Well... I'm trying to build a block for a simple numeric. (Calculating die area and reverse-engineering the Hawaii GPU)
Many clues drive me to one thesis that a full Hawaii GPU employs 48CU which is equivalent to 3072SP / 192TMU while ROP/memory part remains unchanged. Steps are following:
- Start with a Tahiti and Pitcairn die image. Normalize the size of their CU (shader cluster)
- Now we earn 6 GDDR5 controllers in Tahiti as well as 4 in Pitcairn, Tahiti's IMC is about 3 times larger than that of Pitcairn. At this point we should remind that Hawaii's 512bit GDDR5 controller is actually 30% smaller than Tahiti's 384bit controller thank to design-wise simplification of 2^n width bus. So it seems very rational that Tahiti has 3 times bigger IMC than that of Pitcairn.
- ROP partitions are (literally) covering around the shader clusters. Calculated areas (thank to MS PPT: I drawed infinitely many rectangles, get their vertical/horizontal lengths and manually multiply them) are surprisingly similar among Tahiti and Pitcairn, which suggests that they actually have the same amount of ROPs(both employ 32).
- Reverse-engineering is done. We've done almost everything. Etc & Overheads are less than 5% of total area.
- Now the "block-building" time. Start with shaders: Simply multiply 44 or 48 to 'normalized' shader cluster size.
- Estimate the IMC size of Hawaii. In this case we have two ways to approach: Multiply 0.7 to Tahiti's IMC or 2 to Pitcairn's - Fortunately, both earn the same value.
- Estimate the size of ROP partition: I simply multiply 2 to Tahiti's ROP size(which is same as Pitcairn's) because Hawaii have ROPs as twice much as them.
- Apply similar ratio for etc & overhead area (= dead space). In this case I apply 5%.
- Sum them all!
When I apply 44 CUs for Hawaii GPU, I earn 410~415mm2 die size which is quite smaller than its revealed size (438mm2). Interestingly, I could get this number only when I apply the amont of CUs equal to 48, which means AMD's highest SKU as of today (=290X) is being shipped as 'not-fully-capable' status. In other words, a full Hawaii GPU has 3072 SP and 192 TMU at the highest probability.
Well, my thread is over. How do you guys think about that?
Many clues drive me to one thesis that a full Hawaii GPU employs 48CU which is equivalent to 3072SP / 192TMU while ROP/memory part remains unchanged. Steps are following:
- Start with a Tahiti and Pitcairn die image. Normalize the size of their CU (shader cluster)
- Now we earn 6 GDDR5 controllers in Tahiti as well as 4 in Pitcairn, Tahiti's IMC is about 3 times larger than that of Pitcairn. At this point we should remind that Hawaii's 512bit GDDR5 controller is actually 30% smaller than Tahiti's 384bit controller thank to design-wise simplification of 2^n width bus. So it seems very rational that Tahiti has 3 times bigger IMC than that of Pitcairn.
- ROP partitions are (literally) covering around the shader clusters. Calculated areas (thank to MS PPT: I drawed infinitely many rectangles, get their vertical/horizontal lengths and manually multiply them) are surprisingly similar among Tahiti and Pitcairn, which suggests that they actually have the same amount of ROPs(both employ 32).
- Reverse-engineering is done. We've done almost everything. Etc & Overheads are less than 5% of total area.
- Now the "block-building" time. Start with shaders: Simply multiply 44 or 48 to 'normalized' shader cluster size.
- Estimate the IMC size of Hawaii. In this case we have two ways to approach: Multiply 0.7 to Tahiti's IMC or 2 to Pitcairn's - Fortunately, both earn the same value.
- Estimate the size of ROP partition: I simply multiply 2 to Tahiti's ROP size(which is same as Pitcairn's) because Hawaii have ROPs as twice much as them.
- Apply similar ratio for etc & overhead area (= dead space). In this case I apply 5%.
- Sum them all!
When I apply 44 CUs for Hawaii GPU, I earn 410~415mm2 die size which is quite smaller than its revealed size (438mm2). Interestingly, I could get this number only when I apply the amont of CUs equal to 48, which means AMD's highest SKU as of today (=290X) is being shipped as 'not-fully-capable' status. In other words, a full Hawaii GPU has 3072 SP and 192 TMU at the highest probability.
Well, my thread is over. How do you guys think about that?
