Just curious: What does Crystalwell do differently than Turbocache or Hypermemory?

[tipoo]

Both of the latter are technologies from AMD/Nvidia to tap into system memory when their low end chips ran out of video memory. This allowed them to strap less memory onto low end chips.

My curiosity is if this is much different from how Intels Crystalwell works. They both have a smaller faster pool and then revert to the system memory when needed, what's the difference?

Is Crystallwell just smarter about what to cache, whereas the other two just do a simple overflow, fill video memory, then fill everything else in system memory even if you need it more? While Crystalwell would try to keep everything you need more in the eDRAM? Or are they more similar than that?

 

[ShintaiDK]

Crystalwell is a L4 cache with a fast access. turbocache/Hypermemory is just memory over the PCIe bus.

 

[tipoo]

Crystalwell is a L4 cache with a fast access. turbocache/Hypermemory is just memory over the PCIe bus.

I know that bit, but I'm just curious if there's any other difference in implementation. Hypermemory/turbocache put things in video memory first, then resort to system memory if need be, just wondering if the way Crystalwell splits what it puts in each memory pool is different. Were the older two smart enough to decide what the best things to put in the faster memory pool vs the slower would be?

 

[ViRGE]

Both Crystalwell and TC/HM are based around containing as much as they can in their respective local memories, and spilling the rest out to main memory. Functionally they're close enough in operation that specific differences are not worth mentioning (Crystalwell technically is just a pure cache, TC/HM used split memory pools with local having priority), but it should be noted that at just 128MB in 2013+, Crystalwell is far closer to a cache than local memory due to how small its capacity is, relatively speaking.

What killed TC/HM is that accessing main memory over PCIe is slow. Very, very slow. There's a general lack of memory bandwidth (4GB/sec) and latency is through the roof. Whereas Intel's iGPUs get equal priority access to the DRAM directly through the Intel ringbus, which gives them full and fast access to DRAM once you need it.

 

[SPBHM]

I think TC/HM was just replaced by the Windows (vista+) shared memory doing the exact same thing?

I think the "Iris Pro" style cache is different from the GPUs with memory + shared memory because the l4 cache is also shared with the CPU cores? a lower level thing

I think it was 100% possible to use a TC card with just the onboard memory, and not possible with the Iris Pro IGPs?

 

[ViRGE]

I think TC/HM was just replaced by the Windows (vista+) shared memory doing the exact same thing?

I think the "Iris Pro" style cache is different from the GPUs with memory + shared memory because the l4 cache is also shared with the CPU cores? a lower level thing

I think it was 100% possible to use a TC card with just the onboard memory, and not possible with the Iris Pro IGPs?

Correct on all counts.