Congrats @AMD and @UntetherAI on debut @MLPerf results! AMD MI300X matches H100's inference performance, while Untether can beat H200's power efficiency by 3x. Featuring also: @nvidia Blackwell, @intel Granite Rapids and @Google Trillium. My analysis:
The Japanese government will send engineers to be trained in RISC-V and AI at @tenstorrent over the next 5 years. Tenstorrent gets free engineering talent, revenue from the Japanese government, and hopefully, good relationships in the engineers' companies:
OK, I'll bite. Yes, making a transformer chip does mean making a large matmul engine. But there is a lot more to efficient transformer inference than GEMM/FLOPS. You need to be able to feed and use those FLOPS efficiently. (1/x)
Although "baking transformers in silicon" may sound cool, it is mostly just a marketing slogan.
90% of Transformers FLOPS are just GEMMs and 256x256/128x128 systolic arrays in TPU/Trainium are already optimized for GEMMs. Even modern GPGPU (with Tensor cores) are optimized well
.@BrainChip_inc is now offering its #neuromorphic#ai chips in an edge box suitable for security camera systems. Why are edge boxes becoming an increasingly popular path to market for edge AI chip makers? And what is BrainChip's strategy for edge boxes?
I’m Paris with the family this week…but had to tweet this part of the M4 announcement out: apple.com/newsroom/2024/… "The Neural Engine in M4 is Apple’s most capable yet, and is more powerful than any neural processing unit in any AI PC today."
It’s 38 TOPS.
The @Snapdragon X
Neuromorphic IP vendor @BrainChip_inc now has dev kits available. The company converts CNNs to spiking networks so they can run on BrainChip's Akida processor, which reduces power consumption and enables on-chip learning at the edge.
#ai#neuromorphic
