thomas summers

T2 · manager / operator

Founder and CEO of Positron, a chip startup developing inference accelerators using systolic arrays and LPDDR memory; previously shipped FPGA-based Archer product and announced Oracle partnership.

2 calls·2 names·50% bull·last heard 4 months ago·SemiAnalysis
track record

no scored calls yet — needs a stated position or a categorical verdict, with a matured window vs SPY

top calls

highest conviction · one per company
1sthigh conviction
$POSITRONPositronposition

Positron bets on LPDDR memory and 1:1 matrix-vector ratio to slash inference costs

Positron's next-gen Azimov chip uses commodity LPDDR memory to deliver 2.3TB per chip at 400W — 6-8x the memory capacity of Nvidia's B200 at a quarter of the power — while its systolic array achieves a 1:1 matrix-matrix to matrix-vector ratio versus Nvidia's 32:1, enabling far more efficient inference scaling for trillion-parameter models.

2ndmedium conviction
$NVDANvidia

Nvidia's Blackwell worsened matrix-vector ratio to 32:1, neglecting inference efficiency

Nvidia's architectural focus on dense matrix-matrix training workloads caused the matrix-vector performance ratio to degrade from 16:1 on Hopper to 32:1 on Blackwell, even as total flops doubled — a structural disadvantage for inference where attention mechanisms are memory-bound matrix-vector operations.

2 total
$NVDA
···
Nvidia
MEDthomas summers·SemiAnalysis·4 months ago
Nvidia's Blackwell worsened matrix-vector ratio to 32:1, neglecting inference efficiency
Nvidia's architectural focus on dense matrix-matrix training workloads caused the matrix-vector performance ratio to degrade from 16:1 on Hopper to 32:1 on Blackwell, even as total flops doubled — a structural disadvantage for inference where attention mechanisms are memory-bound matrix-vector operations.
"Going from Hopper to Blackwall actually had the ratio of your um matrix matrix, your your gem performance um uh that's when when you go to matrix vector on on Hopper it was a 16 t…"
6:05
$POSITRON
Positron
HIGHthomas summers·SemiAnalysis·4 months ago· position
Positron bets on LPDDR memory and 1:1 matrix-vector ratio to slash inference costs
Positron's next-gen Azimov chip uses commodity LPDDR memory to deliver 2.3TB per chip at 400W — 6-8x the memory capacity of Nvidia's B200 at a quarter of the power — while its systolic array achieves a 1:1 matrix-matrix to matrix-vector ratio versus Nvidia's 32:1, enabling far more efficient inference scaling for trillion-parameter models.
"Yeah, so with Azimov, our our next generation chip, uh we're taking a pretty uh uh counterintuitive or just against the the trend approach of uh actually leveraging LPDDR um as th…"
7:59