Slashdot Mirror
AMD Launches Lower Cost 12- and 24-Core 2nd Gen Ryzen Threadripper Chips (hothardware.com)
MojoKid writes: AMD launched its line of second generation Ryzen Threadripper CPUs over the summer, but the company offered 16-core and 32-core versions of it only at the time. Today however, the company began shipping 12-core and 24-core versions of the high-end desktop and workstation chips, dubbed Ryzen Threadripper 2920X and 2970WX, respectively. All 2nd Generation Ryzen Threadripper processors feature an enhanced boost algorithm that came with AMD's Zen+ architecture that is more opportunistic and can boost more cores, more often. They also offer higher-clocks, lower-latency, and are somewhat more tolerant of higher memory speeds. All of AMD's Ryzen Threadripper processors feature 512K of L2 cache per core (6MB total on the 2920X and 12MB on the 2970WX), quad-channel memory controllers (2+2), and are outfitted with 64 integrated PCI Express Gen 3 lanes. The new Ryzen Threadripper 2920X has a 180W TDP, while the 2970WX has a beefier 250W TDP. In highly threaded workloads, the Threadripper 2920X outpaces a far more expensive 10-core Intel Core i9-7900X, while the 24-core / 48-thread Threadripper 2970WX is the second most powerful desktop processor money can buy right now. It's faster than Intel's flagship Core i9-7980XE, and trailed only AMD's own 32-core Threadripper 2990WX. Pricing for the new chips falls in at $649 for the 12-core 2920X and $1299 for the 24-core Threadripper 2970WX.
Revised Mac Mini, offering an AMD chip.
Maybe even the redesigned Mac Pro...
To me it's been quite odd that Apple is so keen on AMD GPUs, while never using them for primary processors.
"There is more worth loving than we have strength to love." - Brian Jay Stanley
But the single core performance seemed underwhelming on the new AMD processors, especially for the price. It seems like virtually everything I do is limited by single thread performance, with multiple cores mostly being good for multitasking. Am I missing something?
Reports are that the 9900K draws way more than 95 watts when running overclocked for the fiddled benchmarks. A lot of complaints about cooling problems out there. A lot of doubt about accuracy of benchmarks. And the chip is out of stock everywhere, so a lot of people are calling it a paper release. A lot of talk about cancelling orders and going with 2700X or Threadripper instead.
When all you have is a hammer, every problem starts to look like a thumb.
About 6 months ago I built a new budget video editing rig. I was torn between going with an i7 8700 or a an AMD2700 but opted for Intel because of QSV.
QSV allows for decoding and encoding H264 and H265 video in hardware using the on-chip video hardware. It's brillant watching my 6-cores idling while rendering 4K video into H265 files at realtime speeds. Try that with your AMD processor :-)
However, these days I'd probably go for the 1950 Threadripper (cheaper and almost as good as the 2950 because those extra cores *are* useful in good video NLEs such as Davinci Resolve.
What matters is cache size, L2 and L1. Losing a few cores and bolstering cache will improve performance in quite a lot of cases.
The Key is that threads don't talk that much. The amount of shared information needed to justify cores in close proximity and a huge shared cache isn't there a lot of the time.
Cheaper SMP - not difficult with PCI-E's design - would leave much more room for the critical L1 cache, reduce the heat burden on a CPU, and potentially quadruple the number of cores (since 4-way SMP is not too bad).
Close proximity on silicon only matters when you're communicating between units. Totally independent computation can be done anywhere. Port Linux to SystemC and compile it to an ASIC if you want. Should run fine, even if taking no cores at all.
There's zero relationship between half the system services on Windows/Linux and the applications being run, so there's no gain through physical proximity. There's no latency issue to resolve.
On the other hand, those same services reduce the L2 cache space your applications have available, so your applications are fetching from main memory more than they have to, just so that the services can poll for any work, scan your computer for viruses, etc. Not essential activity when playing Elite:Dangerous or Kerbal Space Program.
Not dissimilar to splitting off cores that were connected with graphics, now known as GPUs.
I'm thinking back to a model of an Amdahl mainframe. Similar sort of concept, segregated processing.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)