cache will have little effect on productivity as most productivity workloads tend to favour bandwidth over speed.
gaming on the other hand benefits more from low latency than high bandwidth.
will answer some of your questions…
there’s no mention of amd implementing it on threadripper going by that pdf
A lot of my productivity will spend time on compressing and decompressing. Phoronix seems to suggest that X3D also helps with zstd and other productivity stuff.
In addition, the non-X3D cores are clocked higher. I want to downclock the non-X3D cores to make them more efficient and behave more like Intel’s E-cores. Low frequency, high performance-per-watt, suitable for SMP workloads. In this case, the non-X3D cores lose their frequency benefit to the X3D cores and will always (?) be slower. The benefit of X3D is going to be bigger versus the non-X3D cores this way.
So eco-mode combined with undervolting might push the 7950X3D into a 65W TDP or at least getting into that territory. Then you get a lot of performance for not so much power that can run 24/7 for many years. With Threadripper I wanted to do the same, but only the HEDT (4 memory channel) version supports overclocking. Probably that also means the workstation 8 channel ‘pro’ version will not support undervolting or curved optimizer?
I wanted one system for both gaming and productivity, and also liked that my productivity would become more efficient with caching, particularly if the performance also increases such as with zstd and other productivity workloads. I will need to do my own tests to see how much the X3D favors my own workloads. I run a lot of scripts in VMs. I liked the idea of figuring out which tasks favor X3D and make them run on the X3D cores. I could change my scripts that way.
But if Threadripper 7000 will not get any X3D, that would be a boner. With so many chiplets, it feels stupid that none of them are X3D. Then i cannot assign my workloads that favor the cache to run on those chiplets/cores. Threadripper 7000 becomes less desirable to me in this case. I wanted one system to be both great for productivity and gaming. Threadripper 7000 X3D would be my dream system.
If you spend a lot of time doing compression, your money would be better spent on an intel quickassist card. Get one with a model number above 8950. The early ones don’t support enough codecs. 8970 and up can also accelerate zfs. On eBay they are around $350, from the servethehome benchmarks doing compression they are worth more than 20 cores of compute, sometimes as much as 50 cores of compute for one $350 card.
and x3d is coming, but not till they meet the demand with Ryzen.
when you read though this you will see that the cache accounts for about an 8% uplift. most of the increases in performance are coming from architectural changes to the front end and over provisioning memory bandwidth. (up to 40%).
like i said 3dx cache has limited use for productivity…
thats not to say its useless.
its just not as useful as it is to gaming workloads.
I completely agree. I just bought the epyc 9124, which should arrive tomorrow. I don’t have a task in mind for the computer beyond educating myself on the current hardware and software. I need actual hands on time not just using cloud instances, but deploying hypervisors and container management.
I could have bought a 7xxx series epyc, but in many ways it would have been a duplication of effort. This system will probably last until the next socket becomes available, and I may get several CPU upgrades, so 8 years for the socket and a new cpu every 3 years? something like that.
I’ll also note zstd is supported by NVCOMP so you could be pushing tens of GB/s of compression with an Nvidia GPU instead of relying on your CPU to do that kind of work. (A really rough estimate, typical CPU achieves around 1GB/s per core)