{
 "tracks": [
  {
   "id": "t1",
   "label": "Track 1",
   "title": "FPGA drafter + GPU target",
   "color": "#2a78d6"
  },
  {
   "id": "t2",
   "label": "Track 2",
   "title": "HBM streaming",
   "color": "#0f9d8a"
  },
  {
   "id": "t3",
   "label": "Track 3",
   "title": "Constant fold",
   "color": "#8a4fd6"
  }
 ],
 "models": [
  {
   "id": "eagle4-drafter",
   "name": "EAGLE4 drafter (Llama-3.2-3B dims, 1 layer)",
   "track": "t1",
   "cfg": {
    "hidden": 3072,
    "mlp": 8192,
    "layers": 1,
    "heads": 24,
    "kv_heads": 24,
    "head_dim": 128,
    "vocab": 128256,
    "tied": false
   },
   "fmt": "2-bit body, group-128 scales",
   "cards": "HEAD + TAIL pair (box1 62 HX1 BD + 42 TCLK75)",
   "measured": "record 350.16 tok/s mean batch-1 with the H200 target [M F1125]",
   "note": "QKV input is concat(token embedding, previous hidden) = 2 x 3072; the down projection keeps both halves (to_logits, for_reasoning).",
   "posts": [
    "eagle4-weight-residency",
    "box1-drafter-int8act-20261011",
    "eagle4-arch-evolution-20261005",
    "eagle4-restore-arch-page-20260924",
    "eagle4-weekly-arch-20261005"
   ],
   "tensors": [
    {
     "t": "QKV projection",
     "shape": "6,144 x 9,216",
     "k": 6144,
     "n": 9216,
     "layers": 1,
     "bits": 2,
     "fmt": "2-bit, group 128",
     "params": 56623104,
     "bytes": 14155776,
     "where": "URAM, HEAD card",
     "kind": "attn",
     "note": "input = [embedding ; previous hidden]"
    },
    {
     "t": "o-projection",
     "shape": "3,072 x 3,072",
     "k": 3072,
     "n": 3072,
     "layers": 1,
     "bits": 2,
     "fmt": "2-bit",
     "params": 9437184,
     "bytes": 2359296,
     "where": "URAM, HEAD card",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "MLP gate",
     "shape": "3,072 x 8,192",
     "k": 3072,
     "n": 8192,
     "layers": 1,
     "bits": 2,
     "fmt": "2-bit",
     "params": 25165824,
     "bytes": 6291456,
     "where": "URAM, HEAD card",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "MLP up",
     "shape": "3,072 x 8,192",
     "k": 3072,
     "n": 8192,
     "layers": 1,
     "bits": 2,
     "fmt": "2-bit",
     "params": 25165824,
     "bytes": 6291456,
     "where": "URAM, HEAD card",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "MLP down",
     "shape": "8,192 x 6,144",
     "k": 8192,
     "n": 6144,
     "layers": 1,
     "bits": 2,
     "fmt": "2-bit",
     "params": 50331648,
     "bytes": 6291456,
     "where": "URAM, TAIL card",
     "kind": "mlp",
     "note": "two halves: to_logits + for_reasoning"
    },
    {
     "t": "Cheap vocabulary head",
     "shape": "256 x 128,256",
     "k": 256,
     "n": 128256,
     "layers": 1,
     "bits": 2,
     "fmt": "2-bit",
     "params": 32833536,
     "bytes": 8210350,
     "where": "URAM, TAIL card",
     "kind": "head",
     "note": "vocab sweep for the draft tree"
    },
    {
     "t": "Low-rank LM projection (GLM)",
     "shape": "3,072 x 256",
     "k": 3072,
     "n": 256,
     "layers": 1,
     "bits": 16,
     "fmt": "Q1.15",
     "params": 786432,
     "bytes": 1572864,
     "where": "BRAM, TAIL card",
     "kind": "head",
     "note": ""
    },
    {
     "t": "Exact lm-head (rescore)",
     "shape": "3,072 x 128,256",
     "k": 3072,
     "n": 128256,
     "layers": 1,
     "bits": 16,
     "fmt": "fp16",
     "params": 394002432,
     "bytes": 788000000,
     "where": "HBM, TAIL card (2 copies on G8 N=4: 1.58 GB)",
     "kind": "head",
     "note": "gathered for G8 rescore only"
    },
    {
     "t": "Embedding table",
     "shape": "128,256 x 3,072",
     "k": 128256,
     "n": 3072,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 394002432,
     "bytes": 788000000,
     "where": "HBM, HEAD card",
     "kind": "embed",
     "note": "child-branch rows, depth >= 1"
    },
    {
     "t": "Dequantization scales",
     "shape": "one per 128 weights",
     "k": 0,
     "n": 0,
     "layers": 1,
     "bits": 16,
     "fmt": "Q3.13",
     "params": 0,
     "bytes": 1468006,
     "where": "BRAM, both cards",
     "kind": "norm",
     "note": ""
    },
    {
     "t": "RMSNorm gains x 4",
     "shape": "3,072 x 4",
     "k": 3072,
     "n": 4,
     "layers": 1,
     "bits": 16,
     "fmt": "Q4.12",
     "params": 12288,
     "bytes": 24576,
     "where": "BRAM, both cards",
     "kind": "norm",
     "note": ""
    },
    {
     "t": "K cache (per token)",
     "shape": "24 x 128",
     "k": 128,
     "n": 24,
     "layers": 1,
     "bits": 16,
     "fmt": "int16, post-RoPE",
     "params": 3072,
     "bytes": 6144,
     "where": "URAM, HEAD card",
     "kind": "kv",
     "note": "per token"
    },
    {
     "t": "V cache (per token)",
     "shape": "24 x 128",
     "k": 128,
     "n": 24,
     "layers": 1,
     "bits": 16,
     "fmt": "int16",
     "params": 3072,
     "bytes": 6144,
     "where": "BRAM, HEAD card",
     "kind": "kv",
     "note": "per token"
    }
   ],
   "total_params": 988348416,
   "total_bytes": 1622665236
  },
  {
   "id": "qwen27b-drafter-step21491",
   "name": "Qwen3.8-27B drafter step21491 (BitNet SubLN, 1 layer)",
   "track": "t1",
   "cfg": {
    "hidden": 3072,
    "mlp": 8192,
    "layers": 1,
    "heads": 24,
    "kv_heads": 24,
    "head_dim": 128,
    "vocab": 248320,
    "tied": false
   },
   "fmt": "ternary body (deployed as 2-bit), bf16 head + embedding",
   "cards": "4C1: ATT pair 17/3d + MLPV pair 50/63 (talon2)",
   "measured": "H200 SGLang w2d7: accept 3.365, 132.15 tok/s decode [M, OPEN_RELEASE_DRAFTER_EVAL_20261102]",
   "note": "File /mnt/data/models/qwen27b_drafter_sglang/step21491_bf16ste/model.safetensors md5 32f20d0d (21 tensors); body shown at its 2-bit deployed size.",
   "posts": [],
   "tensors": [
    {
     "t": "fc (3-hidden fuse)",
     "shape": "15,360 x 3,072",
     "k": 15360,
     "n": 3072,
     "layers": 1,
     "bits": 2,
     "fmt": "ternary, stored bf16 in file",
     "params": 47185920,
     "bytes": 11796480,
     "where": "not on the 4C1 cards (depth-0 only)",
     "kind": "attn",
     "note": "file: fc.weight bf16 [3072,15360]"
    },
    {
     "t": "QKV projection",
     "shape": "6,144 x 9,216",
     "k": 6144,
     "n": 9216,
     "layers": 1,
     "bits": 2,
     "fmt": "ternary (BitNet SubLN)",
     "params": 56623104,
     "bytes": 14155776,
     "where": "URAM, ATT pair (TP2: 16 heads per card)",
     "kind": "attn",
     "note": "input = [embedding ; previous hidden] = 2 x 3072"
    },
    {
     "t": "o-projection",
     "shape": "3,072 x 3,072",
     "k": 3072,
     "n": 3072,
     "layers": 1,
     "bits": 2,
     "fmt": "ternary",
     "params": 9437184,
     "bytes": 2359296,
     "where": "URAM, ATT pair",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "MLP gate",
     "shape": "3,072 x 8,192",
     "k": 3072,
     "n": 8192,
     "layers": 1,
     "bits": 2,
     "fmt": "ternary",
     "params": 25165824,
     "bytes": 6291456,
     "where": "URAM, MLPV pair",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "MLP up",
     "shape": "3,072 x 8,192",
     "k": 3072,
     "n": 8192,
     "layers": 1,
     "bits": 2,
     "fmt": "ternary",
     "params": 25165824,
     "bytes": 6291456,
     "where": "URAM, MLPV pair",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "MLP down",
     "shape": "8,192 x 3,072",
     "k": 8192,
     "n": 3072,
     "layers": 1,
     "bits": 2,
     "fmt": "ternary",
     "params": 25165824,
     "bytes": 6291456,
     "where": "URAM, MLPV pair",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "lm-head (128,256-token shortlist)",
     "shape": "3,072 x 128,256",
     "k": 3072,
     "n": 128256,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 394002432,
     "bytes": 788004864,
     "where": "see residency page",
     "kind": "head",
     "note": "d2t/t2d map the 128,256 shortlist to the 248,320 vocab"
    },
    {
     "t": "Embedding table",
     "shape": "248,320 x 3,072",
     "k": 248320,
     "n": 3072,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 762839040,
     "bytes": 1525678080,
     "where": "see residency page",
     "kind": "embed",
     "note": ""
    },
    {
     "t": "SubLN norms (attn_sub + ffn_sub)",
     "shape": "3,072 x 1",
     "k": 3072,
     "n": 1,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 11264,
     "bytes": 22528,
     "where": "BRAM",
     "kind": "norm",
     "note": "attn_sub_norm [3072] + ffn_sub_norm [8192]"
    },
    {
     "t": "RMSNorm gains (input, post-attn, hidden, final)",
     "shape": "3,072 x 4",
     "k": 3072,
     "n": 4,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 12288,
     "bytes": 24576,
     "where": "BRAM",
     "kind": "norm",
     "note": ""
    }
   ],
   "total_params": 1345608704,
   "total_bytes": 2360915968
  },
  {
   "id": "qwen27b-drafter-gptq2-lmhead2",
   "name": "Qwen3.8-27B drafter GPTQ v2 W2A16 + 2-bit lm-head (2026-11-02)",
   "track": "t1",
   "cfg": {
    "hidden": 5120,
    "mlp": 16384,
    "layers": 1,
    "heads": 40,
    "kv_heads": 40,
    "head_dim": 128,
    "vocab": 248320,
    "tied": false
   },
   "fmt": "2-bit body AND 2-bit lm-head (GPTQ v2, group 128), bf16 embedding",
   "cards": "not on FPGA yet (D9 LUT-GEMM ATT image 0xE4B713D9 building)",
   "measured": "H200 SGLang w2d7: accept 3.990 (+18.6 %), 187.74 tok/s decode (+42 %); A8 activations -0.6 % per-token / -0.2 % per-128-group [M, OPEN_RELEASE_DRAFTER_EVAL_20261102]",
   "note": "File W2A16/model.safetensors md5 e41d1646 (49 tensors, 3,030,701,984 B); W4A8/ is a byte-identical copy. Sizes are the packed bytes from the safetensors header.",
   "posts": [],
   "tensors": [
    {
     "t": "fc (3-hidden fuse)",
     "shape": "15,360 x 5,120",
     "k": 15360,
     "n": 5120,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128",
     "params": 78643200,
     "bytes": 19660800,
     "where": "not yet placed",
     "kind": "attn",
     "note": "x1.67 vs step21491 (11.80 -> 19.66 MB)"
    },
    {
     "t": "QKV projection",
     "shape": "10,240 x 15,360",
     "k": 10240,
     "n": 15360,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128",
     "params": 157286400,
     "bytes": 39321600,
     "where": "not yet placed (D9 LUT-GEMM ATT image)",
     "kind": "attn",
     "note": "input = [embedding ; previous hidden] = 2 x 5120; x2.78 vs step21491 (14.16 -> 39.32 MB)"
    },
    {
     "t": "o-projection",
     "shape": "5,120 x 5,120",
     "k": 5120,
     "n": 5120,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128",
     "params": 26214400,
     "bytes": 6553600,
     "where": "not yet placed (D9)",
     "kind": "attn",
     "note": "x2.78 (2.36 -> 6.55 MB)"
    },
    {
     "t": "MLP gate",
     "shape": "5,120 x 16,384",
     "k": 5120,
     "n": 16384,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128",
     "params": 83886080,
     "bytes": 20971520,
     "where": "not yet placed",
     "kind": "mlp",
     "note": "x3.33 (6.29 -> 20.97 MB)"
    },
    {
     "t": "MLP up",
     "shape": "5,120 x 16,384",
     "k": 5120,
     "n": 16384,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128",
     "params": 83886080,
     "bytes": 20971520,
     "where": "not yet placed",
     "kind": "mlp",
     "note": "x3.33 (6.29 -> 20.97 MB)"
    },
    {
     "t": "MLP down",
     "shape": "16,384 x 5,120",
     "k": 16384,
     "n": 5120,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128",
     "params": 83886080,
     "bytes": 20971520,
     "where": "not yet placed",
     "kind": "mlp",
     "note": "x3.33 (6.29 -> 20.97 MB)"
    },
    {
     "t": "Cheap vocabulary head (efficient_lm_head)",
     "shape": "128 x 248,320",
     "k": 128,
     "n": 248320,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128",
     "params": 31784960,
     "bytes": 7946240,
     "where": "not yet placed",
     "kind": "head",
     "note": "NEW: no equivalent tensor in step21491"
    },
    {
     "t": "Cheap-head down projection",
     "shape": "5,120 x 128",
     "k": 5120,
     "n": 128,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 655360,
     "bytes": 1310720,
     "where": "not yet placed",
     "kind": "head",
     "note": "efficient_lm_head_down_proj [128,5120]"
    },
    {
     "t": "lm-head (full 248,320 vocab)",
     "shape": "5,120 x 248,320",
     "k": 5120,
     "n": 248320,
     "layers": 1,
     "bits": 2,
     "fmt": "GPTQ v2, group 128 (was bf16)",
     "params": 1271398400,
     "bytes": 317849600,
     "where": "not yet placed",
     "kind": "head",
     "note": "2-bit instead of bf16: 788.0 MB bf16 (128,256 shortlist) -> 317.8 MB 2-bit + 22.3 MB scales/zeros, over the FULL vocab"
    },
    {
     "t": "Embedding table",
     "shape": "248,320 x 5,120",
     "k": 248320,
     "n": 5120,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 1271398400,
     "bytes": 2542796800,
     "where": "not yet placed",
     "kind": "embed",
     "note": "x1.67 (1,525.7 -> 2,542.8 MB)"
    },
    {
     "t": "Dequantization scales + zeros (body + fc)",
     "shape": "0 x 0",
     "k": 0,
     "n": 0,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16 scales, int32-packed zeros",
     "params": 0,
     "bytes": 9030400,
     "where": "not yet placed",
     "kind": "norm",
     "note": "scales 8,028,160 B + qzeros 1,003,520 B (q/k/v/o/gate/up/down/fc); lm-head adds 19,865,600 + 2,483,200 B"
    },
    {
     "t": "RMSNorm gains (input, post-attn, hidden, final, fc_norm x3)",
     "shape": "5,120 x 7",
     "k": 5120,
     "n": 7,
     "layers": 1,
     "bits": 16,
     "fmt": "bf16",
     "params": 35840,
     "bytes": 71680,
     "where": "not yet placed",
     "kind": "norm",
     "note": "no SubLN norms (step21491 had attn_sub/ffn_sub)"
    }
   ],
   "total_params": 3089075200,
   "total_bytes": 3007456000
  },
  {
   "id": "llama-32-1b-instruct",
   "name": "Llama-3.2-1B-Instruct",
   "track": "t2",
   "cfg": {
    "hidden": 2048,
    "mlp": 8192,
    "layers": 16,
    "heads": 32,
    "kv_heads": 8,
    "head_dim": 64,
    "vocab": 128256,
    "tied": true
   },
   "fmt": "W8 per-tensor, int8 activations",
   "cards": "one V80, 0x1254 s2",
   "measured": "156.1 tok/s batch-1 [M]; 1.48 GiB loaded in 1.8 s [M]",
   "tensors": [
    {
     "t": "embed_tokens",
     "shape": "128,256 x 2,048",
     "k": 128256,
     "n": 2048,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 262668288,
     "bytes": 262668288,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "embed",
     "note": "row gather by token id (one-hot GEMV on card)"
    },
    {
     "t": "q_proj",
     "shape": "2,048 x 2,048",
     "k": 2048,
     "n": 2048,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 67108864,
     "bytes": 67108864,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "k_proj",
     "shape": "2,048 x 512",
     "k": 2048,
     "n": 512,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 16777216,
     "bytes": 16777216,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "v_proj",
     "shape": "2,048 x 512",
     "k": 2048,
     "n": 512,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 16777216,
     "bytes": 16777216,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "o_proj",
     "shape": "2,048 x 2,048",
     "k": 2048,
     "n": 2048,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 67108864,
     "bytes": 67108864,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "gate_proj",
     "shape": "2,048 x 8,192",
     "k": 2048,
     "n": 8192,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 268435456,
     "bytes": 268435456,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "up_proj",
     "shape": "2,048 x 8,192",
     "k": 2048,
     "n": 8192,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 268435456,
     "bytes": 268435456,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "down_proj",
     "shape": "8,192 x 2,048",
     "k": 8192,
     "n": 2048,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 268435456,
     "bytes": 268435456,
     "where": "HBM, 32 port windows (talon card 17)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "lm_head (tied to embed_tokens)",
     "shape": "2,048 x 128,256",
     "k": 2048,
     "n": 128256,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 0,
     "bytes": 0,
     "where": "shares embed_tokens storage",
     "kind": "head",
     "note": "tie_word_embeddings = true"
    },
    {
     "t": "RMSNorm gains",
     "shape": "2,048 x 33",
     "k": 2048,
     "n": 33,
     "layers": 1,
     "bits": 16,
     "fmt": "int16 / fixed point",
     "params": 67584,
     "bytes": 135168,
     "where": "CAB / CSR",
     "kind": "norm",
     "note": "input + post-attention per layer, plus final"
    },
    {
     "t": "KV cache (per token)",
     "shape": "2 x 16 x 8 x 64",
     "k": 512,
     "n": 32,
     "layers": 1,
     "bits": 8,
     "fmt": "int8",
     "params": 16384,
     "bytes": 16384,
     "where": "CAB (KV_MODE 0, 27-32 positions) or HBM (KV_MODE 1)",
     "kind": "kv",
     "note": "per generated token"
    }
   ],
   "posts": [
    "talon-serving-llama-20261015",
    "talon-hbms-engine-arch-cycles-20261015",
    "eagle4-v80-platform",
    "ring8-hbm-llm-serving-20261011"
   ],
   "note": "",
   "total_params": 1235746816,
   "total_bytes": 1235881984
  },
  {
   "id": "llama-32-3b-instruct",
   "name": "Llama-3.2-3B-Instruct",
   "track": "t2",
   "cfg": {
    "hidden": 3072,
    "mlp": 8192,
    "layers": 28,
    "heads": 24,
    "kv_heads": 8,
    "head_dim": 128,
    "vocab": 128256,
    "tied": true
   },
   "fmt": "W8 per-tensor, int8 activations",
   "cards": "one V80, 0x1254 s2",
   "measured": "59.4 tok/s batch-1 [M]; 68.2-81.9 with the 1B drafter [M]; 3,598 MiB loaded [M]",
   "tensors": [
    {
     "t": "embed_tokens",
     "shape": "128,256 x 3,072",
     "k": 128256,
     "n": 3072,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 394002432,
     "bytes": 394002432,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "embed",
     "note": "row gather by token id (one-hot GEMV on card)"
    },
    {
     "t": "q_proj",
     "shape": "3,072 x 3,072",
     "k": 3072,
     "n": 3072,
     "layers": 28,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 264241152,
     "bytes": 264241152,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "k_proj",
     "shape": "3,072 x 1,024",
     "k": 3072,
     "n": 1024,
     "layers": 28,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 88080384,
     "bytes": 88080384,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "v_proj",
     "shape": "3,072 x 1,024",
     "k": 3072,
     "n": 1024,
     "layers": 28,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 88080384,
     "bytes": 88080384,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "o_proj",
     "shape": "3,072 x 3,072",
     "k": 3072,
     "n": 3072,
     "layers": 28,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 264241152,
     "bytes": 264241152,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "gate_proj",
     "shape": "3,072 x 8,192",
     "k": 3072,
     "n": 8192,
     "layers": 28,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 704643072,
     "bytes": 704643072,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "up_proj",
     "shape": "3,072 x 8,192",
     "k": 3072,
     "n": 8192,
     "layers": 28,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 704643072,
     "bytes": 704643072,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "down_proj",
     "shape": "8,192 x 3,072",
     "k": 8192,
     "n": 3072,
     "layers": 28,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 704643072,
     "bytes": 704643072,
     "where": "HBM, 32 port windows (talon card 50)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "lm_head (tied to embed_tokens)",
     "shape": "3,072 x 128,256",
     "k": 3072,
     "n": 128256,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 0,
     "bytes": 0,
     "where": "shares embed_tokens storage",
     "kind": "head",
     "note": "tie_word_embeddings = true"
    },
    {
     "t": "RMSNorm gains",
     "shape": "3,072 x 57",
     "k": 3072,
     "n": 57,
     "layers": 1,
     "bits": 16,
     "fmt": "int16 / fixed point",
     "params": 175104,
     "bytes": 350208,
     "where": "CAB / CSR",
     "kind": "norm",
     "note": "input + post-attention per layer, plus final"
    },
    {
     "t": "KV cache (per token)",
     "shape": "2 x 28 x 8 x 128",
     "k": 1024,
     "n": 56,
     "layers": 1,
     "bits": 8,
     "fmt": "int8",
     "params": 57344,
     "bytes": 57344,
     "where": "CAB (KV_MODE 0, 27-32 positions) or HBM (KV_MODE 1)",
     "kind": "kv",
     "note": "per generated token"
    }
   ],
   "posts": [
    "talon-serving-llama-20261015",
    "talon-hbms-engine-arch-cycles-20261015",
    "eagle4-v80-platform",
    "ring8-hbm-llm-serving-20261011"
   ],
   "note": "",
   "total_params": 3212574720,
   "total_bytes": 3212924928
  },
  {
   "id": "llama-31-8b-instruct",
   "name": "Llama-3.1-8B-Instruct",
   "track": "t2",
   "cfg": {
    "hidden": 4096,
    "mlp": 14336,
    "layers": 32,
    "heads": 32,
    "kv_heads": 8,
    "head_dim": 128,
    "vocab": 128256,
    "tied": false
   },
   "fmt": "W8 per-tensor, int8 activations",
   "cards": "one V80, 0x1254 s2",
   "measured": "28.0 tok/s batch-1 [M]; 39.3-51.1 with the 1B drafter [M]; ~7.7 GiB [M]",
   "tensors": [
    {
     "t": "embed_tokens",
     "shape": "128,256 x 4,096",
     "k": 128256,
     "n": 4096,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 525336576,
     "bytes": 525336576,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "embed",
     "note": "row gather by token id (one-hot GEMV on card)"
    },
    {
     "t": "q_proj",
     "shape": "4,096 x 4,096",
     "k": 4096,
     "n": 4096,
     "layers": 32,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 536870912,
     "bytes": 536870912,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "k_proj",
     "shape": "4,096 x 1,024",
     "k": 4096,
     "n": 1024,
     "layers": 32,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 134217728,
     "bytes": 134217728,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "v_proj",
     "shape": "4,096 x 1,024",
     "k": 4096,
     "n": 1024,
     "layers": 32,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 134217728,
     "bytes": 134217728,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "o_proj",
     "shape": "4,096 x 4,096",
     "k": 4096,
     "n": 4096,
     "layers": 32,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 536870912,
     "bytes": 536870912,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "gate_proj",
     "shape": "4,096 x 14,336",
     "k": 4096,
     "n": 14336,
     "layers": 32,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 1879048192,
     "bytes": 1879048192,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "up_proj",
     "shape": "4,096 x 14,336",
     "k": 4096,
     "n": 14336,
     "layers": 32,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 1879048192,
     "bytes": 1879048192,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "down_proj",
     "shape": "14,336 x 4,096",
     "k": 14336,
     "n": 4096,
     "layers": 32,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 1879048192,
     "bytes": 1879048192,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "lm_head",
     "shape": "4,096 x 128,256",
     "k": 4096,
     "n": 128256,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 525336576,
     "bytes": 525336576,
     "where": "HBM, 32 port windows (talon card 63)",
     "kind": "head",
     "note": ""
    },
    {
     "t": "RMSNorm gains",
     "shape": "4,096 x 65",
     "k": 4096,
     "n": 65,
     "layers": 1,
     "bits": 16,
     "fmt": "int16 / fixed point",
     "params": 266240,
     "bytes": 532480,
     "where": "CAB / CSR",
     "kind": "norm",
     "note": "input + post-attention per layer, plus final"
    },
    {
     "t": "KV cache (per token)",
     "shape": "2 x 32 x 8 x 128",
     "k": 1024,
     "n": 64,
     "layers": 1,
     "bits": 8,
     "fmt": "int8",
     "params": 65536,
     "bytes": 65536,
     "where": "CAB (KV_MODE 0, 27-32 positions) or HBM (KV_MODE 1)",
     "kind": "kv",
     "note": "per generated token"
    }
   ],
   "posts": [
    "talon-serving-llama-20261015",
    "talon-hbms-engine-arch-cycles-20261015",
    "eagle4-v80-platform",
    "ring8-hbm-llm-serving-20261011"
   ],
   "note": "",
   "total_params": 8029995008,
   "total_bytes": 8030527488
  },
  {
   "id": "llama-33-70b-instruct",
   "name": "Llama-3.3-70B-Instruct",
   "track": "t2",
   "cfg": {
    "hidden": 8192,
    "mlp": 28672,
    "layers": 80,
    "heads": 64,
    "kv_heads": 8,
    "head_dim": 128,
    "vocab": 128256,
    "tied": false
   },
   "fmt": "W8 per-tensor, int8 activations",
   "cards": "8 V80s pipeline-parallel, 0x1254 s2",
   "measured": "2.34-2.44 tok/s batch-1, bit-exact 16/16 [M]; 7.98-8.96 GiB per card [M]",
   "tensors": [
    {
     "t": "embed_tokens",
     "shape": "128,256 x 8,192",
     "k": 128256,
     "n": 8192,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 1050673152,
     "bytes": 1050673152,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "embed",
     "note": "row gather by token id (one-hot GEMV on card)"
    },
    {
     "t": "q_proj",
     "shape": "8,192 x 8,192",
     "k": 8192,
     "n": 8192,
     "layers": 80,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 5368709120,
     "bytes": 5368709120,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "k_proj",
     "shape": "8,192 x 1,024",
     "k": 8192,
     "n": 1024,
     "layers": 80,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 671088640,
     "bytes": 671088640,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "v_proj",
     "shape": "8,192 x 1,024",
     "k": 8192,
     "n": 1024,
     "layers": 80,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 671088640,
     "bytes": 671088640,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "o_proj",
     "shape": "8,192 x 8,192",
     "k": 8192,
     "n": 8192,
     "layers": 80,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 5368709120,
     "bytes": 5368709120,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "gate_proj",
     "shape": "8,192 x 28,672",
     "k": 8192,
     "n": 28672,
     "layers": 80,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 18790481920,
     "bytes": 18790481920,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "up_proj",
     "shape": "8,192 x 28,672",
     "k": 8192,
     "n": 28672,
     "layers": 80,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 18790481920,
     "bytes": 18790481920,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "down_proj",
     "shape": "28,672 x 8,192",
     "k": 28672,
     "n": 8192,
     "layers": 80,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 18790481920,
     "bytes": 18790481920,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "lm_head",
     "shape": "8,192 x 128,256",
     "k": 8192,
     "n": 128256,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 1050673152,
     "bytes": 1050673152,
     "where": "HBM, 10 layers per card across 8 cards",
     "kind": "head",
     "note": ""
    },
    {
     "t": "RMSNorm gains",
     "shape": "8,192 x 161",
     "k": 8192,
     "n": 161,
     "layers": 1,
     "bits": 16,
     "fmt": "int16 / fixed point",
     "params": 1318912,
     "bytes": 2637824,
     "where": "CAB / CSR",
     "kind": "norm",
     "note": "input + post-attention per layer, plus final"
    },
    {
     "t": "KV cache (per token)",
     "shape": "2 x 80 x 8 x 128",
     "k": 1024,
     "n": 160,
     "layers": 1,
     "bits": 8,
     "fmt": "int8",
     "params": 163840,
     "bytes": 163840,
     "where": "CAB (KV_MODE 0, 27-32 positions) or HBM (KV_MODE 1)",
     "kind": "kv",
     "note": "per generated token"
    }
   ],
   "posts": [
    "talon-serving-llama-20261015",
    "talon-hbms-engine-arch-cycles-20261015",
    "eagle4-v80-platform",
    "ring8-hbm-llm-serving-20261011"
   ],
   "note": "",
   "total_params": 70552387584,
   "total_bytes": 70555025408
  },
  {
   "id": "trilm-99m-(hbms,-w8)",
   "name": "TriLM-99M (HBMS, W8)",
   "track": "t2",
   "cfg": {
    "hidden": 512,
    "mlp": 1280,
    "layers": 16,
    "heads": 8,
    "kv_heads": 8,
    "head_dim": 64,
    "vocab": 50304,
    "tied": false
   },
   "fmt": "W8 per-tensor, int8 activations",
   "cards": "one V80, 0x1254 s2",
   "measured": "1,011-1,025 tok/s batch-1 [M]; switch to 1B in 3.1-3.9 ms [M]",
   "tensors": [
    {
     "t": "embed_tokens",
     "shape": "50,304 x 512",
     "k": 50304,
     "n": 512,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 25755648,
     "bytes": 25755648,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "embed",
     "note": "row gather by token id (one-hot GEMV on card)"
    },
    {
     "t": "q_proj",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 4194304,
     "bytes": 4194304,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "k_proj",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 4194304,
     "bytes": 4194304,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "v_proj",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 4194304,
     "bytes": 4194304,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "o_proj",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 4194304,
     "bytes": 4194304,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "attn",
     "note": ""
    },
    {
     "t": "gate_proj",
     "shape": "512 x 1,280",
     "k": 512,
     "n": 1280,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 10485760,
     "bytes": 10485760,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "up_proj",
     "shape": "512 x 1,280",
     "k": 512,
     "n": 1280,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 10485760,
     "bytes": 10485760,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "down_proj",
     "shape": "1,280 x 512",
     "k": 1280,
     "n": 512,
     "layers": 16,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 10485760,
     "bytes": 10485760,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "mlp",
     "note": ""
    },
    {
     "t": "lm_head",
     "shape": "512 x 50,304",
     "k": 512,
     "n": 50304,
     "layers": 1,
     "bits": 8,
     "fmt": "W8 per-tensor, int8 activations",
     "params": 25755648,
     "bytes": 25755648,
     "where": "HBM (talon card 17, co-resident with 1B)",
     "kind": "head",
     "note": ""
    },
    {
     "t": "RMSNorm gains",
     "shape": "512 x 33",
     "k": 512,
     "n": 33,
     "layers": 1,
     "bits": 16,
     "fmt": "int16 / fixed point",
     "params": 16896,
     "bytes": 33792,
     "where": "CAB / CSR",
     "kind": "norm",
     "note": "input + post-attention per layer, plus final"
    },
    {
     "t": "KV cache (per token)",
     "shape": "2 x 16 x 8 x 64",
     "k": 512,
     "n": 32,
     "layers": 1,
     "bits": 8,
     "fmt": "int8",
     "params": 16384,
     "bytes": 16384,
     "where": "CAB (KV_MODE 0, 27-32 positions) or HBM (KV_MODE 1)",
     "kind": "kv",
     "note": "per generated token"
    }
   ],
   "posts": [
    "talon-serving-llama-20261015",
    "talon-hbms-engine-arch-cycles-20261015",
    "eagle4-v80-platform",
    "ring8-hbm-llm-serving-20261011"
   ],
   "note": "",
   "total_params": 99745792,
   "total_bytes": 99779584
  },
  {
   "id": "trilm-fold",
   "name": "TriLM-99M constant fold (TUNIT / FOLDRING8)",
   "track": "t3",
   "cfg": {
    "hidden": 512,
    "mlp": 1280,
    "layers": 16,
    "heads": 8,
    "kv_heads": 8,
    "head_dim": 64,
    "vocab": 50304,
    "tied": false
   },
   "fmt": "ternary {-1, 0, +1} as LUT constants, one fp16 scale per matrix",
   "cards": "TUNIT: one V80 (layer 0); FOLDRING8: 8 V80s, weight-resident",
   "measured": "TUNIT 0 / 10,240 mismatches, 266 cycles per projection [M]; FOLDRING8 ~4,550 tok/s [D, xsim]",
   "note": "All 112 projections (16 layers x 7) are exactly ternary; no weight memory traffic once folded.",
   "posts": [
    "eagle4-tunit-constant-fold",
    "ternary-make-possible-20261011",
    "ternary-constfold-20261011",
    "box1-ternary-fold-8card-20261011"
   ],
   "tensors": [
    {
     "t": "q_proj (per layer)",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 2,
     "fmt": "ternary, folded into LUTRAM",
     "params": 4194304,
     "bytes": 1048576,
     "where": "LUT logic (TUNIT: layer 0 only; FOLDRING8: 2 layers per card)",
     "kind": "attn",
     "note": "1.58 bit information, stored as LUT INIT"
    },
    {
     "t": "k_proj (per layer)",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 2,
     "fmt": "ternary, folded into LUTRAM",
     "params": 4194304,
     "bytes": 1048576,
     "where": "LUT logic (TUNIT: layer 0 only; FOLDRING8: 2 layers per card)",
     "kind": "attn",
     "note": "1.58 bit information, stored as LUT INIT"
    },
    {
     "t": "v_proj (per layer)",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 2,
     "fmt": "ternary, folded into LUTRAM",
     "params": 4194304,
     "bytes": 1048576,
     "where": "LUT logic (TUNIT: layer 0 only; FOLDRING8: 2 layers per card)",
     "kind": "attn",
     "note": "1.58 bit information, stored as LUT INIT"
    },
    {
     "t": "o_proj (per layer)",
     "shape": "512 x 512",
     "k": 512,
     "n": 512,
     "layers": 16,
     "bits": 2,
     "fmt": "ternary, folded into LUTRAM",
     "params": 4194304,
     "bytes": 1048576,
     "where": "LUT logic (TUNIT: layer 0 only; FOLDRING8: 2 layers per card)",
     "kind": "attn",
     "note": "1.58 bit information, stored as LUT INIT"
    },
    {
     "t": "gate_proj (per layer)",
     "shape": "512 x 1,280",
     "k": 512,
     "n": 1280,
     "layers": 16,
     "bits": 2,
     "fmt": "ternary, folded into LUTRAM",
     "params": 10485760,
     "bytes": 2621440,
     "where": "LUT logic (TUNIT: layer 0 only; FOLDRING8: 2 layers per card)",
     "kind": "mlp",
     "note": "1.58 bit information, stored as LUT INIT"
    },
    {
     "t": "up_proj (per layer)",
     "shape": "512 x 1,280",
     "k": 512,
     "n": 1280,
     "layers": 16,
     "bits": 2,
     "fmt": "ternary, folded into LUTRAM",
     "params": 10485760,
     "bytes": 2621440,
     "where": "LUT logic (TUNIT: layer 0 only; FOLDRING8: 2 layers per card)",
     "kind": "mlp",
     "note": "1.58 bit information, stored as LUT INIT"
    },
    {
     "t": "down_proj (per layer)",
     "shape": "1,280 x 512",
     "k": 1280,
     "n": 512,
     "layers": 16,
     "bits": 2,
     "fmt": "ternary, folded into LUTRAM",
     "params": 10485760,
     "bytes": 2621440,
     "where": "LUT logic (TUNIT: layer 0 only; FOLDRING8: 2 layers per card)",
     "kind": "mlp",
     "note": "1.58 bit information, stored as LUT INIT"
    },
    {
     "t": "embed_tokens / lm_head",
     "shape": "50,304 x 512",
     "k": 50304,
     "n": 512,
     "layers": 1,
     "bits": 16,
     "fmt": "fp16 on host",
     "params": 51511296,
     "bytes": 103022592,
     "where": "host (embedding, norms, attention and head run on the host in TUNIT)",
     "kind": "head",
     "note": ""
    }
   ],
   "total_params": 99745792,
   "total_bytes": 115081216
  }
 ]
}