Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

hothardware.com

Post reply on HN
Print view View original

81–90 of 97 posts

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#81

Post by sireat »

paulmd wrote: post #77
valarauca1 wrote: post #56

Earlier quoted context omitted.

>server segment where MBs are significantly more expensive False-ish Single socket server boards are on par with home motherboards. They only become expensive when you start wanting 2+ NIC's and A LOT of sockets. But then yes your feature thing is mostly true. > and lack a whole slew of additional features nice for desktop machines. False-ish Intel's server socket (for non E7's) is 2011v3 which is also their enthusia…

You don't need to go to a server board for ECC - a lot of the X99 boards actually do support it. Not all of them (EVGA doesn't, a lot of Gigabyte boards don't) but a lot do (Asrock usually does). And you can find motherboards in pretty much any price range except the barest of dirt cheap. There are motherboards like the GA-X99-SLI that regularly hit $125 or lower. Cheapest thing on Newegg right now is a $150 Asrock b…

Problem is that you cannot have both via current (from 2011+ on) Intel route: high clock speed on the CPU(achieved by OC) AND ECC.

You want ECC you have to get Xeons.

You want OC you have to get regular mainstream K chips or enthusiast -E chips without ECC support.

I think the last chipset which let you OC Xeons and still use ECC was X58 on 1366 socket.

Disclaimer: I am not counting tiny BCLK OC available on Xeons.

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#82

Post by Twirrim »

Sanddancer wrote: post #17
m-j-fox wrote: post #13

I wonder why it needs so many pins? It only has 4 DDR channels, which is going to be a bottleneck keeping those cores fed. 8 channels would be more appropriate for 16 cores and it would explain why so many pins.

It needs a lot of pins because it's an I/O monster. The socket is going to support their high end server chips with 8 memory channels. Plus, all of their high end line supports 128 pcie lanes, and will have on board an unknown number of usb, sata, and networking lanes. All that adds up to needing a very large socket.

That article points out it's a different version of the core on the desktop, SP3r2 instead of the SP3 that is landing server side. SP3r2 is only coming with 4 DDR channels.

It's not clear why they're deliberately using a crippled socket here for the desktop. I don't understand the value proposition, and can only envision it driving up motherboard prices. That CPU combined with a crippled socket is surely likely to end up starved of memory IO

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#83

Post by paulmd »

sireat wrote: post #81
paulmd wrote: post #77

Earlier quoted context omitted.

You don't need to go to a server board for ECC - a lot of the X99 boards actually do support it. Not all of them (EVGA doesn't, a lot of Gigabyte boards don't) but a lot do (Asrock usually does). And you can find motherboards in pretty much any price range except the barest of dirt cheap. There are motherboards like the GA-X99-SLI that regularly hit $125 or lower. Cheapest thing on Newegg right now is a $150 Asrock b…

Problem is that you cannot have both via current (from 2011+ on) Intel route: high clock speed on the CPU(achieved by OC) AND ECC. You want ECC you have to get Xeons. You want OC you have to get regular mainstream K chips or enthusiast -E chips without ECC support. I think the last chipset which let you OC Xeons and still use ECC was X58 on 1366 socket. Disclaimer: I am not counting tiny BCLK OC available on Xeons.

Definitely true.

Tangent: a lot of people wonder who would pay $1700 for a 6950X. The answer I always give is "people want to do intensive work and gaming on the same system". Otherwise the 4/6C chips are better choices for the money.

The great weakness of that system is - as you note - the lack of ECC.

I strongly, strongly recommend building two different systems for this situation. Depending on your needs it can almost be more economical to do it this way. i3s (and now Pentiums) support ECC and make good ZFS servers. The 4C Xeons aren't too bad either. You can also do some cheap E5 Xeon builds using engineering samples. Many-core Xeons are more than sufficient for the kind of stuff you need ECC for, even at relatively low clock rates. A 10C at 2.4 GHz is roughly comparable to 6C at 4.13 GHz (I use mine in Handbrake). And uses 2/3 of the power to boot.

And then you can focus on gaming performance in your gaming system. Like 3x as much for a given level of performance.

It's a lot like laptops. You're going to spend a lot more if you insist on only owning one system that can do everything.

The segmentation here does suck, I would rather see it supported even on consumer hardware, but practically speaking I don't think Intel is going to change this any time soon. Even if AMD does.

(and note: unless the manufacturer is willing to stand behind ECC, Ryzens's "unofficial" ECC support holds about as much weight as using Intel's engineering samples)

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#84

Post by srcmap »

nanistheonlyist wrote: post #20

I break out in cold sweat when I think about routing a 4094 pin socket. How does anyone layout a motherboard for this in a reasonable time frame?

First, you need a computer with a lot of Cores. :-)

All the silicon vendors give out design/layout guideline and reference designs. Most of Motherboard vendor likely copy those with just minor changes.

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#85

Post by NaliSauce »

hmottestad wrote: post #65
simcop2387 wrote: post #52

Well damn this looks like it might be in the right spot for what I'm after. It'll depend on price-point to be exact but I've got a semi-server semi-gaming setup that I want to upgrade from an older i5 and have been waiting to see Naples and early adopter stuff get hammered out. 16 cores is a great spot for me to give a few to a gaming VM, while still being able to have plenty left over for running build/smoke VMs and…

For Intel's Xeon processors, the more cores you have the lower the frequency you get. Quad core might give you an extra gigahertz, which might be of more value to games than 12+ cores.

True for old games, but for them the single core performance is enough (except maybe supreme commander with 5'000 units on the field). Looking forward with DX12 and Vulcan more cores seems to be more future proof.

Plus shouldn't XFR (or Turbo Boost Max Technology 3.0 Frequency for Intel) somewhat alleviate that particular issue?

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#86

Post by valarauca1 »

sireat wrote: post #81
paulmd wrote: post #77

Earlier quoted context omitted.

You don't need to go to a server board for ECC - a lot of the X99 boards actually do support it. Not all of them (EVGA doesn't, a lot of Gigabyte boards don't) but a lot do (Asrock usually does). And you can find motherboards in pretty much any price range except the barest of dirt cheap. There are motherboards like the GA-X99-SLI that regularly hit $125 or lower. Cheapest thing on Newegg right now is a $150 Asrock b…

Problem is that you cannot have both via current (from 2011+ on) Intel route: high clock speed on the CPU(achieved by OC) AND ECC. You want ECC you have to get Xeons. You want OC you have to get regular mainstream K chips or enthusiast -E chips without ECC support. I think the last chipset which let you OC Xeons and still use ECC was X58 on 1366 socket. Disclaimer: I am not counting tiny BCLK OC available on Xeons.

Because ECC and OC are targeted at different audiences. If you want both... you don't understand one or the other.

ECC is to guarantee system stability. Solving rare event that become regular once you have hundreds of boxes running.

OC INCREASES system instability. You put more power into a chip, increasing the probability of incorrect or malformed answers.

Now both these scenarios are rare, especially for single box home enthusiasts. Generally if you have a stable OC you won't hit garbled instruction results. Likewise you may almost never have an error that ECC would present.

---

Having OC and ECC basically says, "I want to solve one extremely rare error while opening myself up to another!" This is an idiotically unsound decision.

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#87

Post by solotronics »

SmellTheGlove wrote: post #40
tscs37 wrote: post #36

Earlier quoted context omitted.

What about a socket that is like USB-C, there is no wrong way to insert it. Though I imagine such a socket would be an insane task to complete and have numerous problems with latency and speed.

Engineering a socket like that has to be a low priority. It's not as though people are reseating their CPUs every day, and generally speaking, the people that do it can handle aligning it properly.

I helped a friend build a computer back in school and when he got the parts in the mail he hurriedly started building (against my advice to wait until I got there) he put the CPU in the wrong way and bent some of the pins. So it does happen more than you would think.

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#88

Post by mschuster91 »

valarauca1 wrote: post #86
sireat wrote: post #81

Earlier quoted context omitted.

Problem is that you cannot have both via current (from 2011+ on) Intel route: high clock speed on the CPU(achieved by OC) AND ECC. You want ECC you have to get Xeons. You want OC you have to get regular mainstream K chips or enthusiast -E chips without ECC support. I think the last chipset which let you OC Xeons and still use ECC was X58 on 1366 socket. Disclaimer: I am not counting tiny BCLK OC available on Xeons.

Because ECC and OC are targeted at different audiences. If you want both... you don't understand one or the other. ECC is to guarantee system stability. Solving rare event that become regular once you have hundreds of boxes running. OC INCREASES system instability. You put more power into a chip, increasing the probability of incorrect or malformed answers. Now both these scenarios are rare, especially for single box…

You can (and should!) decrease error ratios while OCing by providing proper cooling.

Having ECC memory in an OC setup will increase its stability further - and there are perfectly valid reasons for overclocking, for example if you want more computing power than even a 22-core Xeon does not provide.

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#89

Post by astrodust »

BuckRogers wrote: post #50
bitL wrote: post #7

Can't wait! Postponed my Ryzen purchase as rumors started to surface about June launch of 16c version - this could allow running 4x GTX 1080 for ML and provide significant boost for movie rendering.

I don't blame you, I was going to hold off, then realized since I'm all mITX from here on out that I'd just build now. None of this stuff is going to be 65W TDP (at least the 8-core or more parts) so the 1700 will continue to make a lot of sense for those moving to mini-ITX.

These are going to be 150W+ for sure, heat be damned when you need performance.

Re: AMD 16-Core ThreadRipper Enthusiast CPUs to Reportedly Utilize 4094 Pin Socket

#90

Post by astrodust »

cdawzrd wrote: post #58
skummetmaelk wrote: post #41

Earlier quoted context omitted.

Not really, autorouting is generally extremely bad. Especially for applications with such high performance criteria. On the plus side the work only has to be done once since later motherboards are mostly just iterations on a previous design.

The latest high-end PCB tools actually provide useful autorouting features, because they have figured out that if you combine human ability to solve the difficult pattern-recognition and planning problems, you can have a computer solve the details. For example, Xpedition's sketch routing feature: https://www.youtube.com/watch?v=fIi13gI9xEA Routing 4094 signals sounds daunting, but it is a bit less daunting when you r…

Yeah, a good portion of those pins will be wired directly to the power regulator components that are next to the socket, and a good chunk more break out neatly into channels for memory and PCI-e.

Seems like it's almost boring these days since everything's encapsulated in a layer of abstraction. A lot of the more complicated wiring is in breaking out PCI-e channels into things like ethernet, audio, and other miscellaneous ports that involve good chunks of analog circuitry.

Post reply on HN