-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Interesting stacked GPU's idea
- Thread starter bru
- Start date
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Jesus, that is the second tech site I've seen that managed to get the TSMC 'stacked dies' announcement wrong in almost every single way.
Aside from TSVs not being new at all and TSMC and frankly the entire industry talking about AND doing it for years, the tech press went with some press release from TSMC and made up a bunch of their own crap to go with it.
Fundamentally, excessive density works with low power chips, like HBM with chips stacked on top of each other, low power so the top stack is only dissipating the power of up to 8 say 0.5W chips, stacking 200W GPUs is a strict no go and there is fundamentally nothing to be gained from having 2x200mm^2 chips stacked than side by side as has been easy to do for years.
They managed to miss that Zen is a standalone chip that works together in 2 or 4 dies also, you have quite a bit of redundancy and they also stated that this situation works better with highly parallel processing as opposed to serially as with Zen and they are fundamentally and massively wrong here. The thing that makes MCM work in Zen so well is that single threads tend to stay on one core so a thread switching across a die is kept to a minimum and overall control is made much simpler when you just send linked threads to the same core. When you have a GPU, having 4 dies (which they showed side by side as with EPYC rather than stacked?) that need to work on their own or together, you will get a lot of cross communication with a GPU and that is where this method has it's downside. Long term threads doing often the same thing with a huge amount of prediction works great in an MCM and constantly context switching massive numbers of cores needing to be filled without a central control unit is exactly where multi-die packaging fails. The issues xfire has will still be the issues a 4 die package of 200mm^2 GPUs would have, lower latency but fundamentally xfire/sli works best when each gpu works on a different part of the process and one frame is done on each cpu. There is a major reason scaling beyond 2 chips starts to drop off quickly. Until a chip can work as one, more chips makes keeping everything filled up drastically harder. It's already easy to saturate a CPU and very difficult to keep a GPU filled up with work.
TSMCs announcement wasn't about GPUs at all, stacked dies will primarily be the thing of very low power chips for several more nodes. Stacked high power chips would have so many cooling issues you would be talking about incredibly loud AND massively downclocked chips to the point where it simply doesn't work.
GPUs will absolutely go side by side smaller dies at some point soon, they won't go stacked for years, maybe decades.
Stacked chips is incredibly interesting in terms of the process technology and ideas around using it, but GPUs and >30W chips will be the absolute last place this ends up, maybe if ever, because cooling will always be a bigger problem than enough space to fit gpus side by side.
Aside from TSVs not being new at all and TSMC and frankly the entire industry talking about AND doing it for years, the tech press went with some press release from TSMC and made up a bunch of their own crap to go with it.
Fundamentally, excessive density works with low power chips, like HBM with chips stacked on top of each other, low power so the top stack is only dissipating the power of up to 8 say 0.5W chips, stacking 200W GPUs is a strict no go and there is fundamentally nothing to be gained from having 2x200mm^2 chips stacked than side by side as has been easy to do for years.
They managed to miss that Zen is a standalone chip that works together in 2 or 4 dies also, you have quite a bit of redundancy and they also stated that this situation works better with highly parallel processing as opposed to serially as with Zen and they are fundamentally and massively wrong here. The thing that makes MCM work in Zen so well is that single threads tend to stay on one core so a thread switching across a die is kept to a minimum and overall control is made much simpler when you just send linked threads to the same core. When you have a GPU, having 4 dies (which they showed side by side as with EPYC rather than stacked?) that need to work on their own or together, you will get a lot of cross communication with a GPU and that is where this method has it's downside. Long term threads doing often the same thing with a huge amount of prediction works great in an MCM and constantly context switching massive numbers of cores needing to be filled without a central control unit is exactly where multi-die packaging fails. The issues xfire has will still be the issues a 4 die package of 200mm^2 GPUs would have, lower latency but fundamentally xfire/sli works best when each gpu works on a different part of the process and one frame is done on each cpu. There is a major reason scaling beyond 2 chips starts to drop off quickly. Until a chip can work as one, more chips makes keeping everything filled up drastically harder. It's already easy to saturate a CPU and very difficult to keep a GPU filled up with work.
TSMCs announcement wasn't about GPUs at all, stacked dies will primarily be the thing of very low power chips for several more nodes. Stacked high power chips would have so many cooling issues you would be talking about incredibly loud AND massively downclocked chips to the point where it simply doesn't work.
GPUs will absolutely go side by side smaller dies at some point soon, they won't go stacked for years, maybe decades.
Stacked chips is incredibly interesting in terms of the process technology and ideas around using it, but GPUs and >30W chips will be the absolute last place this ends up, maybe if ever, because cooling will always be a bigger problem than enough space to fit gpus side by side.