AMD Showed Off New Threadrippers, 7nm Vega At Computex 2018

[audir8]

The cost for the 32C version is allegedly going to be $1499.

Whilst the socket is octo-channel capable (at least in its server variant), the motherboards aren't, and neither is the CPU itself wired up that way. You can see the current wiring schematic for ThreadRippers online somewhere (shows all the memory channels and InfinityFabric connections between dies within the package substrate).

I think the best we can hope for is faster memory support on the four channels, unless there is an X499 chipset/platform that enables 8 channel memory for TR2.

For $1500, QC support on a 32C/64T processor seems in line with every other AMD offering. It's a trade-off given what's on a CCX and how many can be paired together. TR4 owners should be happy they'll have something worth upgrading to in terms of doubling the highest amount of cores for around $500.

I can't explicitly find if the Vega Nano has 4/8/16GB HBM2?, certainly seems compelling if it does.

[xiando]

Hmmm, wonder how much that Intel 28Core 4-5ghz CPU was. I don't think its realistic to run all cores at 5ghz, must be some significant tdp and heat drawbacks.

Intel demonstrated their 28 core 5 GHz chip with a 1700W mini-fridge sized liquid cooler. That's not exactly your average included "box cooler".

[madscientist159]

Sounds like AMD have a very promising pipeline and it's all underpinned by open source software. If they can execute as planned I can see then starting to get some big wins.

Wow, really? So I can finally download the PSP source and install my own modified version? Or the UEFI interface? Or the power management controller firmware? Or the pre-RAM initialization code? Or the glue between RAM training and UEFI handoff? These are all open source now?

Oh, oops, my bad. I was thinking of the /actually/ open POWER9 systems....

[torsionbar28]

Wow, really? So I can finally download the PSP source and install my own modified version? Or the UEFI interface? Or the power management controller firmware? Or the pre-RAM initialization code? Or the glue between RAM training and UEFI handoff? These are all open source now?

Oh, oops, my bad. I was thinking of the /actually/ open POWER9 systems....

Which POWER9 system do you have? AFAIK the only one available at a consumer price point is the single-socket board from Raptor, which isn't even shipping yet.

[madscientist159]

Which POWER9 system do you have? AFAIK the only one available at a consumer price point is the single-socket board from Raptor, which isn't even shipping yet.

I work at Raptor, so just had the chance to try one of the first Lite boards yesterday. Couldn't resist the chance to take a jab at AMD on their closed, signed, vendor-managed firmware stack.

We've been busy getting the Lite systems up for sale, expect an announcement this week...

If you're interested, we've also been creating and releasing new tools such as overclocking utilities*, all open source. The P9 will run at 4.2GHz easily, without really using much power under light loading. This is all stuff you can't do without explicit vendor permission on modern x86 or higher end ARM, and shows the power of open source!

* Overclocking here is more advanced than the typical "apply flat clock and voltage offset and see what happens". Our tools literally modify the voltage and frequency maps for various load conditions on the processor, and are useful for increasing performance/watt for a particular workload. Added up over racks of servers....you get the idea. https://wiki.raptorcs.com/wiki/POWER9/Overclocking has the details, and we're seeing if we can get some benchmarks up here soon too.

[L_A_G]

I don't think price drops work like that e.g. currently the Ryzen 2600 sells for a bit more than the Ryzen 1600, which is fairly natural considering it's both newer and much better.

You do know they dropped the prices on the original Ryzen parts prior to the launch of the 2000-series to help clear inventory? The 1600 started at an MSRP* of about $220 in the U.S while the 2600 has it's U.S MSRP set to $200.

(*MSRP = Manufacturer Suggested Retail Price)

... and are only moderately more expensive than socket TR4 boards. Downside is that they miss some consumer features like overclocking, audio jacks, and LED headers.

Apart from being about twice the price of the Asus X399 Prime used in the workstation offered by my current employer there's also the fact that the CPU and RAM take up so much space those boards can't fit full size PCIe cards, which make them a complete non-starter for our use cases (as we also do some heavy GPU compute).

[oooverclocker]

<3 AMD

By the way, according to Gamers Nexus the 28-Core with 5 GHz Intel were presenting one day before AMDs 32-Core announcement was just a Skylake-X cooled by a chiller.

Really funny - I can't believe that Intel's marketing people think it's a good idea to fool tech press. AMD ran their 32-Core processor with an air cooler to show a setup that people will likely have at home.

[Adarion]

<3 AMD

By the way, according to Gamers Nexus the 28-Core with 5 GHz Intel were presenting one day before AMDs 32-Core announcement was just a Skylake-X cooled by a chiller.

Really funny - I can't believe that Intel's marketing people think it's a good idea to fool tech press. AMD ran their 32-Core processor with an air cooler to show a setup that people will likely have at home.

Indeed. It went round these days that intel's show and the benchmark were to be taken with a full spoon of salt. Reaching 5 GHz and so on, but hardly on all cores and no longer than a few seconds. They had some really big hoses installed at the backside of the case, that makes you wonder what liquid cooling they were using for the show.
It may be fancy for a show effect, but hardly usable in daily life. Some people also estimated power draws between 400 and 700 W at that speed. Ouch.

It'll be interesting what clocks AMD will have when it enters the markets. But 32 fast cores screams for programs that can make use of it, and even occasional Gentoo compiling won't really fit that monster.

[Tomin]

Indeed. It went round these days that intel's show and the benchmark were to be taken with a full spoon of salt. Reaching 5 GHz and so on, but hardly on all cores and no longer than a few seconds. They had some really big hoses installed at the backside of the case, that makes you wonder what liquid cooling they were using for the show.
It may be fancy for a show effect, but hardly usable in daily life. Some people also estimated power draws between 400 and 700 W at that speed. Ouch.

Anandtech had some pictures. 29 phase power supply for the CPU and 1770 W capable cooling sounds like it could be much more than what those estimations are. I have no idea how long they could run those things on 5 GHz but at least they got a picture of all cores* on around 5 GHz.

https://www.anandtech.com/show/12907/we-got-a-sneak-peak-on-intels-28core-all-you-need-to-know

We confirmed that Intel was using a water chiller in the 5 GHz demo, a Hailea HC-1000B, which is a 1 HP water chiller good for 1500-4000 liters per hour and uses the R124 refrigerant to reduce the temperature of the water to 4 degrees Celsius. Technically this unit has a cooling power of 1770W, which correlates to the fact that a Corsair AX1600i power unit was being used for the system.

*Well, technically that picture just shows some of the cores, so hard to say if the rest were run on such high speeds.

[Niarbeht]

All dies have dram controllers, we do not know if they will have different wiring to utilize 8 channel memory on existing TR2 platform or if they will only Use 2 memory controllers.
I do not know how they could solve it but I've seen nifty pin rearrangement at cpu package level from AMD previously.

https://www.anandtech.com/show/12906/amd-reveals-threadripper-2-up-to-32-cores-250w-x399-refresh

Two of the dies aren't actually directly hooked up to anything except for the other two dies. So, two dies are great for memory-intensive workloads, and the other two dies are great for things that can just spin in place in cache or don't need to really run all that quickly I guess.

I kinda wonder if die-independent clocking, and die-dependent scheduling, might be a possibility to improve performance and power use.

Aside from that, I wonder if it's possible to configure a 32-core EPYC chip in such a way as to "simulate" a 32-core Threadripper chip by selectively not using certain memory channels or whatever.