle their Shopping Assistant with every CPU installation.

Yes, epyc will be competitive, but xeon will still dominate. Mild price drops, or Intel "discovers" (coffee lake/8086k) extra cores/ghz under a mattress. A lot of budget gets spent under the "nobody gets fired for buying xxxx" principal. 10nm changes none of this, sadly. Amd has had better hardware before.

They've also had close to 50% of the market before.

And the "No one gets fired for buying xxxx" works more with software than with hardware. x86 CPUs are essentially a commodity. The best price / performance + performance density + energy efficiency will probably win. Previously Intel had 2/3 (perf density and energy efficiency). With the new EPYC they will probably have 0/3.

Performance isn't a one dimensional metric that you treat it as. Different CPUs have different comparative benchmarks depending on the tests.

I treated it that way for simplicity. In the increasingly virtualized environments at data centers the performance of many more cores will be even more favorable to single core performance. Also, more pci bandwidth with 128 lanes. Not sure if the memory bandwidth is better, but you get the picture.

Some shops will literally buy from the same vendor even if perf/$ is 3x higher going another route. The IT director at place years ago would only go with the most expensive HP stuff, we literally could have had 3x the compute and 4x the storage for going super micro. Even if if his ridiculous MTBF numbers for SM were true, we still would have had 1.5x compute and 3x storage after 3 years for the same price.

So yes, if EPYC isn't offered on the right vendor platforms, they will be shutout even if perf and density are better.

> Some shops will literally buy from the same vendor even if perf/$ is 3x higher going another route.

How many is "some", though? More specifically, what portion of the market do they represent?

I'm actually genuinely curious as to whether anyone has a feel for this, as, if the proportion is particularly high, it would explain why even AWS's prices seem competitive to some managers who I would have otherwise thought would have an idea of how cheap commodity hardware is.

Previously, I thought only "enterprise" shops suffered from this, where the multiplier can easily be 10x for the most vendor-locked-in enterprisey hardware like storage, network, and support (where the multiplier can be infinity!) even if they pay no premium on the commodity compute.

> A lot of budget gets spent under the "nobody gets fired for buying xxxx" principal.

I do not think "Nobody gets fired for buying xxxx" applies to Intel server CPUs. Almost everyone I know that purchases or approves large hardware purchases take a lot of time considering specs and price. AMD and Intel's CPUs are completely interchangeable for anyone writing in a language level C or higher. In the case of cloud providers I am completely ignorant of what CPU my cloud servers are using, and only a vague notion that it's the X86 architecture.

I see "buys xxx" doesn't seem to resonate with hardware purchases. Others have commented similarly. What are the places that's true? Cisco switches & routers, Dell servers, Microsoft Exchange? Hrm... This could make an interesting ask HN.

As for your cloud provider, it's Intel. Almost guaranteed. The server market has been so run by Intel for years now, EPYC just launched leading to the chances of you having amd and it not being fairly big news story is unlikely. [1]

1: https://news.ycombinator.com/item?id=15851470

> I see "buys xxx" doesn't seem to resonate with hardware purchases.

Oh, "Nobody gets fired for buying xxxx" applies to some hardware purchases, just not server CPUs.

I'm excited about ARM chips from Qualcomm and Cavium, hoping to see open source commits to projects to support the architecture.

I too, am interested in the acquired work of Cavium.

There will be a lot of headlines about Intel being doomed. And they will continue to make money hand over fist, with very little loss of market share.

Contrast the times when everyone was certain Microsoft’s days of reign are over. Linux was cheaper, OS C was better, and in the end, nothing really changed.

Xeon is still better for complex workloads like databases since its inter-core latency is lower. Epyc is better when there is no shared data between cores. I'd say public cloud will pick up Epyc for such workloads. They may also use Epyc as a 4x8 core VMs with VM affinity locked to each die. Xeon will continue to be the standard for complex workloads.

> Xeon is still better for complex workloads like databases since its inter-core latency is lower.

Has that been demonstrated to be true, though, especially in light of Epyc's much wider PCIe bandwidth?

For example, would a high-concurrency OLTP-type workload suffer as much from increased inter-core latency as a single-query OLAP workload?

This question would benefit from answers that reflect what the large public cloud providers, such as Google, Amazon, Microsoft choose for their server farms. I would hazard to guess that they would use AMD as a lever on Intel and vice versa.

Reply on news.ycombinator.com