AMD's new desktop CPU oozes cache out of all 16 cores AMD aims to extend its lead in desktop gaming with a new CPU, dubbed the Ryzen 9 9950X3D2 Dual Edition. This top-of-the-line part has 16 cores fed by an absolutely massive 208 MB pool of cache, with memory spread across both CCDs. The hotly anticipated processor chip is essentially a modified version of the 9950X3D announced …
That is the CPU I've delayed my upgrades for... Now if only the RAM would not be so expensive, especially since I prefer real ECC... But that is a different topic...
I know that my workloads will make use of that extra cache. Before it was not worth it, the 5950 is still a very fast CPU and that asymmetric 3D cache the follow up CPUs offered until now is just not my taste. And Threadripper is simply too expensive...
an upgrade for those that already have a AM5 platform
Forgive me for being thick, but a ~10% speed boost for likely north of $750 (assuming you keep everything else the same, and then end up with a $700 part going spare) doesn't strike me as particularly worthwhile in the vast majority of cases.
Also, for games, if core "parking" is still a thing then won't it perform pretty much identically to the existing asymmetric part? With one set of cores parked, you end up on both with half the cores, all of which have V-cache, and the newer part has a lower boost clock too (though again, 100MHz out of 5.7GHz is less than a 2% difference in top speed, so would it even be noticeable?)
I, too, would like to see some lower-end parts from AMD. I'm not convinced of Intel's P/E strategy, particularly as marketing to the public just adds the two numbers together*, but at least they are in the low-end segment; surely there is still a market for processors with recent specs in the <$200 (<$150?) range?
M.
*and it adds a layer of complexity for people like me who have been used to a simple choice governed mainly by number of cores, clock speed and price. How do you compare a 4P+8E part (no HT) to a 6 full-core (with HT) part at a similar price? It depends on workload, but as for most people the workload is "well, it depends"...
I suppose some people might have a low-end CPU from 2023 which also used the AM5 socket and it would be a more substantial update for them, but so would any other high-end modern chip and, depending on their system, it might need other changes to accommodate that. The most important factor is when the cores need to be parked and how much control you have over that, and I don't know either of those answers. The article basically tells me that there's a significant speed boost except some times it's a little worse than before, and without clarification of which times that is and whether I'll be in them, that makes any decision very difficult.
The P/E complexity does make it harder to select a chip, but in my experience, it comes with substantial benefits. A lot of uses either need one or two fast cores but don't benefit from more parallelism or can use as many cores as you have and therefore benefit from having as many non-HT ones as possible with 4 fast and 8 slow cores often being faster than 8 fast cores. The P/E setup works for both of those. Buying a chip based only on clock rate, core count, and price was already missing some important aspects like power levels and cache sizes, which is why people often end up buying on benchmarks which aren't perfect but at least compare more direct information about performance.
You could have 1 TB of RAM over 20 18 years ago too, with x86 technology, and even more than only 16 cores. However those machines started at the 6 digit price range... Lots of noise... Really high power usage... Don't tough those RAM modules during operation, they were, literally, screaming hot.
It turns out you can make terribly programmed software like games run much faster by caching everything into SRAM (but I guess that makes sense, as a game that unpredictably fetches things from RAM over and over is much accelerated if you can cache 10MB chunks of fetches to be almost instant (RAM itself is extremely fast compared to even fast SSD's, but 10,000 RAM fetches does take time)).
Quite an interesting chart design - generally >100% increase means over 2x as fast, but the baseline is 100%, so there's only a 5-7-13% performance increase (meaning the only practical reason to use a 9950X3D2 over the 9950X3D was if the 9950X3D2 had the same price, or less than 13% more, but typically a year old CPU has had a price decrease of more than 13%).
Which is less expensive than that extra cache.
Only if you have one of the existing half populated 3D cache systems of course.
It's actually a win-win because without the OS bouncing workload around cores you get a significant perf. boost from that as well (or rather, don't take the default performance hit).
You can do that on Windows too, even automated if you want. Just a powershell command away. Literally. Example for (smt-)core 0 to 7:
(Get-Process -Name notepad).ProcessorAffinity = 0xFF
same for 16 to 31
(Get-Process -Name notepad).ProcessorAffinity = 0xffff0000
Same for 48 to 63 (throws an error here since I don't have enough CPU cores *sniff*...)
(Get-Process -Name notepad).ProcessorAffinity = 0xffff000000000000
if Intel came add some longevity to the next platform and add a bunch more cache, I might considering migrating back to them. The productivity perf is amazing, but Im a gamer so they need to catch up and match the 10800X3D.
AMD are clearly taking the piss on prices just as intel did, and so need taking down a peg or too.
