Are you considering purchasing a system with non-upgradable HBM memory like CPU chips?
Are you considering purchasing a system with non-upgradable HBM memory like CPU chips?
Imagine a solid foundation of RAM directly built into the chip itself, forming a powerful L4 cache. If AMD or Intel could achieve this, would you invest in it knowing your memory capacity is fixed for the CPU’s lifetime? The concept centers around a bigger substrate, similar to TR4, that also holds the DRAM, removing the need for traditional DIMM slots. For a guess, performance improvements could match or exceed the difference between older and newer 3D caches—potentially doubling speeds in certain cases. Personally, I’d trade flexibility for speed if it meant such gains, as seen in recent laptops and portable devices. The idea is supported by today’s rapid PCIe speeds, which could help offset any downsides of fixed RAM.
Hmm, you're talking about DRAM, right? Basically an Apple Mx. version. It depends on how much it boosts performance, but I think the gains won't be double in most cases. Plus, the cost is a big factor. I usually don’t upgrade much; my starting RAM is enough. The main thing is getting a good price. If it matches the current RAM price, it’s easier. But if you’re paying 5 to 10 times more than regular RAM, that’s not worth it.
Instead of building a huge SoC and adding a GPU there, it could be more practical. This approach works well for certain tasks. I think large language models would gain a lot from the setup I described earlier.
Apple has demonstrated impressive capabilities using their 512-bit and 1024-bit wide memory bus chips with remarkable power efficiency. This isn<|pad|>, not comparable to conventional cache and 3D-vcache solutions, yet offers significant performance gains in specific use cases simply through vast memory bandwidth. AMD's Strix Halo is expected to feature a 256-bit memory bus, and I imagine myself swapping a large desktop setup for a powerful APU—provided the RAM costs remain reasonable. I’m also eager to see exciting developments with Intel’s Lunar Lake MX.
I imagine this could grow in importance for embedded systems and mini computers eventually, though the large quantity of SKUs would limit practical use with higher power choices.
They reduce latency by avoiding RAM sockets. Keeping connections short and resistance low is key for performance.
It wouldn't be unexpected to see this or similar technology in LPDDR5 mainboards with CAMM2 modules for top-tier desktops next year.
Absolutely, you've captured the essence well. It's about building a unified system where components fit together seamlessly, similar to how CPU upgrades used to be. The idea of a monolithic platform would simplify things, especially with the shift from AM4 to more advanced architectures. It's intriguing how far we've come with technologies like HBM and RDNA3, which could pave the way for such designs. I'm exploring this concept to see if it makes practical sense, especially given the challenges of integrating GPUs with high-performance dies.
I need to understand this, but I worry that Intel or AMD might design setups that seem completely unreasonable just to push sales. It’s similar to how Nvidia handles VRAM for their cards. Picture the Ryzen 5 10400 only comes with 8GB of this "HBM" stuff, even though it offers solid performance when RAM isn’t an issue. You’d then have to pay more for the Ryzen 5 10600X to get the full 16GB you need. The Ryzen 7 10700X still only provides 16GB. So to reach 32GB, you’d have to upgrade to the Ryzen 9 10900X. The Ryzen 7 10700X still only has 16GB. If you aim for 64GB, you’d need the Ryzen 9 10950X. And the Ryzen 9 10950XE is what you’d need for 128GB. I just don’t feel confident companies would play this game. It seems better if AMD and Intel let third parties handle RAM production for their chips, for the benefit of consumers.
Beyond a certain limit, its value diminishes significantly, since cooling both with one heatsink reduces performance potential compared to independent cooling. Running power solely to one component leaves insufficient room for data lines due to the massive power plane required. I don’t see much benefit in relying on a single 360-480mm AIO if dual air coolers—separate for CPU and GPU—offer better stability and adaptability.