HEXUS Forums :: 16 Comments

Login with Forum Account

Don't have an account? Register today!
Posted by philehidiot - Tue 06 Aug 2019 15:34
“Standard socket” - does that mean they're going to stop changing a single pin here and there to make us all buy new mobos? Or does it just mean the socket isn't stupidly big?
Posted by DanceswithUnix - Tue 06 Aug 2019 15:41
Did anyone else read “bloat16 support”? In the era of custom inference & training silicon this AI feature really seems beyond too little too late.
Posted by Tabbykatze - Tue 06 Aug 2019 18:01
philehidiot
“Standard socket” - does that mean they're going to stop changing a single pin here and there to make us all buy new mobos? Or does it just mean the socket isn't stupidly big?

Nope, new socket.

You can see what they're doing, they want people to buy the cooper lake so they can upgrade to ice lake. But the mobile parts really aren't that promising on 10nm at low power, will it scale high power?
Posted by Output - Tue 06 Aug 2019 18:45
philehidiot
“Standard socket” - does that mean they're going to stop changing a single pin here and there to make us all buy new mobos? Or does it just mean the socket isn't stupidly big?

I'd imagine that's their way of saying the consumer-focused socket, rather than server ones.

There's no way they're going to give up on their insistence of making a new socket as often as they can.

EDIT: Scratch that. I've now actually read the article it was mentioned in. I guess it means that the server board you put it in doesn't have to have any particular bells and whistles that may not be available on others.
Posted by philehidiot - Tue 06 Aug 2019 18:46
Tabbykatze
Nope, new socket.

You can see what they're doing, they want people to buy the cooper lake so they can upgrade to ice lake. But the mobile parts really aren't that promising on 10nm at low power, will it scale high power?

I'll be honest, my question was hopeful but mostly just purely sarcastic.
Posted by Tabbykatze - Tue 06 Aug 2019 18:57
philehidiot
I'll be honest, my question was hopeful but mostly just purely sarcastic.

Your powers of text based sarcasm continue to serve you well!
Posted by chj - Tue 06 Aug 2019 21:10
Tabbykatze
philehidiot
I'll be honest, my question was hopeful but mostly just purely sarcastic.

Your powers of text based sarcasm continue to serve you well!

As shown on the AMD/Samsung galaxy article :P
Posted by 3dcandy - Wed 07 Aug 2019 08:22
dubbed LGA4189….

Wait 4189. Can't wait to see the boards now second hand - spares or repairs, can be fixed easily, only just over a thousand slightly bent pins…
Posted by scaryjim - Wed 07 Aug 2019 08:33
Output
… EDIT: Scratch that. I've now actually read the article it was mentioned in. I guess it means that the server board you put it in doesn't have to have any particular bells and whistles that may not be available on others.

My reading is more that current AI-focussed stuff tends to be add-in boards, while Intel are baking that into a CPU that you drop in a socket. Hardly earth shattering. I'd also assume this is an MCM of the existing 28-core Xeon scalable silicon: apparently the whole “glued together is bad” position has gone out the window now we all know that AMD will be releasing 64-core server CPUs this year… :O_o1:
Posted by Tabbykatze - Wed 07 Aug 2019 09:39
scaryjim
My reading is more that current AI-focussed stuff tends to be add-in boards, while Intel are baking that into a CPU that you drop in a socket. Hardly earth shattering. I'd also assume this is an MCM of the existing 28-core Xeon scalable silicon: apparently the whole “glued together is bad” position has gone out the window now we all know that AMD will be releasing 64-core server CPUs this year… :O_o1:

Intel are quietly eating their words in a corner hoping no one will notice…
Posted by philehidiot - Wed 07 Aug 2019 11:26
Tabbykatze
Intel are quietly eating their words in a corner hoping no one will notice…

I'd personally say that Intel have been furiously stuffing anything related to their previous comments into their prison wallet and that they are boisterously shouting about how amazing and innocent they are to anyone who will listen whilst a bunch of people who aren't blind or stupid stand around watching unimpressed.

EDIT: Reminds me of reality cop shows on telly where everyone has seen him selling meth but he's “hidden” it, thinking everyone else is stupid and that if he denies it loudly enough people will believe him.
Posted by Tabbykatze - Wed 07 Aug 2019 13:58
Oops, Intel hit by another vulnerability. So this 56 core is now a non hyper threaded SKU for security purposes! XD

SWAPGS if you do news.
Posted by DanceswithUnix - Wed 07 Aug 2019 15:16
Tabbykatze
Oops, Intel hit by another vulnerability. So this 56 core is now a non hyper threaded SKU for security purposes! XD

SWAPGS if you do news.

https://arstechnica.com/information-technology/2019/08/silent-windows-update-patched-side-channel-that-leaked-data-from-intel-cpus/

AMD are coming out surprisingly well out of these things.
Posted by Tabbykatze - Wed 07 Aug 2019 15:29
DanceswithUnix
https://arstechnica.com/information-technology/2019/08/silent-windows-update-patched-side-channel-that-leaked-data-from-intel-cpus/

AMD are coming out surprisingly well out of these things.

It is very surprising, but not without merit to their design decisions and quality of engineers.
Posted by Output - Wed 07 Aug 2019 18:46
scaryjim
My reading is more that current AI-focussed stuff tends to be add-in boards, while Intel are baking that into a CPU that you drop in a socket.

That does make a lot more sense than what I was thinking.
Posted by DanceswithUnix - Wed 07 Aug 2019 20:48
scaryjim
My reading is more that current AI-focussed stuff tends to be add-in boards, while Intel are baking that into a CPU that you drop in a socket.

Intel have added 16 bit floating point support to their AVX 512 instructions.

That allows 32 float values to be processed in a single instruction. That's nice, but I can't imagine it will keep up with Vega let alone Turing and specially a dedicated card. It might help with control programs for packing data up for Intel's own AI card which seems to support the same fp16 format.