{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreig2mp5kp4wwo3h7f3kor3saycztcvm3pl6nc4qj2dl4jqdz3h6z3i",
"uri": "at://did:plc:ax37ypwwfpozwkc552d7scv5/app.bsky.feed.post/3mgwn332b3dv2"
},
"path": "/threads/2026-gpu-accelerator-table-give-me-some-feedback.2634025/",
"publishedAt": "2026-03-13T09:36:04.000Z",
"site": "https://forums.anandtech.com",
"tags": [
"Graphics Cards",
"https://chipsandcheese.com/p/amds-cdna-4-architecture-announcement"
],
"textContent": " R200=Rubin 1250=MI455X\nIn the case of Rubin, it appears that adaptive compression(achieved 50) has replaced 2:4 structured sparsity.\nI am uncertain whether the value is 25 or 35. On the other hand, I've put together a table where the MI455X is listed with the same value.\n SemiAnalysis stated that FP4 is 32,768, but it seems highly likely that this is a dense value. (https://chipsandcheese.com/p/amds-cdna-4-architecture-announcement) The matrix value seems to be based on sparsity?\n\nSo, to conclude, how accurate would you say the table above is? (25 or 35 etc)",
"title": "2026 GPU accelerator table, give me some feedback"
}