1.. _skiboot-6.0.6: 2 3============= 4skiboot-6.0.6 5============= 6 7skiboot 6.0.6 was released on Thursday July 19th, 2018. It replaces 8:ref:`skiboot-6.0.5` as the current stable release in the 6.0.x series. 9 10It is recommended that 6.0.5 be used instead of any previous 6.0.x version, 11especially in the case where NVLINK2 GPUs and/or Mellanox CX5 adapters are 12being used. 13 14Over :ref:`skiboot-6.0.5` we have several important performance related bug 15fixes and one stability bug fix: 16 17- phb4/CAPI: Reallocate PEC2 DMA-Read engines to improve GPU-Direct bandwidth 18 19 We reallocate additional 16/8 DMA-Read engines allocated to stack0/1 20 on PEC2 respectively. This is needed to improve bandwidth available to 21 the Mellanox CX5 adapter when trying to read GPU memory (GPU-Direct). 22 23 If kernel cxl driver indicates a request to allocate maximum possible 24 DMA read engines when calling enable_capi_mode() and card is attached 25 to PEC2/stack0 slot then we assume its a Mellanox CX5 adapter. We then 26 allocate additional 16/8 extra DMA read engines to stack0 and stack1 27 respectively on PEC2. This is done by populating the 28 XPEC_PCI_PRDSTKOVR and XPEC_NEST_READ_STACK_OVERRIDE as suggested by 29 the h/w team. 30- phb4: Disable nodal scoped DMA accesses when PB pump mode is enabled 31 32 By default when a PCIe device issues a read request via the PHB it is first 33 issued with nodal scope. When accessing GPU memory the NPU does not know at the 34 time of response if the requested memory page is off node or not. Therefore 35 every read of GPU memory by a PHB is retried with larger scope which introduces 36 bandwidth and latency issues. 37 38 On smaller boxes which have pump mode enabled nodal and group scoped reads are 39 treated the same and both types of request are broadcast to one chip. Therefore 40 we can avoid the retry by disabling nodal scope on the PHB for these boxes. On 41 larger boxes nodal (single chip) and group (multiple chip) scoped reads are 42 treated differently. Therefore we avoid disabling nodal scope on large boxes 43 which have pump mode disabled to avoid all PHB requests being broadcast to 44 multiple chips. 45- npu2/hw-procedures: Enable parity and credit overflow checks 46 47 Enable these error checking features by setting the appropriate bits in 48 our one-off initialization of each "NTL Misc Config 2" register. 49 50 The exception is NDL RX parity checking, which should be disabled during 51 the link training procedures. 52