Hot Chips: Microsoft Xbox One X Scorpio Engine Live Blog (9:30am PT, 4:30pm UTC)

– This week it’s the Hot Chips conference in Cupertino. We’re sat nice and early, with the first talk today from Microsoft. John Sell, a Microsoft hardware veteran, is set to talk about the Scorpio Engine, found in the Xbox One X. It’s practically the only talk this week where the slides were not given out early, so I wonder what will be discussed, especially given the large amount of interest in what the Scorpio Engine is. So never mind the eclipse, let’s talk consoles.

– Despite the big hall booked, there’s a few hundred people here

– Hot Chips is a specialised conference, after all. Most people here are from tech companies, not press

– A number of familiar faces in the crowd, however. Talks later today from AMD, NVIDIA and Intel. Talks tomorrow on Server hardware too

– Just me on text and images today, it might get a bit fast paced

– Now starting with an intro to the conference

– 10% more attendees this year, 20% more paper submissions

– Had to restart my laptop, keeps freezing with intermittent wifi which is odd

– Eclipse viewing in 35 minutes – the conference were going to provide glasses, but they got stuck at customs

– Scorpio engine, 7b transistors, 359mm2 die, 50mm module

– Green around the outside are the 12 GDDR5 controllers

– 12 GB of main memory, dev systems have 24GB

– 285 GB/s peak in the lab for mem bandiwdth

– Coherent traffic is vertical to the crossbar

– 64 outbound and 40 inbound 256-bit data paths for drawing

– 2MB L2 cache with bypass and index buffer access

– out of order rasterization, 1MB parameter cache, delta color compression, depth compression, compressed texture access

– 4 MB shared L2 (2MB per quad cluster), lower main memory average latency (up to 20%)

– 12 channels and 192 banks of main memory (3x and 6x)

– 2048 entry L2I TLB and L2D TLB for 4KB pages

– 32 entry L1I TLB for 4KB pages, 8 entry L1I TLB for 2MB pages

– 40 entry L1D TLB for 4KB pages, 8 entry L1D TLB for 2MB pages, 256 entry L2D TLB for 2MB pages

– Page Descriptor Cache of nested translations (up to 4.3% perf over last gen)

– 4k60 HEVC video encoding for DVR and streaming

– Support for 4K display, 3-surface resize/blending, pre-multiplied FP alpha

– DP 1.2a, HDMI 2.0b, HDCP 2.2, two stream MST

– Audio with new firmware features, such as spatial surround (will come avail for Xbox One S)

– Video showing now, rendered on the system

– Will upload pictures when we can, something just isn’t working properly here with either my devices or the data

– ‘Any problems with backwards compatibility?’

– ‘No, the additional bandwidth will help and the previous will be emulated’

– ‘While there is added latency, there’s no compatibility issue. The GPU used to consider the SRAM just as another part of the main memory’

– ‘Q: Page descriptor cache was giving +4.5% perf in SPEC. Was that running in a VM?’ ‘Yes, the benefit we get negates the VM loss’

– ‘Does designing the SoC give better options for power management?’

– ‘Yes, to a significant degree with Scorpio but since the original Xbox One – we want every system to give the same perf, and we don’t want to give people an unfair advantage in MP games (like in PCs), so we strive we don’t have variation which makes power management more challenging, and we don’t want to through chips way. Every chip is married to its circuit board, variations in the power supply, and trying to tune out all the margin possible’

– ‘Shader power gating?’ ‘We have clock gating throughout the design, but for power gating, we do not do any power gating within the shader array. We change to operate at lower frequencies in certain modes, e.g. BluRay discs can go through fixed function hardware we clock gate the shaders’

– That’s all for the talk, let me upload the pictures if I can

– Now that everyone in the hall has left to watch the eclipse, some photos are going through. Of course, photos that were sent straight to upload aren’t saved on my devices. I’m waiting for the slide deck and we’ll do a proper analysis on this.

– That’s all for the live blog. There’s a Knights Mill talk later today that we’ll try and get a better experience for πŸ™‚