New GPU Compute System - Page 9

diep · 2017-12-07, 23:56

Ok i'm digging in older articles.

https://www.anandtech.com/show/4455/...-for-compute/4

Here is see first time in my life picture with 16 ALU's for each SIMD as description for older GCN.

diep · 2017-12-08, 00:02

Ok if i understand it correctly AMD moved to 16 alu's in each SIMD, which is pathetic little, to sail around an old problem they had, that's that they can't execute different kernels (wavefronts) at the same time.

By having now 4 different SIMDs in each CU, it allows within the same CU different wavefronts to execute. So instructions that take longer than 1 clock throughput then also can get handled meanwhile the other execution units still can execute simple instructions then, instead of needing to wait long time.

edit: Ok so i need to dig more into how many wavefronts can most ideally get executed simultaneously (not the maximum as that's clear) to still get good IPC. Blindfolded i would gamble at 8 for GCN.

diep · 2017-12-08, 00:13

Ok that opens possibilities! More than 4 wavefronts @ 64 streamcores doesn't make sense it seems if i estimate it here.

Means 64KB / 4 = 16KB LDS available.

For the internal iterations of the FFT that's more than sufficient.
That's 2K doubles or 2^11 is 11 iterations.

2017-12-08, 00:02	#90
diep Sep 2006 The Netherlands 3·269 Posts	Ok if i understand it correctly AMD moved to 16 alu's in each SIMD, which is pathetic little, to sail around an old problem they had, that's that they can't execute different kernels (wavefronts) at the same time. By having now 4 different SIMDs in each CU, it allows within the same CU different wavefronts to execute. So instructions that take longer than 1 clock throughput then also can get handled meanwhile the other execution units still can execute simple instructions then, instead of needing to wait long time. edit: Ok so i need to dig more into how many wavefronts can most ideally get executed simultaneously (not the maximum as that's clear) to still get good IPC. Blindfolded i would gamble at 8 for GCN. Last fiddled with by diep on 2017-12-08 at 00:05

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Cost to compute prime	Unregistered	Information & Answers	30	2013-12-18 03:34
Doubled compute power for a day?	Christenson	PrimeNet	19	2011-10-26 08:29
New Compute Box	Christenson	Hardware	0	2011-01-15 04:44
Compute billions of digits of Pi using GMP	M0CZY	Miscellaneous Math	5	2010-10-14 09:40
My throughput does not compute...	petrw1	Hardware	9	2007-08-13 14:38

2017-12-07, 23:56	#89
diep Sep 2006 The Netherlands 1100100111₂ Posts	Ok i'm digging in older articles. https://www.anandtech.com/show/4455/...-for-compute/4 Here is see first time in my life picture with 16 ALU's for each SIMD as description for older GCN.

2017-12-08, 00:13	#91
diep Sep 2006 The Netherlands 3·269 Posts	Ok that opens possibilities! More than 4 wavefronts @ 64 streamcores doesn't make sense it seems if i estimate it here. Means 64KB / 4 = 16KB LDS available. For the internal iterations of the FFT that's more than sufficient. That's 2K doubles or 2^11 is 11 iterations.