819 GB/s
P(T) = F(H, R, K, L, E, X ; T)
00 ·
01 ·
≈26–28
10¹⁴
10¹⁵
02 · Toy compute
| ≈8.3×10¹¹ | |||
| ≈2.3×10¹² | |||
| Marmoset | ≈7.4×10¹² | ||
| M3 Ultra | ≈26–28 TFLOP/s FP16 | 2.6–2.8×10¹³ | |
| Macaca | ≈7.4×10¹³ | ||
| H. sapiens | 10¹⁵ |
03 · Hominins
| A. afarensis | ≈446 ml | ≈3.4–3.8×10¹⁴ | |
| H. habilis | ≈609 ml | ≈4.6–5.0×10¹⁴ | |
| ≈825 ml | ≈6.2–6.6×10¹⁴ | ||
| ≈959 ml | ≈7.2–7.5×10¹⁴ | ||
| H. heidelbergensis s.l. | ≈1227 ml | ≈9.2–9.3×10¹⁴ | |
| ≈1415 ml | ≈1.06×10¹⁵ | ||
| ≈1499 ml | ≈1.11–1.13×10¹⁵ | ||
| ≈1330 ml | 10¹⁵ |
04 ·
R
K
L
E
Embodiment
X
04B ·
48
≈300
91,67%
100%
≈202 GiB
| Qwen3 235B MoE | |
| DeepSeek-R1 70B | |
| Qwen3.6 35B-A3B | |
| DSV4-Flash | |
| R1 671B | |
| C | I | U | T | ||
|---|---|---|---|---|---|
| R1 70B | 12 | 5 | 3 | 4 | 20/24 = 83.3% |
| R1 671B | 13 | 2 | 0 | 9 | 15/24 = 62.5% |
05 ·
LLMoffline training → inference
06 ·
H. erectus
H. erectus
H. erectus
P(T) = F(H, R, K, L, E, X ; T)
07 ·
- Qwen3 235B-A22B · ≈22B
- DeepSeek-R1 Distill 70B · dense
- Qwen3.6 35B-A3B ·
- DSV4-Flash · Long-context / Diversity / Fallback
- DeepSeek-R1-0528 671B ·
- · ≈40B
→
- health probes
→
- Neo4j Knowledge Graph
- SageMath · PARI/GP · Macaulay2
- Lean 4 · Julia
- SearXNG /
≈2 ×
≈44
GFLOP/token*
Qwen3-235B-A22B
≈141
GFLOP/token*
DeepSeek-R1 Distill 70.6B dense
≈80
GFLOP/token*
GLM-5 family · ≈40B active
Inference ≠ Training
08 ·
01
Raw compute
02
03
04
Task first
09 ·
10 ·