I'll sure buy one more iteration of Skylake with 10 (20) cores but not until it matches or exceeds Socket AM4+ pin count, with 1350 pins minimum and 40 PCIe lanes direct from the CPU... oh wait, that's LGA2066 and Core i9-9900X. OK then, so probably never.

10 cores would be under the 12 that AMD already offers at very respectable single-core performance. Not to mention that 2020 is supposed to be the year when zen 3 launches, so there's even the possibility that these 10 cores will be put against faster 16 core processors. Considering the fact that Comet Lake is 14+++nm, I think this will be DOA, unless there's some very deep price cuts and they can somehow squeeze even more performance from their 14nm process. I mean, they've been doing it for what, 5 years now? How much more refined can it get?

Oh how little I care at this point. Until they get 10nm to work or scrap in favor of 7nm I don’t care to see another product from Intel.

Isn't Intel's process more refined though the 10nm process seems to have been incredibly troubled and I don't know how Intel anticipated AMD's Zen, Zen+ and current Zen2 architecture and how it would stack up. Then again the nm stuff is mostly marketing anyway from what I recall and for users it's more about features and performance, pricing too of course and then things like sockets and getting a fan or water kit mount and of course motherboards and all that stuff. Guessing Intel will also start doing something about the various exploits and potential performance issues from workarounds and fixes here but no idea if that's 2020 or later and who knows what sort of hardware changes will be needed. EDIT: "Isn't Intel's process more refined...they've sure had a lot of issues with it though." *Thumbsup* Yeah..about that.

DmitryKo:

I'll sure buy one more iteration of Skylake with 10 (20) cores but not until it matches or exceeds Socket AM4+ pin count, with 1350 pins minimum and 40 PCIe lanes direct from the CPU... oh wait, that's LGA2066 and Core i9-9900X. OK then, so probably never.

Why exactly does pin count matter? Most of the pins in a CPU are ground, voltage, and dummies. If the CPU's wattage decreases, you can actually shrink the pin count, rather than add more. And what are you going to do with 40 PCIe lanes on a mainstream platform?

Loophole35:

Oh how little I care at this point. Until they get 10nm to work or scrap in favor of 7nm I don’t care to see another product from Intel.

I just want to see something that's actually new and interesting. By the time Comet Lake comes out, it'll have been 5 years since we've seen something mildly interesting, in a good way. I say "mildly" because going from Haswell to Skylake wasn't a big jump, but, at least it actually had noteworthy changes. As far as I'm concerned, Intel's 10nm has been ready over a year ago, but they can't switch to it because they can't achieve high enough clock speeds on it. Right now, high clocks (both for core and memory) and single-threaded performance is where Intel has the most leverage over AMD, otherwise, they're pretty much worse in every other way (I'm not sure if Zen2's AVX performance is caught up with Intel's or not, but, most things don't need AVX). The gains from shrinking by 4nm isn't going to make up for the performance losses.

Is it so that Intel won't switch to a new architecture (would be about the time) before they can make the CPUs using the smaller process technology? So, if they can't get it up, they will just keep refining the old Core arch, and increase the core count by two, the MHz by 200, or something.

karma777police:

Intel is making a big mistake here. If I was them I would kill socket for 9900k and lower SKUs and bring 2066 socket to mainstream therefore people who already have that socket do not have to buy new mobo. New chipset is fine but same socket so newer CPU can work on new and older mobo. That would be the smartest thing to do. Also lower the price of 2066 socket CPUs to match Zen 2 and case closed. They don't have to worry about core count cause they already have that with 2066 socket. If they want to release an updated CPU sure, to gain some performance and have higher clock speed. Since 2066 socket already have core count, if they are any smart they would offer 18/36 core CPU for $499 to undercut Zen 2 big time. AMD cannot squeeze more than 16 cores on the current AM4 socket and Intel's 18 Cores for $499 would seriously hurt AMD. Then offer 16/14/12/10/8 cores and each SKU seriously undercut Zen 2. In nutshell Intel Strategy for near future should be More cores for less money Quad Memory Channel vs. Dual Memory Channel again for less money Cheaper 2066 socket motherboard compared to X570 Single CPU performance given clock speed won't suffer compared to Zen 2 as it already beats Zen 2 in gaming This would hurt AMD big time. In mean time work on new gen. CPU (10nm).

Seems an extreme plan. This would be everything but not smart.

How does the clock speed a proc can hit work in relation to fab size in nm? Are they not directly related? Always wondered how Intel can reach such high clock speeds when there are smaller nm dies that can't push past a certain point.

NCC1701D:

How does the clock speed a proc can hit work in relation to fab size in nm? Are they not directly related? Always wondered how Intel can reach such high clock speeds when there are smaller nm dies that can't push past a certain point.

That's pretty much exactly why I made my comment about Intel's 10nm being ready, but not capable of sustaining the high clock speeds. There's only so much you can do to force physics to behave the way you want. Intel can make 10nm chips but because of stuff like quantum tunneling, they can't push the voltage high enough to maintain the clock speeds people are expecting. And since they aren't doing anything to improve IPC (beyond shrinking the die by 4nm), it doesn't make sense to release any high-end or overclockable chips because the overall performance will be worse than their existing 14nm node. Ironically, they've just been digging themselves into a deeper hole by bumping up the clocks every generation they couldn't get 10nm out the door.

NCC1701D:

How does the clock speed a proc can hit work in relation to fab size in nm? Are they not directly related? Always wondered how Intel can reach such high clock speeds when there are smaller nm dies that can't push past a certain point.

It's semi related. Cell libraries can be optimized for switching speed, density, power, etc. Critical path weighting/optimization becomes a big factor too - keeping the entire chip in sync, designing stages for that, etc. Typically a transistor can switch a lot faster then what a processors rated clockspeed but everything else keeps it limited.

schmidtbag:

Why exactly does pin count matter? what are you going to do with 40 PCIe lanes on a mainstream platform?

I was basically saying that even though Intel obviously planned their response for Zen2 announcement well in advance, yet they probably underestimated the impact and underdelivered on the Comet Lake specs. So it would probably take them another iteration in the form of Tiger Lake silicon (Willow Cove architecture) to get closer to AMD SocketAM4+ which I suppose comes with Zen3/Zen4 and introduces DDR5 memory - and match AMD's 1350+ pin count too 🙂 As for PCIe lanes, everything is PCIe these days - SSDs, video cards, chipset links, even flash memory cards (QXD/CFexpress and SD Express), as well as USB4/Thunderbolt and DisplayPort 2.0. The old days when chipset was handling DRAM, expansion cards and all kinds of legacy BIOS 'super I/O' (for COM LPT PS/2 8254 FDD PATA PCI etc. based on LPC bus) will soon be gone. Yes, say goodbye to booting your 1987 DOS 3.3 USB-floppy images. Honestly we don't even need a chipset anymore - all we need is PCIe to connect our peripheral devices directly to the processor. 16x videocard, x8 expansion slot, 4 x1 expansion cards, x1 2.5G Ethernet port, 3 x4 M.2 /SATA RAID, 8x USB4/Thunderbolt and USB3/USB2 ports... here go your 40 PCIe lanes offered by mainstream AMD X570 chipset (itself a scaled-down version of EPYC 'Matisse' I/O die).

DmitryKo:

I was basically saying that even though Intel obviously planned their response for Zen2 announcement well in advance, yet they probably underestimated the impact and underdelivered on the Comet Lake specs. So it would probably take them another iteration in the form of Tiger Lake silicon (Willow Cove architecture) to get closer to AMD SocketAM4+ which I suppose comes with Zen3/Zen4 and introduces DDR5 memory - and match AMD's 1350+ pin count too 🙂

I get what you're saying, but you can't use pin count as a way to determine that. I don't disagree that it'll take another iteration for them to really catch up, but pin count is not a good/reliable metric for performance.

As for PCIe lanes, everything is PCIe these days

Sure, but how often do you find yourself running out bandwidth or lanes? In a lot of cases, motherboards use PCIe multiplexers for on-board 3rd-party controllers like ethernet, audio, more USB controllers, or more SATA ports, because most of those things don't need that much bandwidth. Sometimes they'll even use multiplexers for x1 slots. If you really need all those extra lanes, a mainstream platform like socket AM4, 1151, or 1200 isn't meant for you.

Honestly we don't even need a chipset anymore - all we need is PCIe to connect our peripheral devices directly to the processor. 16x videocard, x8 expansion slot, 4 x1 expansion cards, x1 2.5G Ethernet port, 3 x4 M.2 /SATA RAID, 8x USB4/Thunderbolt and USB3/USB2 ports... here go your 40 PCIe lanes offered by mainstream AMD X570 chipset (itself a scaled-down version of EPYC 'Matisse' I/O die).

Why stop there? Why not just solder the CPU to the motherboard?

schmidtbag:

pin count is not a good/reliable metric for performance

I was talking about features, like PCIe/SATA/USB lanes direct from the processor, which allow you to only upgrade the processor but leave the system board and memory.

how often do you find yourself running out bandwidth or lanes?

Right now, I've just added an M.2 PCIe x4 adapter card with an NVMe SSD to a second x16 slot on my motherboard, but I cannot install a USB3.0 controller card because the only spare PCIe x1 slot is obstructed by 2-slot cooling system of my video card, and the rest of my slots are conventional PCI.

If you really need all those extra lanes, a mainstream platform like socket AM4, 1151, or 1200 isn't meant for you.

As I said, Socket AM4 offers up to 40 PCIe lanes with AMD X570 chipset. Isn't that a 'mainstream' platform?

Why stop there? Why not just solder the CPU to the motherboard?

Last time I checked, motherboard makers are different entities from CPU makers, so what would be their incentive to offer such packages for the traditional desktop? Of course there are embedded all-in-one desktop and headless server motherboards where the CPUs come in BGA packages and they are indeed soldered to the motherboard. The point was that AMD EPYC CPUs do not require a chipset even in multiprocessor motherboards, they have all the necessary I/O on board.

DmitryKo:

As I said, Socket AM4 offers up to 40 PCIe lanes with AMD X570 chipset. Isn't that a 'mainstream' platform?

Yes, it is. My point is if you expect more than that, you're not really fitting mainstream needs, and need to go higher-end.

Last time I checked, motherboard makers are different entities from CPU makers, so what would be their incentive to offer such packages for the traditional desktop?

The chipset is half the reason you buy a motherboard. If you have a SoC (which is basically what you're asking for), you start to complicate things for the user because you're removing the user's ability for budgeting and customization. If you buy a high-end CPU with all the bells and whistles, you kinda have to buy a high-end motherboard to socket it in and interface with those features or else you wasted your money. Meanwhile, if you buy a low-end CPU (implying a crippled chipset) and stick it in a high-end board, half the features won't work. Take that i5 for the X299 platform, for example. For PC enthusiasts, having the chipset external allows much more variation and freedom. This doesn't matter so much for servers since they don't care about aesthetics or peripherals, they just want a bunch of I/O, cores, and memory. I'm not necessarily saying this is a bad idea (ARM, MIPS, and RISC-V platforms work like this) but it's a bit of a slap in the face to PC enthusiasts since you're removing choices from people.

The point was that AMD EPYC CPUs do not require a chipset even in multiprocessor motherboards, they have all the necessary I/O on board.

Well if you also take a close look at many server motherboards, they don't have much in the way of rear-panel I/O. In a lot of Epyc boards, I'm pretty sure most of the I/O is from 3rd party controllers integrated into the motherboard, rather than fed directly from the CPU. So, Epyc doesn't need a chipset because people who buy those platforms are expecting to either use their own expansion cards, or, whatever the 3rd party controllers are that were included by the manufacturer.

schmidtbag:

if you expect more than that, you're not really fitting mainstream needs

I don't need more than that. The point was these 40 PCIe lanes are only found in high-end desktop (HEDT) Core X parts (LGA2011/2066 and LGA3647/4189, and not mainstream parts with Socket 115x/1200.

The chipset is half the reason you buy a motherboard.

The reason why I buy them is my CPUs cannot start without a motherboard. 😉 Then there are additional factors like PCIe expansion slots, USB ports, and DIMM sockets. Nowdays all these are connected directly to the CPU, at least in case of AMD Socket AM4 processors. Though personally I would gladly buy an X570 motherboard with any of 3000-series Ryzen 3/5/7 - it would still be far better that my current Intel config, and still better than anything Intel can offer in its current mainstream range.

If you have a SoC (which is basically what you're asking for), you start to complicate things for the user because you're removing the user's ability for budgeting and customization.

it's a bit of a slap in the face to PC enthusiasts since you're removing choices from people

On the contrary - even if every CPU would encompass the 'chipset' on the I/O die, so that low-end CPUs include a low-end 'chipset' and high-end CPUs include a high-end 'chipset' onboard , this should give the enthusiasts more choice by providing an unified upgrade path across the entire range - something Intel processos lost with the transition from Socket 775. It does not take away the choice of physical slots and sockets, ATX form factor, onboard Ethernet and HD audio, multi-phase voltage converter and overclocking options, fan/LED control. You could still buy a high-end motherboard with an entry-level CPU and upgrade later (the most logical choice for me), or buy a mid- to high-end CPU with a basic chipset in a small form factor, then change the motherboard for more I/O and epansion options (which is nearly impossible today). Or tie-in the CPU with respective low-end, mid-range, or high-end motherboard and still have some ways to upgrade a single part when needed.

if you also take a close look at many server motherboards, they don't have much in the way of rear-panel I/O

USB3 , USB-C, USB4/Thunderbolt3/DisplayPort2 will soon all be coming directly from the CPU. LAN controllers, WiFi M.2 cards and HD audio controllers sit on standard PCIe lanes as well. With the end of legacy BIOS and complete switchover to UEFI firmware, this reduces the 'chipset' to a LPC bus master connecting whatever custom controllers are installed on the motherboard, including SPI flash ROM, voltage/fan control, sensors, and LED lighting .

i must be reading this wrong, but wtf does Wi-Fi 802.11ax have to do with cpu? do they plan to add that stuff into the cpu? or is going into the chipset? which last i check wifi/ethernet were never part of chipset but extras on the motherboard

karma777police:

Intel is making a big mistake here. If I was them I would kill socket for 9900k and lower SKUs and bring 2066 socket to mainstream therefore people who already have that socket do not have to buy new mobo. New chipset is fine but same socket so newer CPU can work on new and older mobo. That would be the smartest thing to do. Also lower the price of 2066 socket CPUs to match Zen 2 and case closed. They don't have to worry about core count cause they already have that with 2066 socket. If they want to release an updated CPU sure, to gain some performance and have higher clock speed. Since 2066 socket already have core count, if they are any smart they would offer 18/36 core CPU for $499 to undercut Zen 2 big time. AMD cannot squeeze more than 16 cores on the current AM4 socket and Intel's 18 Cores for $499 would seriously hurt AMD. Then offer 16/14/12/10/8 cores and each SKU seriously undercut Zen 2. In nutshell Intel Strategy for near future should be More cores for less money Quad Memory Channel vs. Dual Memory Channel again for less money Cheaper 2066 socket motherboard compared to X570 Single CPU performance given clock speed won't suffer compared to Zen 2 as it already beats Zen 2 in gaming This would hurt AMD big time. In mean time work on new gen. CPU (10nm).

No offense, but that is the worst plan I have ever seen! Intel uses a monolithic design for the XCC chips, meaning yields are low - selling a huge 18-core monolithic chip for $499 would be disastrous for their margins and bottom line. This would hurt them far more than it would hurt AMD.

tsunami231:

wtf does Wi-Fi 802.11ax have to do with cpu? do they plan to add that stuff into the cpu? or is going into the chipset?

Full-featured Wi-Fi/Bluetooth controllers come as standard M.2 2230/2242 E-key card, using PCIe lanes for WiFi and USB2 for BlueTooth. They fit in the M.2 E-key slot positioned next to the rear I/O panel. These expansion cards are commonplace as they replaced mini-PCIe Wi-Fi cards for notebook computers. They come bundled with a few desktop motherboards which include antenna connectors bracket on the rear I/O panel; you can also buy them separately and mount the antenna bracket on the free PCIe expansion card slot. However some Intel Wireless AC M.2 cards (in particular 9560, 9461, 9462) only contain analog signal processing and RF curcuits, while the actual Wi-Fi controller was moved to the CPU/chipset. These cards use Intel's proptietary CNVio interface and will only work with recent Intel desktop chipsets and mobile CPUs.