How are they defining "workstation class"? You can get SMT-4 18-core and 22-core POWER9 workstations with 72 and 88 hardware threads respectively in a single socket.
Probably based on manufacturer marketing and IBM only markets POWER9 as server class CPUs. You can always find a manufacturer that just pulls a part from a higher target class and sells it as the top end lower class.
In general I prefer cost range based comparisons (or first for any cost) for this reason.
Maybe as in "You can actually get off the shelf software for the thing". Power9 may be great for a very few things in particular, but for actual workstation use in an everyday environment x64 wins hands down.
It would help if you could cite any workstation software, rather than just generic desktop applications. Aside from developing software to run on POWER servers, what workstation tasks is a TALOS II good for?
"Workstation" is a pretty vague term that has shifted over the years, but anyone who's using it as a distinct category from mainstream desktop computers will have at least something in mind that a workstation would be better at than a regular desktop computer. Web browsing, simple office apps and games definitely don't put you in workstation territory. Krita can be used in a workstation/professional context, but I don't think the choice of CPU is relevant there—GPU, monitor and digitizer are what a professional user of Krita is more likely to care about.
With the shift from desktops to phones as personal computing devices, I've seen the beginning of usage of "workstation" to call any box-type computing device used for work.
"Workstation software" to me means anything hardware intensive that you use to build things. Compiling large applications, video editing, music production, 3d modelling/CAD etc
Talking about a CPU like this to anyone three years ago and you would be called a madman.
But now it’s actually here. Kudos to AMD for proving how Intel has held back the industry for years and years.
Now though, what’s next? AMD has shocked (disrupted?) the CPU scene with a completely unthinkable increase in computing power at several affordable pricing levels. Everyone can upgrade their rigs to something immensely better, no matter the budget.
Current cpu chiplets are internally split in two groups, each group has its own L3.
We know that Zen3 does away with that split. That's potentially double the L3 for single cpu, which helps some loads. Additionally, better process might mean they can also make the L3 cache bigger.
AMD hasn't talked much about Zen3 because it's a ways off (H2, potentially Q4), but from the information they've given out, we know that IPC increase is over 15%.
This will widen the gap with Intel, and it's assumed that Intel's negligible advantage in some videogames, to do with memory access latency, will be gone by then.
Perhaps the craziest part of these graphs is the 3970X, this generation's second best, versus the 2990WX, last generations very best Threadripper processor. Both have 32 cores/64 threads, but the 3970X has 2x the multithreaded performance, literally 100% faster under many workloads. Its base clock is 25% faster but there's still an enormous microarchitectural improvement. Part of this is moving away from the asymmetric NUMA topology of 2990WX. What a feat.
It's automatic overclocking, implemented by extending the normal power management and turbo mechanisms to work with power draw limits the user can set beyond the normal specs.
Sure, these are great CPUs (I have a 3950x myself), but there is something needs to be mentioned: AMD has artificially restricted the memory limit. They have a max configuration of 256GB. I get that they want to segment these from Epyc, but it already only has four memory channels instead of eight (not a lot with 64cores/ 128 threads). Nor is 256GB a lot for such a CPU (the Ryzen 3000's support 128GB on two channels), previous Threadrippers supported 2TB!
I think AMD went too far here with their artificial limit, so it is worth mentioning.
(I'm not really a fan of TRX40, I think it is overpriced for what you get and they don't even throw in 10gbe when they start at £400+)
Looks like results of a GPU company being in bed with a CPU company. Wonder how much this all stems from the AMD's acquisition in ATI in 2006, and what they learned over their years together...
21 comments
[ 3.8 ms ] story [ 55.0 ms ] threadIn general I prefer cost range based comparisons (or first for any cost) for this reason.
There is certainly more and better optimized software for x86_64, but this makes it sound like there's nothing, which is false.
But now it’s actually here. Kudos to AMD for proving how Intel has held back the industry for years and years.
Now though, what’s next? AMD has shocked (disrupted?) the CPU scene with a completely unthinkable increase in computing power at several affordable pricing levels. Everyone can upgrade their rigs to something immensely better, no matter the budget.
So where does AMD go from here?
We know that Zen3 does away with that split. That's potentially double the L3 for single cpu, which helps some loads. Additionally, better process might mean they can also make the L3 cache bigger.
AMD hasn't talked much about Zen3 because it's a ways off (H2, potentially Q4), but from the information they've given out, we know that IPC increase is over 15%.
This will widen the gap with Intel, and it's assumed that Intel's negligible advantage in some videogames, to do with memory access latency, will be gone by then.
Intel has a couple of significant advantages here i.e. AVX-512 and DLBoost.
The most important for AMD right now is to maintain that position and level, so they can get market share (especially in server space).
It's automatic overclocking, implemented by extending the normal power management and turbo mechanisms to work with power draw limits the user can set beyond the normal specs.
Sure, these are great CPUs (I have a 3950x myself), but there is something needs to be mentioned: AMD has artificially restricted the memory limit. They have a max configuration of 256GB. I get that they want to segment these from Epyc, but it already only has four memory channels instead of eight (not a lot with 64cores/ 128 threads). Nor is 256GB a lot for such a CPU (the Ryzen 3000's support 128GB on two channels), previous Threadrippers supported 2TB!
I think AMD went too far here with their artificial limit, so it is worth mentioning.
(I'm not really a fan of TRX40, I think it is overpriced for what you get and they don't even throw in 10gbe when they start at £400+)