What do you think (or would like to see) AMD's next step for HEDT will be?

[cbn]

The current AMD platform is X399 with the 180W 16C/32T 1950X competing against Intel's X299 and the 140W 10C/20T Core i9 7900X.

However, Intel has announced even more cores to the X299 (up to 18C/36T with the Core i9 7980XE):

So what do you think AMD will do next?

Introduce a X499 motherboard (when Pinnacle Ridge arrives) with all four dies active (ie, eight channels of memory)? Maybe also use some left over 14nm Zeppelin dies make to another four active die (ie, eight channel) processor?

What about TDP? 180W max (wiith unlocked multiplier)? Or increase to a higher level?

Validated ECC Support? Or ECC enabled by AMD with validation up to the OEM?

P.S. With a full coverage cooler (like this one rated at 500W TDP), I personally would like to see the stock TDP higher than 180W. Maybe even as high as 300W to 380W (380W = four 8C/16T dies each at the full 95W specification). This with validated ECC (but without all the Server and Security features specific to EPYC)

 

[osgorth]

I expect slightly higher clocks (10% ?), hopefully a bit less power draw and a leaner TDP (165W?). I think the core count and pricing is going to remain the same for Pinnacle Ridge. But yeah: higher clocks and less power - that's what we need.

What I want is validated ECC at higher speeds than 2133. Surely it must be possible to make at least 2666 or hopefully 3000MHz ECC DIMMs. If AMD would go all the way and include full & validated ECC support for at least the Threadripper line, that would go a long way to thoroughly annoy Intel. And it would make it a serious competitor to Xeon workstations. Right now it's not.

That said, just better memory support is what I want. I want higher clocks overall, and I want to be able to go out and buy, say, a 3600 or 4000 kit and expect it to work straight out of the box. Intel is way ahead in this area, and it annoys me greatly. They need to make sure that at least all the major brands and kits work on day 1, e.g. the Corsairs, the G.Skill's and what have you.

For Zen 2 I'm hoping for 8 cores per CCX to massively reduce crosstalk. Hey, I can dream right?

 

[crashtech]

New steppings that fix bugs and increase max clocks would seem the logical route at this time. Memory support seems to be more about AGESA revisions than memory controller issues, afaik.

 

[LightningZ71]

I expect zen+ with higher clocks as the next step. Then, for zen2, 12 cores per die for a 24c/48t TR. There is also the possibility of an in-between product for a special purpose processor between TR and Epyc that has all 4 dice active, but only one active memory channel per core. For those non-memory bound but highly threaded jobs. 48 cores, 96 threads, that's a lot of threads and l3 cache (96MB).

 

[cbn]

There is also the possibility of an in-between product for a special purpose processor between TR and Epyc that has all 4 dice active, but only one active memory channel per core.

I like that idea of a high core count quad channel for existing X399, maybe AMD 14nm Zen for that as a first step?

... I'll bet by this time AMD has a pile of Zeppelins with one broken memory controller. (re: Notice each memory controller is quite large (about the size of two Zen cores with L2 cache, see below)).

 

[cbn]

If AMD would go all the way and include full & validated ECC support for at least the Threadripper line, that would go a long way to thoroughly annoy Intel. And it would make it a serious competitor to Xeon workstations. Right now it's not.

I would really like to see full and validated ECC support as well.

P.S. Back in the LGA 1366 era.....HP, Dell and Lenvo 1P Workstations (Z400, T3500, S20) all used a consumer Intel chipset (X58). This with ECC UDIMM support. Interestingly enough some of the Xeons (eg, W3690, W3680, W3670) used in these computers also had unlocked multipliers....though the motherboards in these OEM workstations did not actually allow overclocking. So maybe AMD is not that far from getting Socket TR4 hardware in machines like the HP Z4?

 

[osgorth]

So maybe AMD is not that far from getting Socket TR4 hardware in machines like the HP Z4?

Yeah, let's hope.

My personal hope and gut feeling is that Dell brings out some sweet TR workstations. Since they're already using them in their Alienware boxes, the step from there to a Precision workstation isn't usually that big.. If only they could come up with a working ECC solution - that could be really interesting.

I think the TR is much more interesting as a Xeon competitor than anything else, seeing as Xeons can't be overclocked and are usually at lower clock speeds to begin with. A 4GHz 16-core Threadripper could certainly be a tough battle for a lot of Xeons, particularly if price is taken into consideration. The Xeon infrastructure is really way too expensive, and that's where AMD wants to hurt Intel if they want to be truly competitive.

 

[cbn]

I think the TR is much more interesting as a Xeon competitor than anything else

+1.

(They both have ECC and the ability to address a large amount of RAM).

P.S. Here is a list of the 1S LGA 2066 Workstation processors:

https://www.anandtech.com/show/11775/intel-launches-xeon-w-cpus-for-workstations

(The top 18C/36T Xeon actually is clocked faster than I expected at 4.3 Ghz turbo, though the base clock is only 2.3 Ghz)

 

[cbn]

As a comparison point, here are the 1P EPYC processors:

https://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade/3 (scroll down to the bottom of the page)

EPYC 7551P: 32C/64T Base clock 2.0 GHz, All 2.6 GHz, Max 3.00 GHz, 64MB L3, 8 channel DDR4 2666, PCIe 8 x 16 128 PCIe, 180W TDP, $2100
EPYC 7401P: 24C/48T Base clock 2.0 GHz, All 2.8 GHz, Max 3.00 GHz, 64MB L3, 8 channel DDR4 2666, PCIe 8 x 16 128 PCIe, 155W/170W TDP, $1075
EPYC 7351P: 16C/32T Base clock 2.0 GHz, All 2.4 GHz, Max 2.90 GHz, 64MB L3, 8 channel DDR4 2666, PCIe 8 x 16 128 PCIe, 155W/170W TDP, $750

 

[cbn]

Comparing the EPYC 7551P (previous post) to the Threadripper 1950X the EPYC has twice as many cores for slightly more than twice as much money.

That seems like a good deal with the exception being the clocks are so low.

Maybe for a future HEDT platform (with four active dies and eight channels) AMD could increase TDP/clocks and/or add an unlocked multiplier, but strip out the server management and security features as well as the validated ECC ? (This so it won't compete directly against a 2P EPYC)

 

[cbn]

I expect zen+ with higher clocks as the next step. Then, for zen2, 12 cores per die for a 24c/48t TR. There is also the possibility of an in-between product for a special purpose processor between TR and Epyc that has all 4 dice active, but only one active memory channel per core. For those non-memory bound but highly threaded jobs. 48 cores, 96 threads, that's a lot of threads and l3 cache (96MB).

Maybe to get around the bandwidth issue for a very high core count (but only quad memory channel) processor* AMD could create a technology similar to Intel's "Memory Drive". This where multiple NVMe M.2 SSDs in RAID-0 could function as part of the memory pool and thus also add bandwidth.

http://www.zdnet.com/article/optane-ssd-fast-enough-to-be-used-as-memory-extender-intel/

https://www.intel.com/content/www/us/en/solid-state-drives/optane-ssd-dc-p4800x-mdt-brief.html

Intel® Memory Drive Technology is a revolutionary software that extends system memory transparently. Combined with an Intel® Optane™ Solid State Drive (SSD), Intel® Memory Drive Technology transparently integrates the SSD into the memory subsystem and makes it appear like DRAM to the OS and applications. Intel® Memory Drive Technology increases memory capacity beyond DRAM limitations and delivers DRAM-like performance in a completely transparent manner to the operating system and application. In addition, no changes are required to the OS or applications.

There are two key scenarios in which it is beneficial for an IT environment to use Intel® Memory Drive Technology:

• Displace a portion of DRAM to reduce overall memory cost
• Grow the memory pool beyond DRAM capacities when large system memory is required.

https://www.pcworld.com/article/322...adripper-cpus-get-free-nvme-raid-support.html

Of course, the endurance of NAND is much less than Optane, but if only a small part each NVMe SSD is used for the memory pool I think the SSD's wear leveling could make it work for consumer/workstation environment. This especially if the System RAM is not exceeded for capacity or bandwidth reasons very often.

*Or perhaps eventually a large Workstation class APU (16C, large iGPU with 1/2 rate double precision floating point, quad channel memory, HBM2 (with ECC) stacks). This especially with multiple PCIe 4.0 or PCIe 5.0 NVMe M.2 SSDs in RAID-0 (see table below).

https://www.nextplatform.com/2017/07/14/system-bottleneck-shifts-pci-express/

P.S. PCIe x 16 cards that work with Threadripper's NVMe RAID are very affordable (Eg, ASUS Hyper M.2 x16). However, at one time this kind of tech was expensive, not bootable and exclusive to certain machines via BIOS lock. (Eg, HP Z Turbo Drive Quad Pro)

 

[NostaSeronx]

The Bulldozer cores will arise again!



///

So what do you think AMD will do next?

Release a 16-core AMD Excavator on a planar node under some new packaging name. Same platform, single die, lower cost... Sempron of HEDT. (*cough* Revisioned 40h-4Fh *cough*)

 

[raghu78]

I think AMD's next two generations of TR are pretty straightforward.

1. Zen on 12LP with higher clocks.
2. Zen 2 on GF 7LP with 24C/48T TR

I think AMD needs to make some IMC performance tweaks to reduce memory latency and support for DDR4 4000+ speeds with Pinnacle Ridge.

 

[TahoeDust]

What they need to do is get those clocks up. Even if they equal IPC, until they can compete in frequency, I won't be interested.