Intel’s Xeon Platinum 8284 CPU: When 300 MHz Cost $5,500 | Hacker News Reader

Intel’s Xeon Platinum 8284 CPU: When 300 MHz Cost $5,500

anandtech.com

58 pointsvarshithr63 comments

It's going to be interesting to see what the new AMD 'Rome' brings to the table when it gets into peoples hands.

upto 64 Cores/128 Threads and 128 PCI express lanes (that's on all of them irrespective of the core count) - 8 channel memory.

Given AMD's insanely aggressive pricing on the Zen 2 consumer line I'm expecting a proper knockdown drag out fight.

Does AMD gain a lot by driving down prices heavily? Should make sense only if they expect Intel won’t match their price cuts.

If Intel follows, then they have both just lost money. Might make sense if ruining the market drives the competitor out, but that’s not likely with Intel.

AMD CPUs should be cheaper to manufacture, because their chips are much smaller. Xeons are built using giant monolithic dies, so the defect rate per wafer matters a lot. EPYC processors from AMD are built using 9 separate chips under one package [1]: one I/O die and several chiplets with multiple cores each, so the yields should be much higher.

[1] https://images.anandtech.com/doci/13561/amd_rome-678_678x452...

Seems like there's only up side for AMD. They're not getting Intel-level prices now because they're not being used, so as long as they're not selling below cost, it's all upside to them. For Intel's part, they only have to follow suit if/when the strategy starts working for AMD, by which point AMD may have clawed back enough enough market share to have justified the strategy. At that point the race to the bottom probably ends, and prices might incrementally inch up somewhat, but the increased competition will still probably depress Intel's prices over what they would have been, while AMD still has gains.

That’s certainly my hope, AMD increasing core counts forced intel to follow suit or lose on multibench workloads to sometimes hilarious extents even when they where ahead on IPC per core.

I don’t care which one wins I just want proper competition again.

Same. I never particularly cared whether I had an AMD or Intel proc when price and quality were comparable. I'll happily choose whichever has the better value proposition. Though Info think AMD had a particularly uphill battle in the server market dominated by clouds that are mainly standardized on Intel. But with such price benefits, I can imagine various providers experimenting with blocks of AMD based compute.

In my opinion they are not driving down prices heavily. They are just pricing their chips competitively so people will buy AMD over Intel, but they are not super cheaper. The pricing of AMD chips just makes sense if you look at the lineup and compare it to Intel.

The number one chip on passmark right now is the 3900X which is beating intel units at nearly 4 times the cost.

Now passmark isn’t 100% accurate but it’s usually pretty close.

https://m.cpubenchmark.net/high_end_cpus.html

I’d call that aggressive myself.

Intel had been considered to get insane premiums on their chips for a long, long time...

Yeah, but we're talking server chips in TFA.

True but AMD hasn’t released Rome yet, given that in Benchmarks I posted the 12 core 24 thread is out performing the Xeons and given that Rome is a zen 2 based core I think it’s safe to extrapolate that AMD have a competitive or better processor portfolio compared to intel at this point.

It does work because price cuts give Intel an allergy. Intel has only recently hinted at reducing prices of some high-end CPUs "up to 15%” 2 years after zen came out and after it had already raised prices on the same CPUs by at least 15% since it started struggling with manufacturing shortages last year.

AMD should go as aggressive as possible with the pricing to steal market share from Intel because it has nothing to worry about in regards Intel responding with a price war - it just won't happen. Intel couldn't even bring itself to do that with it's mobile chips - twice. First with Strong-arm and then with Atom. It lost the mobile market twice because of its insatious margin greed.

Since they’ve lead the market so consistently for so long I think they’ve never been in a protracted situation where aggressive price cuts where required.

AMD to their credit have played everything blindingly well with nary a misstep in the last two years.

It must be hard to be intel and have your seeming invincibility in the x86 market challenged by a company that nearly went bust not that long ago.

If I were in charge of Intel,I'd have bought AMD when they were struggling...

That would invite a monopoly investigations in the EU and possibly the US pretty much instantly.

Ironically intel needs AMD to prove they don’t own the whole market, it’s just blown up this time because AMD nailed it on the architecture and fab side.

I was thinking about but still went posting my comment.You are right on monopoly.

Antitrust enforcement would probably take issue with that

My biggest question is... what percentage of the market is really buying the latest generation of processors upon release? Gamers, sure. I'd imagine the number of "every day use" laptops sold is a lot larger than people trying to eek out the last few frames per second. Would these budget machines also have this new advanced AMD CPU? How long does it take "the kind of laptops you pick up at Walmart" to deprecate old CPUs and implement new ones?

>My biggest question is... what percentage of the market is really buying the latest generation of processors upon release?

That really isn't relevant. Most consumers buy a new computer when their old one breaks or becomes unbearably slow. Most corporate buyers have a consistent 3-5 year procurement cycle. AMD don't need everyone to rush out and buy new computers, nor do they particularly want it - huge spikes in sales are detrimental to the supply chain, as we saw with the GPU mining fad.

AMD want and need to consistently beat Intel on value over a protracted period, so that whenever a customer does choose to upgrade their computer they choose an AMD processor. At the moment, they're looking good in terms of both architecture and fabrication; the main problem AMD is likely to face is with partner relationships, because of Intel's historical dominance and use of anti-competitive practices.

Majority of data center users replace their gear within 3-5 years. There are outliers, of course, who replace hardware with each generation or those who run their servers into the ground, but 3-5 years is far more common, so 4 years on average.

This rough math means if AMD takes half of the new CPU orders, they should have about 10% server CPU market share within a year, consistent with some analyst estimates [1]

[1] https://www.digitimes.com/news/a20190328PD200.html

AMD couldn’t have timed bringing a genuinely competitive product to market either on the backside of the Meltdown and MDS and Intel struggling on process.

Will be interesting to see if intel begin pulling the same kind of shenanigans they did back the last AMD was competitive (early Opteron/Athlon days).

I don’t think the market structure is the same and given they got slapped for it last time by the regulators I’m curious to see if they’ll try.

Anecdotally I know several people whose 'main' desktop PCs are now right around 3 years old. Mostly stuff from one or two intel generations prior to the introduction of the Ryzen 1xxx series several years ago. After going through all of the reviews for the Ryzen 3xxx (3700, 3800X, etc) and X570 chipset motherboards, they've upgraded and are really pleased with the performance.

It's fun seeing something that can chew through kdenlive video editing of 2160p, 60fps video and encode a project into HEVC like a hot knife going through butter.

The only problem is that such increases in speed are really only needed either for gamers or a small subset of people who work with graphics/engineering/video.

Lots of other uses for an 8-core desktop workstation... Lots of guest VMs inside virtualbox, a host operating system such as Qubes which is a xen hypervisor, etc. Not just gaming and very cpu intensive things like 2160p60 video encoding.

then you can get a 3200g which has 4 fast cores and a decent gpu for $99! its amazing.

For a lot of the server market, I suspect Intel, being a known quantity, will remain where it is. Bringing a new design to market is expensive (tooling, supply chain, warranty, maintenance, product life cycle, etc) and I'd need to know I'd be able to sell at least a certain number of boxes over a given time to make it worth it.

If all my tooling and supply chain is geared towards Intel on the server lines, it'll take a much better value proposal, volume guarantees, and the threat my competition will do it first to make me add a new line geared towards AMD.

OTOH, for cloud powerhouses that design their own hardware, even a marginal improvement in TCO/performance will tip the scales.

Then also the market isn't only about the highest-end CPU. There's plenty of market for lower-end CPU's. Both AMD and Intel have their offerings and production capacity. The market shifts will be marginal I think.

I'm having a hard time figuring out who would be buying this. Only thing I can think of is that the chip has some extra features not really advertised that the customer needed.

The finance industry have some really whacky requirements. HFT firms often use heavily overclocked servers with HEDT processors, because microseconds of latency on a server request have direct monetary implications. I can't think of an application that would justify paying $5000 for an extra 300MHz of base clock, but I can imagine that someone needs it.

If your DC had limited rack space it might start to make sense. Buying a cheaper processor means you would need more of them to accomplish the same amount of work, but also you need more space.

>If your DC had limited rack space it might start to make sense

Not just rack space. If workloads have to be on the same machine or socket (not NUMA/RDMA friendly), your only choice is to get bigger and bigger processors.

$5k is barely a 1 week fully loaded cost for an experienced SSE.

If I can spend $5k to save a week of optimization work, it's a good deal. Also buying clockspeed can't create new bugs either, which is a risk with any code change.

In my experience, most companies don't think this way, but definitely should think more like this.

That said, companies that need more speed often don't need that extra speed on just one node. So that's really $5K x N nodes. When you consider that just the CPU costs $15K, and not the complete cost of a system, I think it's a lot harder to justify. You better have that next lower tier $10K CPU and know that you need just that little bit more. And then... you probably ought to be investing in scaling out horizontally while that $5K x N buys you time.

> Also buying clockspeed can't create new bugs

Famous last words

> Also clockspeed buying can't new create bugs.

> buying clockspeed can't create new bugs either

Some software does not like high clockspeeds. For example, game engine in GTA V starts acting up when framerate gets too high.

It’s similar to why you used buy RISC vs Intel. Sometimes decisions like this are driven by by software licensing. $10-20k on a server may offset $200k in Oracle.

If you look at a longer time horizon, there are other dynamics. Your contract may not allow you to reduce license spend in three years, but you can shrink your hardware spend.

There's some stuff that isn't horizontally scalable either intrinsically, or the cost benefit hasn't gotten there yet. Back in the day I heard rumors.that Borg scheduler nodes needed a lot of cores on the local box _and_ a high per thread IPC. I could see Google happily spending $5k for 300 more MHz.

There have got to be customers out there for whom 10%+ base frequency performance is worth well north of $5K.

No reason not to capture margin for a handful of top-binned chips.

There's plenty of times that extra compute capacity saves money, but not spending $5K to get a relatively small bump in clock on one chip. $5K will get you way more compute power if you just spend it another machine.

If you pay licences per machine/cpu socket/cpu core, those $5k might be your cheapest option.

Sure, if what you’re doing is easily scalable to another machine.

There’s a lot of software out there, legacy and otherwise, where the throughput ceiling is what you can do with one box.

Just the mere fact that such a processor exists,is not in the official catalog and costs substantially more, tells me that there's adequate demand for these or at least the particular customer has very deep pockets if they can ask for custom build.

Reminds me of the mystery largest buyer of glitter: https://www.nytimes.com/2018/12/21/style/glitter-factory.htm...

Here are some good comments and discussion about the possible industries.

https://www.reddit.com/r/UnresolvedMysteries/comments/aop378...

tldr; Most likely food, makeup, toothpaste etc. Especially food safe and biodegradable glitters. They are also located close geographically (New Jersey) to those industries.

> I can't think of an application that would justify paying $5000 for an extra 300MHz of base clock, but I can imagine that someone needs it.

Here's one: EVE Online servers. Back in 2013, CCP had at least one 4.4GHz Xeon they used to host fleet fights with thousands of players:

https://forums-archive.eveonline.com/message/3917593/#post39...

Why? Because EVE is mostly written in Python, which is both very slow and, thanks to the GIL, effectively single-threaded. It's a fundamentally limiting architecture, but that extra single-core performance let them push it just a little further – which for them was (is?) well worth it.

maybe someone who runs rendering farm with CPU utilization is always in the 99%? baffles me too.

Render farms scale horizontally, you can just buy more CPUs.

$5,500 doesn't matter in this context which is why the pricing is structured this way. This is destine for a server in a datacenter, the cost of the hardware is pretty much a blip in comparison to rackspace, power and networking costs.

This largely comes down to a problem of folks assuming that 'price' is correlated strongly with 'cost'. It's not and never has been.

I assure you that $5500 is a substantial addition to the TCO of buying and running the server.

The reason these parts exists is for optimising for expensive software that's charged per core. For the same performance bump, you can add two cores and pay Intel $2000 and Oracle $10000, or you can buy a chip clocked higher and pay Intel $5000.

> I assure you that $5500 is a substantial addition to the TCO of buying and running the server.

I assure you it's not. This thing is going to go into a 2U server, which is going to run ~$400 month for the rackspace, power and bandwidth. It's also not going to be purchased by folks building their own stuff from Newegg based on supermicro chassis and misc parts, it's going to be bought from Dell and HP as one part of the expense of a $30k+ server.

Could I build something cheaper and find cheap local colo to make it a 'significant' portion of TCO? Yes, doesn't change anything about the reality that this is not the target market so it makes no sense to be concerned about the cost in such a market. TBH if this segment of the market is your concern, you should buy a threadripper and hope the ECC ram you bought works with it. But, it's not the same market.

The numbers you're quoting don't match your claims. You said "the cost of the hardware is pretty much a blip in comparison to rackspace, power and networking costs."

Then once you start listing numbers, the server costs $30k+ and the data center $400/mo. I'm pretty sure that $30k for a server is not a "blip" on ($30k + $15k) assuming a three year lifetime.

Trying to make out like I'm coming it from some kind of hobbyist market segment point of view is a puzzling straw man. These products, which optimize for performance over performance/TCO$, are built with very valid niches in mind, and where the math works out it saves you a ton of money in software licensing or software development.

They are not, as you seem to be claiming, positioned as sensible choices for wide deployment in datacenters. If that were the case, why bother with the rest of the SKUs!?

Would someone buying in volume even have to pay $5500?

Hmm, isn't the new Mac pro up to 28 cores? :-D

I built a dev computer out of some old Xeon E5-2670 V2s (10cores/20threads @2.5ghz) each. I thought it would be way faster than a single core i7 4 core at 4.5ghz. To my surprise, only a few special types of things are the 20 core machine faster than the 4 core machine.

I mean, there are few special types of applications that bothered to implement any kind of multithreaded architecture until very recently; and even today its often not the default (-j for gnu).

This is precisely the reason that moore’s free lunch ending matters so much; everything needs to be rearchitected such that the server is faster than single core across normal usage

But we’re not at all close to such a world

A very funny illustration of you point: https://m.9gag.com/gag/a6OvN3q

Did you find that `make -j20` (or -j40 if you have a total of 40 threads) was five times faster (ish) than the quad-core machine?

My tangentially related experience: when I switched (after 4 years) from a Core i5-3570K (4C/8T, 3.4 GHz nominal) to a Ryzen 7 1700X (8C/16T, 3.4 GHz nominal) last year, multi-threaded build time of a smallish C++ project I work on went down from 1 minute to 30 seconds.

I'd really like a comparable speed-up on my next HW upgrade; hoping for affordable 16-core CPUs 3-4 years from now.

Releasing in September: https://www.amd.com/en/products/cpu/amd-ryzen-9-3950x

The 3570k is not hyper threaded from what I understand. If it was, I would have been thrilled because I am still using that professor myself, waiting for the new threadripper.

Of course, my mistake.

I also updated the 1700X to 2700 this year, the "marginality problem" ([0]) affecting some of the early Ryzens (random segfaults after some minutes of heavily parallel workloads with lots of process creation - e.g. builds!) was getting more annoying when Rust needed to download and compile a bunch of crates. I'm very satisfied - more or less the same performance (3.2 vs 3.4 GHz), but 65 W instead of 95 W TDP.

[0] https://www.phoronix.com/scan.php?page=news_item&px=Ryzen-Se...

Yes, in fact, that is one of the few times I get to hear the glorious sound of all 20 cores come to life and it does some serious number crunching. But very often I find that one thread will be working 100% for a bit while the rest just idle. (it's especially obvious with heavy javascript parsing or large text file editing for example).

Reply on news.ycombinator.com