I believe AMD should reinstate HBM support for STRIX halo regions.
I believe AMD should reinstate HBM support for STRIX halo regions.
This could be considered a GPU matter, though its unified memory approach makes it fit here. Strix Halo stands out because of its impressive memory capacity, which is the main selling feature. Being able to switch to VRAM adds an extra advantage, but it also restricts upgrades since the design can't be changed easily. This supports the case for HBM, as it was too advanced to be modular and Strix Halo shares that limitation. Faster memory helps, but simply increasing speed isn’t enough—having more capacity when loading data is crucial. Given its high cost, people would likely pay more for better performance and long-term reliability. HBM was innovative but didn’t reach its potential; since 2017 improvements have made it more affordable now.
AMD likely weighed various factors when building Strix Halo. Perhaps cost, power consumption, or performance balance wasn’t ideal for that product. It could be a mix of these considerations plus others. The specifics of memory technology aren’t crucial as long as the device functions properly. I’m unsure about the current state of HBM, but it seems Strix Halo wouldn’t fully leverage it. A better approach might be expanding the Halo lineup to include larger GPUs with at least double the capability of today’s models, making LPDDR less viable and favoring GDDR or HBM instead.
Given the severe issues with HBM GPUs and their fragile thermal systems, it seems unlikely they'll be adopted in consumer products anytime soon. Perhaps only data centers will see such advancements.
It remains unchanged except for consumer goods. The cost is too high to make a profit. Nvidia's data center solutions employ HBM3, yet they offer superior profit margins.
HBM is expensive due to its advanced features. It’s a technology likely to remain in data centers because it pushes consumer GPU prices even higher, while also reducing overclocking options—something that could increase frustration with the brand and discourage further interest.
I believe it has both pros and cons. HBM delivers great speed but higher memory needs have risen too. Would you opt for a 32GB HBM setup with better performance versus a 128GB LPDDR5 system in the same budget? Giving more options would likely raise the cost because of shipping, development, and support expenses.
AMD frequently offers additional memory channels when they wish to maximize performance. This boosts data transfer rates and expands available storage options. For quite some time, no laptop has supported more than dual-channel RAM. The last model I recall was a desktop upgrade with an LGA 2011 socket in 2013. Today it's still feasible, though it would demand significant architectural changes to fit a quad-channel controller. Or perhaps even more ambitious: 8-channel RAM delivering roughly 300GB/s speed with 512GB of storage. That would be truly remarkable. Far beyond current standards compared to HBM technology.
It wouldn't necessarily raise the cost; it would require a complete shift to two distinct platforms at the silicon level. HBM would also add significant complexity to implementing performance tiers without simply manufacturing full 32GB HBM-halo devices and adjusting firmware. Likely, it would be more affordable than HBM.
I've been thinking about this and looked into it. The expenses for HBM are really steep. I questioned myself about how much more complex the IMC would be for 4 or 6 DDR channels, but the reality is that HBM could nearly double the product cost. I ran through a hypothetical using HBM2e—it appears to offer the best balance of price and performance. Still, adding a silicon interposer with HBM adds another layer of expense. The biggest challenge isn't just the bill of materials; it's the huge drop in yield when you start etching on silicon. I think AMD still uses organic interposers for CPUs because silicon would give better clock speeds (FCLK) and lower power use, which both boost performance. The problem is, if you lose 10% of your production due to defects, especially when supply is tight, it becomes a real issue. Even before accounting for the extra material costs, having an interposer and stacked RAM means you have to account for a roughly ten percent increase just to handle defects. This would cut supply and likely push prices up further.