Intel held its Architecture Day on Tuesday. A trove of disclosures that I am still digesting. Stay tuned for analysis. One easy thing that I can say with certainty is that it will drive discrete GPU revenue and share. That’s not as easy to assess as you may think. $INTC
Gaming, Intel fields a performance desktop GPU "Arc" in Q1 that's up there with NVIDIA & AMD. Not at the top but close. NV & AMD will likely update so at worst case, as I have said for the past 4 years, Intel will have a very competitive mid-range, upper mid-range. TSMC N6. $INTC
Intel's gaming DGPU, Arc, has two features that are more similar to NVIDIA's arch than AMD's:
-Discrete ray tracing blocks
-Super-sampling driven by discrete AI blocks (XMX)
$INTC
For HPC & AI, Intel's going straight after NVIDIA's A100 with Ponte Vecchio with XeHPC arch. Claims of 45 TFLOPS FP32 and ResNet train and infer put it on top. Can't assess based on 2 benches, but these two are impressive. 64 RT for visualization interesting. Appears scalable.
Ponte Vecchio is wild:
->100B transistors
-3D Foveros stack PLUS EMIB
-Combo Intel 7 base tile & TSMC N5 compute/link tile
-"will be released in 2022" perf based on A0 silicon
PV is GPU-based, Habana Gaudi is ASIC-based. Intel hedging its bets. $INTC
Now onto Intel Architecture Day CPU cores:
-E core for "efficiency" aka "Gracemont"
-P core for "performance" aka "Golden Cove"
-Hybrid CPU arch mixing E & P in Alder Lake (client) and Sapphire Rapids (DC)
-"Thread Director" to manage uhhh "threads" between P & E
$INTC
Intel "E" Core:
-"efficient" core
-Vs Skylake 1T, either 40% better integer perf at same F -or- 40% lower power at same F
-Vs Skylake 4T, 80% better integer perf at same F -or- 80% lower power at same F
-higher perf than I would have expected for an E core
$INTC
Whoops broke the thread:
Share this Scrolly Tale with your friends.
A Scrolly Tale is a new way to read Twitter threads with a more visually immersive experience.
Discover more beautiful Scrolly Tales like this.