9800x3D 3D v-cache verified beneath the CCD. Intel will never recover from this...
9800x3D 3D v-cache verified beneath the CCD. Intel will never recover from this...
There's probably a reason they changed their naming scheme, the 14900k beating the 15900k looks far worse than the 14900k beating the 285k. That's the hope though, that Arrow Lake is a Ryzen 1000 moment for Intel. The problem is that Ryzen at least provided a multicore and value advantage back then that at least made their initial architecture useful for productivity. Ryzen's gaming dominance didn't start happening until Ryzen 3000, then they overtook Intel with the 5800x3D. 7950x3D merged these two advantages and likely the 9950x3D will be the absolute king of processors for 2025. This change to 3D v-cache confirms that the 9950x3D will have +64MB for both CCDs, since there's no reason not to with the SRAM die below the CCD. At this point, I wouldn't doubt if Intel sort of gives up on the gaming desktop end, something a lot of us though AMD would do until 3D v-cache came along. L3 cache wasn't even a major discussion point until that happened since there had never been a scenario where we could control for the amount of L3 cache, usually being a small bump with architectural improvements. 5800x vs 5800x3D providing a control for what happens when you increase the L3 cache by 3x. Publicly traded companies already on the down can't afford patience.
Consider the expense of integrating these chiplets. Sales prices don't always show the full picture; comparing similar setups with and without them helps. Typically, larger caches can slow performance. Pushing capacity further amplifies this effect. Bandwidth becomes more efficient, delays shrink. Which tasks suit each scenario? https://chipsandcheese.com/p/amds-7950x3...ets-vcache Would appreciate seeing how this plays out, though cost remains a key factor. It seems less appealing for budget-conscious segments. Premium options like the upcoming Strix Halo might offer better value.
I believe this highlights why halo products are important. Neither Intel nor AMD aimed to create the absolute top gaming CPU; they focused on delivering the best CPU available. In contexts like these discussions, gaming often takes a very high priority. You don’t have to play perfectly on ARL if it’s lagging behind—AMD has demonstrated over the years that even with weaker GPUs, you can still compete.
AMD probably experimented with it before my time. The idea of boosting 3D v-cache might not justify the benefits, or gains may slow down as L3 cache sizes increase for most programs. A 32MB likely falls short compared to a 64MB version, and adding another 128MB could be too costly—especially when alternatives exist mainly for price. They’d probably need specialized SRAM modules tailored to each die size, possibly using BGA spacers for graphics components. It seems feasible to design APUs with consistent dimensions across Zen5 generations, similar to their CCD designs, enabling better compatibility. This approach might work better in mini PCs or high-end desktops than on handheld devices.
I’m uncertain about the exact task, yet I need to be present since AMD offers CPUs with 3D V-cache.
In basic terms about how systems work: If information isn’t in the fastest memory (L1), check L2, then L3. If it’s still missing, go to RAM. Cache size influences how often data is found, but bigger caches cost more and slow things down. Builders must balance these layers for best results. CPUs have become much faster, yet RAM can’t keep up, which is why we’re looking at bigger caches or other tricks. High-end chips often sacrifice socket design to boost performance. When you switch from socketed to high-end models, it makes performance upgrades more expensive. Getting data quickly is just as important as processing it.
The EYPC 7003 and 5800x3D came out together in a short timeframe, but I wouldn't claim the 3D v-cache was the first to appear on servers. There could be specific situations where a 9800x3D performs better than a 9700x, but how many would benefit from a 9950x or a Threadripper/EPYC setup? For scenarios where the additional 64MB of L3 memory makes a real difference, it seems to be mainly for gaming—where that’s the primary use for 3D v-cache on AM4/AM5. MCM and socketed chips are different concepts.
In the case of Phoenix APUs, there were various versions: socketed LGA AM5 models with the 8600G/8700G, as well as BGA variants like the Z1 Extreme, 7840U, 7940HS, etc. We’ve also encountered soldered BGA versions of standard Zen MCMs, such as the 7945Hx3D.
AMD Ryzen 9 7945HX3D details show that even with a custom Zen 5/RDNA3.5 IOD+GCD setup, it’s hard to fit a full MCM-based Ryzen design inside an AM5 chip with three dies. The extra L3 memory is more relevant for gaming workloads, where those are the most common use cases for 3D v-cache in AM4/AM5.
In markets beyond gaming, a 9950x3D or Threadripper/EPYC usually offers better performance. A few applications might benefit from a 9950x3D, but they don’t dominate enough to justify a full product line shift. AMD seems to agree that the 9800x3D is mainly for gaming—very limited other uses exist where extra L3 memory matters.