AMD Instinct MI300X
Compare AMD Instinct MI300X against another machine in the Arena
Reference: AMD Instinct on Wikipedia
- Class
- GPU
- Memory
- 192 GB HBM3
- Runs (Q4_K_M)
- —
- Bandwidth
- 5300 GB/s
- TDP
- 750 W
- Released
- 2023-12-06
- Price (US)
- $22298 used as of 2026-08
Sourcing and disambiguation notes
OAM module requiring a dedicated 8-GPU baseboard (e.g. Lenovo/Supermicro universal baseboard server), not a PCIe slot card -- form_factor: server to match how the published SXM4 A100 is categorized. AMD never published an MSRP; leaked/reported per-unit prices ($10k-$15k) vary too much between sources to record as a fact, so price is omitted. 2026-08 METHODOLOGY UPGRADE: the previous $20,000 came from GPUDojo aggregating a SINGLE eBay listing, and was flagged here as "a single listing is not a market". It is replaced by GPU Poet's tracker, which averages the three lowest-priced listings across major marketplaces daily. Broker quotes for MI300X-class hardware remain wildly inconsistent (gpucost.org showed 9 active listings spanning $21,333-$448,000, the top end clearly bundled multi-GPU systems), so treat this as directional, not a precise market price. No new price exists (enterprise-quote only). 2026-09 update: a price-refresh panel re-read the same July page and proposed $21,333, quoting the page's "starting at $21,333" deal-teaser line -- verified directly against the page and rejected: that figure is an undated single-cheapest-listing snapshot from AFTER July ("since July 2026"), not the July 2026 monthly average the record cites. The July page's own dated section still reads $24,554 unchanged. Replaced instead with August 2026's own dated section (same lowest-average-of-three methodology): $22,298, range $21,333-$25,650 -- a genuine ~9% drop from July, not the panel's misattributed figure.
Specification sources
- https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html — vendor, 2026-08-22
- https://www.amd.com/en/newsroom/press-releases/2023-12-6-amd-delivers-leadership-portfolio-of-data-center-a.html — press, 2023-12-06
Measurements
Decode is token generation — the speed you feel while an answer streams. Prefill is prompt processing — the wait before it starts. Why bandwidth predicts decode speed.
No records yet for this hardware. Know of a published benchmark on it? Submit the link.
What verified, single-source and estimated mean, and the same rows with every filter and sort in the benchmarks explorer.
Speed over time
What it can run
Fit is arithmetic, not a measurement — how it is computed. Against 192 GB; models that fit are listed largest first.
No modeled quant fits in 192 GB.