Thread by @IanCutress

Live at #isscc is @AMD @LisaSu with the plenary talk

@AMD @LisaSu Mainstream Server Performance - 2P SpecInt Rate. Includes AMD, Intel, Arm. Doubling every 2.4yrs.

@AMD @LisaSu GPU single precision FP32 GF

@AMD @LisaSu Top supercomputers. Zettascale by 2034 ish on current trends

@AMD @LisaSu CPUs are 2x perf every 2.4 years
Supercomputers 2x perf every 1.2 years
Systems aren't small, systems aren't cheap!

@AMD @LisaSu Time to train LLMs

@AMD @LisaSu Performance is all well and good, but efficiency matters more. Efficiency isn't increasing at the same pace. It's a harder problem to solve

@AMD @LisaSu Given constant efficiency gain, Zettascale will need 500MW in 2034. Not going to be possible

@AMD @LisaSu That's scaling from 52 GF/watt today to 2140 GF/watt in that time. It means in the next decade, efficiency matters more than perf to continue scaling

@AMD @LisaSu Process technology continues to be important, harder to get density and efficiency perf

@AMD @LisaSu IO doesn't scale. Energy per bit is reducing, but not when normalised to channel quality

@AMD @LisaSu Memory power also going up substantially due to bandwidth and capacity requirements of the new workloads

@AMD @LisaSu AMD focuses on efficiency through advanced architecture. Using the right technology for the right node. Heterogeneous compute with the right solution

@AMD @LisaSu Exploit arch innovations, packaging advances, and silicon advances. MI250 systems power 4 out of top 5 Green500 spots. 2.5D integration with HBM, blended HPC and AI perf

@AMD @LisaSu 2.5D and 3D packaging innovation

@AMD @LisaSu 3D integration technology still maturing for high volume. Logic on logic would be a massive breakthrough. Reducing the cost of communication

@AMD @LisaSu Also, use the right math for the right operation for efficient compute

@AMD @LisaSu Next gen GPU accelerator
HPC and AI
MI300
CPU+GPU in one package
5nm, 3D Stacking
Quantized format support
Unified memory APU architecture

@AMD @LisaSu Cache and Fabric on the bottom
CPU and GPU execution units on top

!!!!

@AMD @LisaSu CDNA3 Unified Memory APU Architecture

@AMD @LisaSu Next set of innovations: Compute and Memory
TSMC already showcased 12-high hybrid bond

@AMD @LisaSu Samsung with processing-in-memory HBM (PIM HBM2). Some Xilinx FPGAs support this

@AMD @LisaSu Reduce access energy by 85% using PIM

@AMD @LisaSu IO efficiency as well, s

@AMD @LisaSu Long reach IO, AMD has transceiver paper at conference.

@AMD @LisaSu "The package is the new motherboard" - Dr. Lisa Su. @AMD #ISSCC2023

@AMD @LisaSu How to reach zettascale. Need 10000GF/watt

@AMD @LisaSu 'Still a lot of technical innovation to go'

@AMD @LisaSu 'Need to leverage AI more holistically, not just menial tasks.'

Basically use AI to get the best power/perf/area/cost on chips. Insert EDA vendors

@AMD @LisaSu AI accelerated HPC. Lots of work to be done in the algorithms. @FelixCLC_ this is your bag

@AMD @LisaSu @FelixCLC_ Reduce Zettascale today (21GW) down to 100MW

@AMD @LisaSu @FelixCLC_ 'Coming to ISSCC is like coming to speak to the smartest people on the planet'

What about Hot Chips Lisa 🥹😢

@AMD @LisaSu @FelixCLC_ Innovate in every dimension

@AMD @LisaSu @FelixCLC_ 5-10yr vision on semiconductor R&D for US needs all disciplines to work together, no silver bullet, lots of work on arch, materials, process, system innovation. Need to bring all together to solve large challenges. Brings out the best in all.

@AMD @LisaSu That's a wrap of Lisa's talk!