Hacker News new | past | comments | ask | show | jobs | submit login

It is disappointing only for gamers.

For scientific and technical computing, a 9950X provides the greatest improvement in performance per dollar since five years ago, in 2019, when the first 16-core desktop computer, 3950X, has been introduced by AMD.

This is caused by the doubling of the AVX-512 throughput per core in desktop and server Zen 5.

The new Intel Arrow Lake S desktop CPU, 285K, also provides a 50% increase in AVX throughput over Alder Lake/Raptor Lake, but this increase remains small in comparison with the throughput doubling in 9950X, which ensures a 4/3 throughput in 9950X (actually even more than that in practice, due to the better AVX-512 instruction set).

For games and for applications that are not dominated by array operations, for instance for software project compilation, the performance of AVX or AVX-512 code does not matter much, but there are a lot of professional applications where 9950X will provide a great jump in performance.

For things like software project compilation, it is likely that Intel 285K will provide the best performance per dollar, due to its 24 cores. Unlike the older Intel E-cores, the new Skymont cores have a multithreaded performance that can be similar with that of the old Zen 4 cores or of the P-cores of Alder Lake/Raptor Lake, except for the applications where a high contention between threads for their shared L2 cache memory happens.

So this time one can truly consider Intel 285K as a 24 core CPU (without SMT) from the point of view of the applications dominated by integer/pointer computations, while in Alder Lake and Raptor Lake the E-cores were better viewed as half cores, due to their lower performance, so the top models of Raptor Lake were better thought as equivalent with a 16C/32T CPU, not with a "24-core" CPU, as advertised.

The 9950X has become more important than the past desktop CPUs, because the prices of the server CPUs have greatly increased during the last decade, so now, for most small businesses or individuals, it has become not cost-effective to use a "real" server CPU. Instead of that, it is much more economical to use servers with 9950X (and ECC memory). Multiple servers with 9950X are much cheaper than a single server of similar capacity with an Epyc CPU.




I believe also that the 9000x3D series (from my memory of rumours) also has the 3D cache on both CCXs, meaning no latency with cross-CCX communication.


I think that has the possibility to make it worse honestly. It’s not like the contents of the cache is duplicated. Instead it’s split across a ccx boundary and if the data is in the wrong cache you’ll be hit. Now clever thread management can help avoid this but so far the 9xxx series has shown terrible thread affinity choices with many existing games and apps. I’ll wait and see how the 3D cache helps here.


AMD claims that the 9000X3D series is the product that will provide the game performance increase expected by gamers for a new product generation.

Of course, that remains to be seen, but it is plausible.


Long term as games start using AVX512 I expect the 9000 series will be seen as a big step up against previous generations. One of those "fine wine" things.


That doesn't matter now. Gamers will vote with their dollar, and they are fickle.




The deadline for YC's W25 batch is 8pm PT tonight. Go for it!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: