Coffee Lake Review: Initial Numbers on the Core i7-8700K and Core i5-8400 | Hacker News Reader

Coffee Lake Review: Initial Numbers on the Core i7-8700K and Core i5-8400

anandtech.com

91 pointspolskibus66 comments

"The problem here is the same problem we’ve seen with big core parts and Windows 10, however: these large processors can only take so much before having to move threads around, to keep both the frequency high and the energy density low. All it takes is for a minor internal OS blip and the single-thread power begins to dissipate. Windows 10 famously kicks in a few unwanted instruction streams when you are not looking, and as a result the average single thread performance seen on the 8700K might be equal or lower than that of the previous generation. It becomes an infuriating problem to debug as a reviewer."

I didn't know this was a big issue with multicore processors, very interesting. Does this mean that if you OC a single core to the limit to get the best single threaded performance, you will see drops in performance when your other cores start getting utilized?

I can't quite tell from the current replies if the mechanism of what is going on is clear. Apologies if this is redundant.

What is happening is a thermal management issue. "turbo" lets you boost briefly to speeds that will cause the chip to overheat if run long enough. Their idea to manage this was to play hot potato with which core gets baked. As soon as it reaches critical temperatures, they turbo up another core and the OS migrates threads.

So one part of the chip gets hot while the other cools off, and then they continue back and forth.Personally I'm surprised this doesn't crack silicon or pop bonds.

With regards to overclocking, I guess I didn't know you could overclock a single core? Is this pinned or is it adjusting what the "turbo" frequency is?

In anycase, if there is a differential in speeds among cores, the hot potato thing will probably happen. And if you overclock it will probably happen more.

The only way to combat this would be to either mess with the windows cpu driver and put your cpu at risk, or make the speed homogenous among all cores.

> "turbo" lets you boost briefly to speeds that will cause the chip to overheat if run long enough. Their idea to manage this was to play hot potato with which core gets baked. As soon as it reaches critical temperatures, they turbo up another core and the OS migrates threads.

This is not an accurate description of how turbo works. You don't have to migrate away from a core and switch to another.

Turbo simply recognizes that if you have a multi-core processor, and many of the cores stay asleep and don't contribute to power and thermal demands, then the remaining cores can take up the slack. The more cores stay asleep, the faster the remainder can go.

> mess with the windows cpu driver and put your cpu at risk

AFAIK, the thermal throttling is hardware controlled. I doubt you can override the protections in software. Driver tweaking may help

Does this mean that if you OC a single core to the limit to get the best single threaded performance, you will see drops in performance when your other cores start getting utilized?

This is how all CPUs with some form of "Boost" work, yes. People who overclock generally lock the multiplier to avoid it. So you will tend to have this problem when you do not overclock.

To further complicate things, some of the larger many cored parts have a favored core which will turbo higher than the others, but only if the OS schedules work on it.

Yes.

At least for their high-end server parts, Intel has docs on how many turbo bins you get. AVX can also slow things down all on its own (it reduces the available turbo). A co-worker of mine is (hoping to) speak at KVM forum about this as relates virtualization.

Ah, found some docs for the Core i5: https://www.intel.com/content/www/us/en/support/processors/0...

There's a nicer version of these somewhere with graphs, but I'm on my phone and it's early :)

It is called thread migration (or, more abstractly, process migration) and high memory clock speeds favor environments where this is expected to occur frequently.

As such, there is significant performance degradation when high-powered CPUs are paired with low-powered RAM. Incidentally, most OEMs/ODMs are susceptible to this as they frequently use high-powered CPUs as advertisement anchors but cut costs with low-powered RAM.

It is the source of unpleasant lag spikes for end-users and, as mentioned elsewhere, they are difficult to work around.

When you say low-powered ram, are you talking about clock speeds?

Another chip, another socket. If you were considering upgrading your CPU in place you can forget it. You're going to need a new motherboard, with almost exactly the same set of chips on it but a very slightly different socket.

This socket is probably only going to stick around for this generation too, the 10nm chips will almost certainly use a new socket.

What is the business advantage for Intel? Swapping out a CPU is trivial, the whole MB and CPU is a job and time commitment. Are there that few of us willing to make the upgrade or some major business advantage that outweighs the upgrade market?

Moving to 6 cores required different power delivery, so they needed a new socket. It just happens to have the same number of pins as the old CPU. I would be willing to bet the number of people who want upgrade their CPUs and currently have a Kaby Lake CPU would number in the tens of thousands. Engineering around that number of people would be nuts.

If you know you won’t do a cpu swap, you may spend more up front to have a more future proof system.

its the same socket 1151, but the pinout is different. According to intel because of different power requirements for hexacore and faster DDR4 rating.

Sure those are the technical reasons, but the real reason is to sell more mobos.

I'm sure they could have done that anyways. AMD kept backwards compatible sockets with the physically similar AM2, AM3 and AM3+, while increasing memory speeds, introducing new chipsets, and introducing their Hex cores.

I would prefer the option to accidentally buy an old motherboard and have my CPU not work at optimal settings than Intel's solution where the CPU won't work at all even though it fits.

I took the liberty of comparing the benchmarks online for the i5-8400 vs the Ryzen 5-1600: https://goo.gl/R4AQZe

For single-threaded performance, the i5 wins (no-brainer due to the higher turbo clock), but for multi-threaded performance, the ryzen is still a winner, for less money.

Power consumption _is_ higher for the Ryzen though. I'm thinking that the Ryzen is still the best bet for now as far as bang for your buck, considering it has hyperthreading AND can be overclocked, and the i5 does not.

Not sure if the i5 can be overclocked. If not, then the Ryzen will probably beat the i5 in single-threaded performance as well if you OC it.

Computerbase concluded that the R7 1700 and the R5 1600(X) beat (respectively) the 8700K and 8600K in price/performance for non-gaming uses.

Pretty graphs: https://pics.computerbase.de/8/0/0/8/6/diagramme/70-630.1507...

Somewhat surprising to see AMD in this situation given the console design wins and those being 8 core systems, but for now gaming remains Intel's forte, though.

That's because DirectX 11 and older have a single thread draw submission queue which is being a bottleneck.

DX11 still dominates - right now there are five (5) DX12-only games and only a handful announced, because it's so completely different you have to re-engineer significant part of game engine to get the benefits.

The i5 is cheaper, and actually outperforms the 1600 in most productivity benchmarks as well.

It also has an iGPU and if you are buying a 180$ CPU it actually counts.

If you care about integrated graphics. Most desktops will probably have a dedicated GPU, as well though. Which makes the integration graphics completely useless.

I would like to see the benchmarks that show the i5 outperforming the Ryzen 5 1600X in a multi-core setting.

In the <200$ CPU range it's not uncommon to see no dedicated GPUs, both form a budget perspective and for practical perspective an iGPU comes quite in handy.

As for the productivity benchmarks the Anandtech ones put it ahead or tied with the 1600 with most benchmarks other than CBMC. For more common productivity uses such as office work and video encoding the 8400 is faster, sure it won't beat the 1600 in Cinebench but how many people render in Cinema 4D on a sub 200$ CPU? And even in Cinema4D MC it's not that far behind.

At this point it's really not a "no-brainer" it's about price, if you aren't going for a dGPU the 1600/1600X might not be worth the premium unless the initial shortage of CoffeeLake is going to hike the price above the 180$ MSRP.

If you are buying a dGPU and will be gaming it's again not a non-brainer since the 8400 is faster, although at that point paying 70$ more for the 8600K and pushing it to 5-5.2ghz would be much better.

If you are looking to build an SFF PC or a NUC again the 1600 is not a non-brainer since you will need to account for a dGPU.

At best you can say is that the 1600/1600X are still competitive, but they are quite far from being a non-brainer. In fact they are only competitive because of the recent price cuts, if the 1600X was still priced at $250-260 the only non-brainer here would be the 8600K.

Heh, you made me realize that unlike AMD previous gens, Ryzen has no GPU inside. I wonder if they tried to make a Ryzen APU in labs.

Ryzen APU's (Raven Ridge) are coming this year or early next year, but they are limited to 4C/8T only.

Thanks I knew none of that.

Disregarding brand loyalty and whatnot, there are very few situations in which AMD's Ryzen chips aren't a "no-brainer" over Intel's.

Most rational people should be getting AMD's chips and AMD Ryzen-based devices.

I don't think the situations are as few as you report.

The heat/consumptions and single-core speed aspects are not negligible.

In addition (although this is definitely something for a minority of users), I suffered blocking VFIO scheduling issues with AMD (A10) while I didn't with Intel. Even Ryzen seems to be immature for VFIO (but I didn't personally test this).

Doesn't Ryzen have better power efficiency compared to these new Coffee Lake's?

The issue is a bit muddled because sites are (for some weird reason?) ignoring the most common Ryzen SKU, which is 65W (R7 1700).

I've noticed this as well. It seems outright bizarre that many benchmarks have been leaving out the 1700, when anecdotally, it seems to be the Ryzen that people are buying. It also seems to be the price/performance/wattage sweet spot for Ryzen, since it has low wattage, and also comes bundled with a cooler.

But the Ryzens being compared... consume quite a bit more power than the intel parts they are being compared against.

That said, yes, if power consumption is no item, the Ryzens look better. If you are looking for power efficiency, go with the intel parts.

No, no, not at all!

That's exactly what I was pointing at: most reviews are using the most power hungry Ryzen SKUs (95W), so of course it looks bad in comparison. Meanwhile, the price/perf leader and most sold Ryzen SKU is only 65W. This one is ahead of Intel in power efficiency.

Really?

I'm honestly looking for opinions here. I'm currently looking at the i7-7820X for a gaming/developer/GP machine. The Ryzen 1800X has only 16PCIe lanes, and a dual channel memory controller that is not as fast as Intel's. The Ryzen 1920X has what I need, but costs about $200 more, and I'm a bit concerned about performance consistency due to the NUMA architecture.

With this in mind, I'm really leaning towards the 7820x. But I keep seeing these posts saying that beyond brand loyalty, Ryzen is the way to go. Is the balanced ST/MT workstation niche really in Ryzen's court?

I took a hard, rational look at Ryzen for my latest workstation purchase. Ultimately I went with Intel due to all the reports of crashes under heavy SMT load on Linux.

That's understandable, but it's also worth pointing out that that issue is now understood and affects only processors manufactured before June 2017. Any affected processor should be returned for replacement or RMA.

You can RMA affected chips (I've done so), which were from early batches (apparently). It's inconvenient but it's not the first and won't be the last piece of hardware that has to be RMA-ed. At least you can RMA them! Intel didn't offer that when TSX turned out to be broken.

After compiling things on an 8 thread Haswell Xeon, and now a 16-thread Ryzen, I really want a ThreadRipper...

I considered ThreadRipper, but I'm going to overclock 8700K. Multiple cores are good to have sometimes (clean rebuild, reindexing), but 99% of time there are 1-2 threads and their performance is most important. If AMD would improve ThreadRipper IPC and frequency at least to 5 GHz, it would be much more interesting choice.

I'm looking at building a PC for deep learning, and I'm just learning about the importance of PCIe lanes for GPUs. I'm confused by the Coffee Lake numbers - 16 on the CPU, 24 on the motherboard? Does this mean that you could have 2 GPUs are without slowdown, or does the 'x16 per GPU' advice only take into consideration the CPU lanes?

reference: https://twitter.com/karpathy/status/869289384687714304

I'm on the same boat as you, building my first deep learning box as we speak. What I've understood is the following:

if you want to build, or grow your build over time to multi-GPU deep learning machine, you are better off server-targeted Xeon family processors instead of this consumer market oriented Core line. Xeon processors typically support 32-40 PCIe lanes, while Core processors support 16. I read that with typical deep learning loads, utilizing 8 PCIe lanes per GPU is good enough, so with 32-40, you can go up to 4 GPUs.

However, instead of spending tons of money and building a perfect basis for a hypothetical future DL box from the get-go, I decided to just start with cheaper parts and go with single GTX 1080 and Core processor. When I hit the limits of it, I probably know more what I need and likely the technology has advanced too.

Go for an AMD chip, you get 60-128 general purpose Pcie lanes.

Well, Threadripper AMD anyway.

The key is if PCIe lanes are important, you need to buy a Motherboard that supports a high number of PCIe lanes. AMD's Threadripper is cheap (for a workstation) and supports 64+ PCIe Lanes.

Intel's "extreme" line (both chips and motherboard) are the high PCIe count workhorses, as well as their Xeon chips. But they are much more expensive than AMD's solution. Even then, an Intel 6950X "only" has 40-PCIe lanes. I wonder if Intel's new line of "Extreme" will up the PCIe count to better compete vs AMD.

I understood that for a hobbyist, who doesn't hand-optimizate preprocessing, good single thread performance is better with numpy. From that front Intel is better.

You should probably get a threadripper and max out with whatever GPU/ML processor boards you need.

Will you be running Linux?

I find it fascinating how inexpensive CPU processors really are. You pay approximately same amount of money for top-of-the-consumer-line processor as you pay for 32GB of RAM or 1TB SSD drive. Modern consumer GPUs are twice as expensive.

Well the problem might be price fixing of RAM (speculation, btw). There's competition in CPU world but its a different matter in the RAM/HDD world. RAM has actually been getting more expensive.

https://www.reddit.com/r/buildapc/comments/6xth78/is_ram_jus...

Six-core Coffee Lake is 151 mm^2. 1TB of 64L 3D NAND flash is 1888 mm^2. 32GB of 21nm DDR4 is 1715 mm^2.

CPUs are way more complicated to design, but once they're in production, you get a lot of them per wafer.

It says Socket 1151. Can I just drop a Coffee lake CPU into My 1151 motherboard (potentially after a bios upgrade) or will there be caveats such as chipset model?

Unfortunately the strict answer to your question is actually yes.

The CPU won't work, mind you, but it does physically fit...

No, the socket is LGA 1151 but the chipset is different: 300-series vs 200. You need a new motherboard.

Why didn't they just use a new socket number for this? Even something like LGA1151B would have made this clear and easier to search product.

Technically its the same socket. But the voltage is different. Coffee Lake consumes more.

You could use Kabylake('200 series') processors in 100 series (Skylake) motherboards with a BIOS update. Both 1151 socket.

Shame this didn't continue.

Its because the power profile of the connectors are different. Its subtle, but it makes it incompatible.

Need a new mobo Z370

What's the usual step when companies brand their product generations with numbers and they get too big? I can't imagine calling something an "Intel 10400" in two years.

Nvidia and AMD both reset their numbering schemes when that happened:

After the GeForce 9 series (ie, the 9800 GT), Nvidia moved to the 100 series (GeForce G 150), which continues up to this day (ie GeForce GTX 1080).

AMD did something similar after the HD 8000 series, and switched to the RX 200 series.

Interesting that the jump from 900 to 1000 doesn't seem to be that much of a problem, but companies avoid the jump from 9000 to 10000. At least I can think of no example right now. Nvidia stopped after the 9000 series, Intel never crossed the 10000 mark in their "Core 2" CPUs ("highest I can find is Core 2 Duo T9900"), and so on.

It's easy to say ten eighty or ninety eight hundred, but 10800 lacks a convenient phonetic form.

"Ten, eight hundred"? But I see what you mean. Especially when you reach the teens, e.g. "17700", it would sound pretty bad.

286->386->486->Pentium

Pentium->Pentium II->Pentium III -> Pentium IV -> Core

If history is any guide they'll come up with a completely new naming and numbering scheme soon enough.

If I had to guess the new marketing name will be introduced at 10 or 7 nm, and be metallic or light-themed as a reference to the major process changes at that node (silicon->exotic semiconductors and/or conventional->extreme ultraviolet lithography).

I thought Pentium was just because the Trademark office was scoffing at registering just a string of numbers. So instead of 586, they called it Pentium.

Core was a total change in design representing a very different lineage than the current at the time Pentiums, so it makes sense that they'd change the name.

I'm not sure that I see a similar external pressure in this case.

More to the point, they were running away from the horrible Pentium 4 chips and wanted people to know that their machines wouldn't run like a dog anymore.

It's kind of hard for Intel to switch up the numbering on a product that is changed so little between revisions. Plus, having the 10nm chips use the 10 prefix seems like a good coincidence.

The ex-celeron.

You can take a look at graphics cards companies. ATI/AMD went from Radeon to Radeon HD to Radeon R5/R7/R9/RX. Nvidia went from Geforce to Geforce GT. In all cases, the name change happened after the 900/9000 series of cards, and the numbers were more or less reset.

Nvidia didn't reset after the recent 900 series of GPUs though.

Wake me up when Intel starts offering ECC support in their desktop processors. Until then Ryzen looks very tempting.

Reply on news.ycombinator.com