Hacker News Reader

Comment by gigatexal | Hacker News Reader

In an interesting (to me at least) move intel instead of competing on price when AMD put out a competitive offering they decided to keep prices the same more or less or increase them by adding value. Faster AVX units, more cores, more SKUs (not sure about that last one). And while intel has margins to maintain AMD has put features into their chips for years. For example AES has been shipping in most every desktop and server CPU while only some intel ones get it. Most AMD chips can do VT-D etc while the same feature is relegated to higher end chips generally on intel.

View on HN

I could be wrong but I assume that there's enough inertia in the high-end business-to-business side of things (the part that matters for high-end server processors) that they probably can keep milking that cow for a few years more.

People who have only been buying Intel CPU in the last decade won't jump ship immediately. Nobody gets fired for buying Intel CPUs, at least until very recently. You know you won't have compatibility issues etc...

Of course within the next few years Intel will have to react somehow, either by outperforming AMD significantly to justify the premium or by slashing prices.

I know I hope AMD can keep up the pressure and keep iterating.

> Nobody ever got fired for buying IBM.

Replace IBM with Intel for the 2010s version.

And yet there are still people out there buying IBM and somehow not getting fired for it.

Do you frequently see people getting fired for purchasing decisions?

Systems Engineer here;

Absolutely.

Good example is that we bought 1,000 servers, 256G memory, dual socket Xeon E5-2687w v4 etc;

The things didn't come with a TPM module and it was a business requirement to run:

A) Windows

B) Full Disk Encryption

TPM modules were not available to us in such quantities.

So, um, ... what happened to those now "useless" servers. Wondering for a friend ;-)

Available software constraints choices in many of those cases. Put two choices in front of an application team exec: rewrite and port an application from POWER/AIX, POWER/System i, POWER/System z for $X million, with all the attending risks that entails. Or spend same $X million on new IBM hardware, and get more capacity, performance, etc., while the system administrators/system programmers update the underlying OS, with lower magnitude of risks and impacts, and if the sysadmins screw up, that's another department, not yours, anyways. Even if porting is a one-time charge, and saves Y * $X million in N years (where N usually <5).

The choice to avoid risk is easy.

This is why, by the way, selling to C-suite, or above middle management, as much as it is hated in technical circles, is used. The higher you go in management, the more risk is able to be shouldered.

If Intel hypes up any AMD compatibility issues Intel can find, then that urge to avoid risk will play to Intel's advantage.

Port away from? We moved everything TO POWER/AIX.

If its not confidential, could I ask why? IBM has been pivoting POWER to Little-Endian and has been focusing more on Linux with Red-Hat, SUSE and now Ubuntu available.

Power is big iron reliable. if you can pay for it it likely pays off. Probably easier to manage a handful of million dollar Power/AIX setups than 1000s of servers.

I don't follow, this is the same POWER 'big iron' just running Linux instead of AIX.

The hardware is where the magic is not the software.

To complement siblings answer, there are several types of failures where PC hardware will just crash and burn, while POWER will just work as usual.

Yep, I've seen that, too. There are some use cases where IBM still excels. There is a popular perception that IBM is a has-been dinosaur, especially with the System z (mainframe) stuff, but I still haven't heard that Visa International has moved even one of their three world-wide super data centers off of that platform. The Linux world could avoid reinventing a lot of wheels in the future looking to the lessons IBM has publicly published about in the mainframe world. Even POWER/AIX is scoffed at in many Linux circles.

I've certainly seen my share of underutilized System p servers, but in the right hands, they're a powerful option, with features that Linux still hasn't caught up to yet. For a cool feature that I would love to see on Linux someday, peek at checkpoint-restart, which enables some really sweet capabilities [1]. Generally, anytime I hear a business tell me uptime reliability is nearly a blank-check requirement (when downtime losses are described in 5+ figures per minute and up, and it is a 24x7x365 operation), but they must still stay on a Unixen box because that is what their IT shop knows, and it's not a parallelizable application, I start looking in AIX's direction. It can be a right PITA to work with if you grew up with Linux or BSD, but once you figure out some add-ons to use, it gets much more pleasant to use.

[1] https://www.ibm.com/developerworks/aix/library/au-aix7.2.1-l...

I find sad that many aren't aware how many cool features are available in IBM and Unisys mainframes, still not available in mainstream computing.

The AS/400 (IBM i) and ClearPath are my favourites.

One with its TIMI architecture and kernel level JIT, the other with an history going back to the 60's using an almost memory safe systems programming language, while people are still debating how to replace C.

Is there an article that compares the features present in Power vs x86 and this almost memory safe language?

I don't touch Aix since 2003, but really enjoyed using it.

Just curious, do they still support the Windows/Symbian style of dynamic libraries, with import libraries?

Sometimes IBM are better than the alternative. Hard to believe in the HN bubble, I know, but there you are.

(The trick is not to end up buying a bunch of rubbish IBM products while you're getting the one you want.)

I suspect anyone ordering top-bin CPU skus by the pallet-load has an ongoing program for evaluating performance-normalized TCO of at least a few candidate platforms based around a few candidate processors.

Heck, why has Google been investing heavily in OpenPOWER? I bet it's as much as keeping the option there and making it as attractive as possible: gives them a far stronger negotiating position with Intel.

I saw mentioned in another thread that all AMD processors support ECC RAM, provided you pick a motherboard that does it.

This alone could save you thousands.

Not really. Low-end Xeons with ECC cost >$300.

Sure, but with a $300 budget on Intel you're choosing between an i7-7700K or a Xeon E3-1245v6, which are all but identical.

The Xeon has a slightly higher base clock, slightly lower turbo (all clocks are changed by 0.1 GHz), ECC, and a locked clock multiplier. It appears to benchmark consistently a couple of percent below the i7.

What's even the point of that product segmentation existing? To annoy their customers? I'd rather just buy a $300 processor that checks all the feature boxes instead of all but one.

http://www.cpu-monkey.com/en/compare_cpu-intel_xeon_e3_1245_...

Best I can come up with is binning. Maybe some CPU's fail on the ECC front, so they sell those as 7700 and up the factory boost clock to compensate and 'market to consumers'.

Or maybe the MSRPs vary wildly, but in practice prices converged?

> Best I can come up with is binning

Bingo.

The memory ECC is a tiny bit of logic right next to all other memory logic in the MC, it seems very far fetched to me that there would be a binning process for specifically the instance of that logic being broke but the data logic next door working just fine.

Note that the ECC on the processor's internal data buses and caches is completely independent of whatever ECC the memory controller may or may not use.

The product segmentation is basically for different markets, but also for process maturity.

Mobile chips are very small, under powered, and under clocked. They get fabbed first.

Then your core i3/i5/i7/E3's are all fabbed. As the process matures and can be expected to regularly handle 4GHz.

Next your E5/E7's get ran, normally in a few waves. Your _enthusiast_/workstation cores are generally fabbed with E5/E7 engineer samples which get sent out for testing. These have [1] 20+ cores so you need to know your process is stable so you don't have to sell >50% of them as 8-16core Xeons.

---

But yeah the product segmentation is mostly to gate access. Only Xeons offer ECC [1].

And only [1] i5/i7 offer unlocked multiplers.

Ultimately this started to be a organization issue. Consumer sales (i3/i5/i7) doesn't even talk to enterprise sales (E5/E7). Consumer support can't answer questions about enterprise chips.

My personal opinion is the only reason Intel is still in business is they have a monopoly. Their anti-consumer practices are down right twisted.

[1] can/only are strong words there are a lot of exceptions/gotchas in Intel's product lines. Celerons/Pentiums sometimes have ECC, i3's and Pentiums have had unlocked Multiplers. There is a line of Xeons with unlocked multiplers. IDK

> Only Xeons offer ECC

> And only [1] i5/i7 offer unlocked multiplers.

These statements are false, which you yourself acknowledge later in your post. You don't get credit for noting a caveat when the caveat is that your point is entirely false, at least in my book.

Intel no longer manufactures 2-core Xeons so Pentiums and i3s support ECC nowadays. This isn't uncommon, this has been the norm literally since they stopped making 2C Xeons (Haswell?).

And there are currently unlocked i3s in the lineup, just like there have been unlocked Pentiums in the past. It's the norm that Intel offers a barnburner overclockable SKU somewhere in the low end. They don't always, but they do it probably every other generation.

There has been one Pentium and i3 that are unlocked since the start of the Nehalem era.

What's up with the i3-6320 supporting ECC but the i3-7320 no longer supporting it? Intel is at the least, very inconsistent with ECC on the consumer side.

ECC was probably not for consumers. I think they needed to have an offering for processors smaller than E3 for pro needs, and somehow merged it with some i3. Also check for embedded options: those are basically LTS, so low volume and/or high latency designed specialty electronics can use them without having to update all the time. I've not checked if there still are some offering on some newer gen i3, but Intel can very well have internal restructuring between their consumer and their pro-branch, which I remember having read are not very integrated...

I once heard there is certain mentality that when given more options, more people end up making purchases where they otherwise would not. For instance, if you have one fast food chain along a short stretch of road, less people will go to that one location than if there were multiple, competing chains in the same area.

I believe this is called agglomeration[0][1], and may have parallels beyond physical, competing stores, into realms like similar, but differently named CPU's.

[0]http://journals.sagepub.com/doi/abs/10.1177/0042098017694131 [1]https://en.wikipedia.org/wiki/Economies_of_agglomeration

edit: rplst8 mentions binning, and that seems like another likely factor as well.

Sure, if you're tinkering at home where the stakes are low and your time is free. You can also save a bunch of money running Xeon engineering samples.

But if your production environment goes tits-up then your savings go out the window and a whole lot more besides. And that's what you're risking by running a hardware configuration that is non-validated/unsupported.

People constantly make the mistake of assuming that businesses need to pinch pennies on their money-generating assets, just like they do on the rigs they use for entertainment at home. In the business world, reliability and official support trump an extra $1k in your pocket every time.

Forget Xeons entirely: there are still a lot of customers who buy Sun/Oracle/IBM hardware even though it costs tens or hundreds of thousands of dollars extra. $1k does not even register for a lot of businesses.

(ECC is validated on Epyc, of course - but the prices are a lot closer to Xeons there, it's no longer a slam-dunk. You have to look at it on a case-by-case basis to determine which is actually a better deal for your workload. And frankly AMD has been doing an undue amount of massaging on their benchmarks here, like in their launch data where they just straight-up removed 43% from all the Intel benchmarks. )

http://techreport.com/review/32125/amd-epyc-7000-series-cpus...

A fair point, and well articulated!

ECC is not an officially supported/validated feature on Ryzen. Support comes and goes across various motherboards and BIOS revisions and if it doesn't work you have no leg to stand on. Also, UDIMMs only, no RDIMMs.

It is officially supported on TR and Epyc, as well as Pentiums, i3s, and Xeons (entry-level Xeons are nowhere near as expensive as people think they are).

If you think you need ECC my advice is to build out around a platform where you won't be the first person to try something and run into the bugs. There is not a large community of ECC-users on Ryzen.

Ryzen doesn't support RDIMMs, only EPYC do. AFAIK ECC support is broken on Ryzen even with UDIMMs.

I suspect Intel is not too worried about AMD because they've seen this script before. When AMD's Opteron server CPUs came out they offered vastly superior performance and scaleability compared to contemporary Xeons. Even with a clear technical advantage, AMD never came close to matching Intel's volume in the server market.

Today, AMD's Eypc CPUs are a much closer match to their Xeon competitors than Opteron was. So if history is anything to go by, Intel has little to fear from AMD's latest efforts.

AMD wasn't able to match Intel's volume because Intel engaged in outright bribery of PC and server vendors on a impressively massive scale (they literally paid out hundreds of millions per year — and that's just to one vendor — as an incentive not to buy AMD). Other factors obviously didn't help, but this was clearly a huge one.

Given they've been slapped down for this, I wonder if they'll try it again. If they don't, AMD could have an easier go this time round.

I don't completely understand the "need" for bribery. Could they just not lower their prices? (which could have resulted in the same balance in the end)

I think you're confused about "couldn't match their volume". Parent means (I think) that AMD couldn't _sell_ enough units. Not that they couldn't _make_ enough units (which is I think your interpretation).

They couldn't sell enough because someone was paying their customers not to buy their (superior) product.

I think temac was suggesting Intel could sell cheaper chips. Why would Intel need to bribe the competition?

To maintain product pricing and value perception.

And to block competitors from selling out their own products (AMD in that case).

Yes

There are few things at play. And this time is a little different.

1. Opteron in the early 2003 to 2004 did have a good run. It forced Intel to lower prices as well as using dirty tactics to compete. Let's not forget Opteron was AMD first entry into Server market. AMD then failed to execute after that.

2. Software compatibility was a thing, and may still be a thing judging by the comments here. But remember Datacentre is now largely dominated by Cloud and Web Host Providers, and most of them depends on Open Sources Software. All of a sudden, Google Amazon and Microsoft have incentives to improve software compatibility, if there is such problem exist.

3. IO compatibility / performance used be a problem for AMD. I am surprised it is not mentioned anywhere. Intel has the best IO ever with their Chipset. From PCI-E, SATA, AGP etc. But now none of these IO implementation are secrets anymore. Thanks to the SoC IP industry you can get the best and leverage the battle tested IP from other vendors. And not to mention the really only IO left is PCI-E. Which is very well defined.

I expect there will be lots of interest to test out EPYC, but the real volume will be Zen 2, when the industry has settled on EPYC being good enough. And I cant wait to see how Intel would react, after all if you look at their recent management teams changes it seems they dont have a clue what they are doing.

Reply on news.ycombinator.com