AMD Ryzen Threadripper 2990WX

[Biscuit]

I'm really, really, turned off by game mode. But on the workstation front, I'd definitely opt for one of these over an Intel for my work computer which does no gaming at all.

Depends on perspective, I guess - AMD has launched it into an HEDT platform, and from a platform point of view it's being tested appropriately. I thought Hexus did a reasonable job of pointing out that this was an outstanding workstation chip, but if you were looking for an HEDT platform for gaming you should look elsewhere.

Looking around at some of the benchmarks in other peoples testing, specifically the linux testing, it does seem like there are significant room for optomisation.

Perhaps the "gaming mode" is more a dirty work around for AMD to get around Microsoft's issues with scheduling than it is for them to get around their own shortfall.

I mean look at the state of this chart from Phoronix:


Thats an outragrous difference between the results using the same software on a different OS.

Edit: For comparison, here are the results when using a 7980XE:

They had an MCM in the consumer market in 2006. All Core 2 Quads were an MCM of 2 Core 2 Duo dies. It worked very well for them, but makes their current line on Threadripper a bit hypocritical...

Very good point, but then they just ran everything over the FSB with the memory controller located outside the CPU dies. I guess the process of transitioning a Core ix series architecture into an MCM would be a significantly more difficult challenge.

[DanceswithUnix]

Sorry to be pedantic, but there are 4 x die on an interposer, each die has 2 x 4 core CCX.

Is it really on an interposer? The die are intended to bond onto a conventional carrier, there isn't any high density HBM connections, so AIUI these things use an MCM carrier.

Also,looking at this Win10 against Linux comparison,it seems Linux works better with the 2990WX:

The likes of IBM have been doing high core count (and 4 thread per core at that) Power based Linux boxes for a long time, so some of the scaling issues will be long since addressed.

Edit: Some of those benchmarks are a sound thrashing though, notably the zip test where I presume Windows' poor small file performance is killing it. Was surprised to see Blender going significantly faster though, and amuse to see Intel's Linux distro (usually one of the fastest) trailing.

[scaryjim]

… I guess the process of transitioning a Core ix series architecture into an MCM would be a significantly more difficult challenge.

Well, they can do 2P systems easily enough, so they've got the links in place to connect at least 2 dies in an MCM, and they're currently using MCM packaging technology for the Kaby G, so there's absolutely no good reason they couldn't just slap two dies on a substrate and call it quits. But given how much shade they've tried to throw about AMD using MCMs for their high core count servers, my guess is they won't do that.

[DanceswithUnix]

Even better, can this be fixed by setting processor affinity for games? If so, then no reboot required.

I don't see why you couldn't park cores in the Windows scheduler to get the same sort of effect as gaming mode, but that would be a Windows change not a BIOS hack so probably not an option for AMD to do and probably not something Microsoft would be interested in maintaining.

[Biscuit]

Is it really on an interposer? The die are intended to bond onto a conventional carrier, there isn't any high density HBM connections, so AIUI these things use an MCM carrier.

Perhaps my understanding of terminology is wrong then, but I thought the interconnections between the die were on the "carrier" which would make it a passive interposer.

I mean at its basic level, this is an interposer (according to wiki):

[DanceswithUnix]

I mean at its basic level, this is an interposer (according to wiki):

I find that picture quite odd, down to the 1980's gold bonding wires rather than flip chip bumps.

AMD usually talk about an interposer for things like the Vega chips where there is a passive silicon layer between the organic substrate of the carrier and the die that contain the logic.

So TR is the top MCM setup in this picture from Intel, wiring within the substrate, well established technology so comparatively cheap:

[Biscuit]

I find that picture quite odd, down to the 1980's gold bonding wires rather than flip chip bumps.

AMD usually talk about an interposer for things like the Vega chips where there is a passive silicon layer between the organic substrate of the carrier and the die that contain the logic.

So TR is the top MCM setup in this picture from Intel, wiring within the substrate, well established technology so comparatively cheap:

Fair enough, I guess there is a technology/terminology thing I have got a bit confused with... I'll just stick with "substrate"

[devBunny]

… It feels like the part was reviewed for usage it isn't designed for ...

. . .

The monolithic core/thread counts are now so high that I don't see many games ever swamping them. No doubt you could write a game that would hammer 16 threads, but I doubt it would be an efficient way to code - many tasks just don't parallelise that well...

I'd just like to point out that, not only do sites like this put a lot of emphasis on gaming with technology that isn't for gamers, "gaming" is limited to video gamers. A 32-core computer would be the dream of someone playing games in which the core component is the neural net rather than the frame buffer; games such as Backgammon and Chess.

It wouldn't surprise me if some of these chips end up in the hands of World Class and other top-level players, maybe even strong club players with deep pockets. In Backgammon, the "rollout" of a particular play (a Monte Carlo sampling of the paths from the given position to the end of the game) can take anything from minutes to hours. For me, 64 faster threads would be a huge improvement on the 12 threads that my 3930K provides.

I do admit, however, that this segment is about as extreme as "niche market" gets. :-D

[CAT-THE-FIFTH]

LOL:

https://www.google.co.uk/amp/s/wccft...cpu-hedt-/amp/

So Intel is releasing another socket to compete with Threadripper.

[Biscuit]

LOL:

https://www.google.co.uk/amp/s/wccft...cpu-hedt-/amp/

So Intel is releasing another socket to compete with Threadripper.

It's like they're deliberately being completely ignorant to the complaints of the market. Either that or completely out of options. Perhaps both.

Does make me laugh that they stuck with X*99 but just jumped to 599... like... what are they both hoping to achieve with these numbering systems?!

such a bizarre petty naming battle

[Dashers]

As a HEDT consumer, these sort of cores and platforms are way out of my league it's not even worth reading about.

Entry level thread ripper and i9 is where it's interesting.

[Muaadib]

It's like they're deliberately being completely ignorant to the complaints of the market. Either that or completely out of options. Perhaps both.

Does make me laugh that they stuck with X*99 but just jumped to 599... like... what are they both hoping to achieve with these numbering systems?!

such a bizarre petty naming battle

Amd did it to steal Intel's brand recognition (Bx50,Xx99 and to some extent Xx70 for Zx70). Now they are just leapfrogging each other.

[CAT-THE-FIFTH]

It looks like the NV driver might be the reason gaming performance seems a bit meh with the 2990WX:

https://www.golem.de/news/32-kern-cp...08-136016.html
https://www.hardocp.com/news/2018/08...nvidia_driver/



Loads of other games are also affected too it appears.

[Tabbykatze]

It looks like the NV driver might be the reason gaming performance seems a bit meh with the 2990WX:

https://www.golem.de/news/32-kern-cp...08-136016.html
https://www.hardocp.com/news/2018/08...nvidia_driver/



Loads of other games are also affected too it appears.

Would love to see that qualified by other reviewers. Would be a poor show for Nvidia if that was the case.

I doubt they did anything actively for this to happen but maybe >32 thread CPUs weren't in their driver considerations.

[DanceswithUnix]

I doubt they did anything actively for this to happen but maybe >32 thread CPUs weren't in their driver considerations.

Maybe on the OpenGL Quadro card driver, but I wouldn't expect the DirectX driver to be tested let alone optimised for lots of threads.

If this is true, Nvidia won't want the message that you want a Vega 64 in your high end desktop box so I expect this will get fixed pretty damned fast

[Tabbykatze]

Maybe on the OpenGL Quadro card driver, but I wouldn't expect the DirectX driver to be tested let alone optimised for lots of threads.

If this is true, Nvidia won't want the message that you want a Vega 64 in your high end desktop box so I expect this will get fixed pretty damned fast

Abso-frickin-lutely. That also might make sense why the Linux benches were so much faster which could be down to the nvidia linux drivers being far better at working with large numbers of threads!