4 ms·
Really would like to know more about the GPU and Ram. At 128 gpu cores vs the 8 in the M1 you talking 16 times the potential GPU power which could put it up the
by SigmundA 5y ago
Really would like to know more about the GPU and Ram. At 128 gpu cores vs the 8 in the M1 you talking 16 times the potential GPU power which could put it up there near the top of discrete GPU's.
- willis936 5y agoIt takes a lot to scale a GPU up that far. Memory throughput requirements (the actually expensive part) increase linearly with number of cores. It will either cost a large fortune and use a Lot of power or be memory throughput starved (and cost a small fortune).
- SigmundA 5y agoTile based deferred renders help with memory bandwidth requirement (partly why UMA mobile GPUs and the M1 use it). The M1 has 16mb of cache coherent with CPU and GPU and then there is tile memory in each core.
- johncolanduoni 5y agoThat would be interesting for the Mac Pro since the kind of applications most people who would buy one would be running have probably never even been run on a tile based GPU, let alone optimized for it.
- SigmundA 5y agoIf however you have developed for mobile and specifically iOS you have a big head start: https://developer.apple.com/documentation/metal/gpu_features/understanding_gpu_family_4 https://developer.apple.com/documentation/metal/gpu_features...
- paulpan 5y agoCurious about this too. Anything more than 16 GPU cores seems like it'd too big to fit on die (share with CPU cores and everything else), so may need to be a discrete solution. Would Apple then make discrete GPU solutions using PCIe5? Makes sense for the Mac Pro given the edge cases for requiring multi-GPU solutions.
- ksec 5y ago>Anything more than 16 GPU cores seems like it'd too big to fit on die 32 HP Core with 8 HE Core, 128 GPU Core give you ~650mm inclusive of L2 and SLC. Adding in I/O, NPU, ISP, VPU etc this chip should still be within the reticle limit of ~850mm2.
- tyingq 5y agoFor some very loose comparison, the Fujitsu A64FX is a 60mm square package and has 48 ARM cores and 32Gb of memory, though no GPU cores. Block diagram over a die shot: https://3s81si1s5ygj3mzby34dq6qf-wpengine.netdna-ssl.com/wp-content/uploads/2019/11/fujitsu-a64fx-block-diagram.jpg https://3s81si1s5ygj3mzby34dq6qf-wpengine.netdna-ssl.com/wp-...
- ksec 5y agoThe A64FX has much smaller cores though using the similar design to the N1. Apple's HP Core are nearly double the die size ( Hence why it is faster, it is a trade off between transistor budget )
- SigmundA 5y agoThey have been talking up the advantages of unified memory so my guess is they will stick with integrated, but I would imagine they will end up supporting AMD or Nvidia discrete to some degree.
- lvl100 5y agoThought this was the most important piece of info in that article. Could be a game changer for a number of applications considering the softwares and dev support Apple can provide.