Tenstorrent Grayskull e75/e150 specifications page - cores, SRAM, LPDDR4, bandwidth, TFLOPS, TBP, and the software-support-discontinued notice

FIRST-PARTY HARDWARE SPEC PAGE. Read in full 2026-10-05. The Grayskull record this supports did not previously exist, although engines/niche-tt-metal.json names 'tenstorrent-grayskull' as a supported backend - so the name was in the corpus with nothing behind it. VERBATIM: 'Tensix Core Count: 120 / SRAM: 120 MB (1MB per Tensix Core) / Memory: 8 GB LPDDR4, 256-bit memory bus'. e75/e150 comparison table, VERBATIM: Specification | e75 | e150 -- Part Number TC-01001 | TC-01002; Tensix Cores 96 | 120; AI Clock 1 GHz | 1.2 GHz; SRAM 96 MB | 120 MB; Memory 8 GB LPDDR4 | 8 GB LPDDR4; Memory Speed 3.2 GT/sec | 3.7 GT/sec; Memory Bandwidth 102 GB/sec | 118 GB/sec; TeraFLOPS (FP8) 221 | 332; TeraFLOPS (FP16) 55 | 83; TeraFLOPS (BLOCKFP8) 55 | 83; TBP (Total Board Power) 75W | 200W; System Interface PCI Express 4.0 x16 | PCI Express 4.0 x16. DATA PRECISION FORMATS, VERBATIM: Floating point FP8, FP16, BFLOAT16; Block floating point BLOCKFP2, BLOCKFP4, BLOCKFP8; Vector VTF19. THE MOST IMPORTANT SENTENCE ON THE PAGE, and the reason this record's status is deprecated rather than verified: 'NOTE: Software support for Grayskull has been discontinued. The last supported versions of Tenstorrent's software for Grayskull are as follows: TT-System-Firmware: fw_pack-80.14.0.0.fwbundle, TT-KMD: ttkmd_1.31, TT-Metalium: v0.55.' So the silicon is documented and the software is end-of-life. A record that carried this part as a live option would be wrong in a way the spec table alone cannot show. THE SRAM NUMBER IS THE SAME FALSE-FRIEND SHAPE AS THE WORMHOLE RECORD, and is recorded here rather than glossed: 120 MB is on-die SRAM acting as an L2-like feed tier, and the page's own 1 MB per Tensix Core arithmetic (120 cores x 1 MB = 120 MB) proves it is on-chip, not a memory tier. It must not be filed in vram_gb. The 8 GB LPDDR4 with its 256-bit bus is the part's actual memory tier. NEGATIVE FINDINGS: no process node is published anywhere on this page, and the 8 GB / 256-bit / 118 GB/s trio is internally consistent for LPDDR4 at 3.7 GT/sec (256 x 3.7 / 8 = 118.4), so no figure had to be derived.