Radeon Instinct MI6

AMDdatacenter
Overview
ArchitecturePolaris (GCN 4) (2017)
Launch dateJune 2017
GPU target (gfx)gfx803
Memory
VRAM16 GB GDDR5
Memory bandwidth224 GB/s
ECCNo
Compute — vector
FP640.358 TFLOPS
FP325.73 TFLOPS
FP165.73 TFLOPS
Cores & clocks
Compute Units36
Shader cores2,304
Board & system
TDP150 W
Form factorFull-Height, Single-Slot
CoolingPassive
PCIePCIe 3.0 x16

Compatible ROCm versions

none — no ROCm release has ever listed gfx803 as a supported GPU target (gfx). ROCm support is granted per target, so this card is not covered by any version, including the current one. The unofficial HSA_OVERRIDE_GFX_VERSION workaround is the usual route people take — see ROCm on consumer Radeon GPUs for what it does and where it breaks.

Added 2026-09 as the entry-level member of AMD's original three-card Radeon Instinct launch lineup (alongside MI8 and this collection's existing MI25), predating this collection's other Radeon Instinct entries by nothing in calendar time (all three launched June 2017) but representing an older, lower-tier architecture — Polaris (GCN 4th Gen) rather than MI25's Vega10 (GCN 5.0). All figures read directly from AMD's own live, current "Accelerator Specifications" comparison tool (amd.com/en/products/specifications/accelerators.html), which still carries structured spec data for this decade-old card; extracted from the page's embedded JSON data model (not every field AMD's schema defines is populated for this old a product — e.g. Compute Units and Matrix Cores are present as columns but empty for MI6, so neither was taken from that source — Compute Units was later derived from the published shader-core count, see below, and Matrix Cores stays empty). fp16TFLOPSVector (5.73) is numerically identical to fp32TFLOPS (5.73) in AMD's own table — Polaris predates Rapid Packed Math (introduced with Vega/GCN 5 the following generation), so there's no 2x FP16 speedup and no separate matrix/tensor throughput field applies (Polaris has no Matrix Cores at all, introduced with CDNA/MI100 years later). eccSupport is "no" per AMD's own table (Memory ECC Support: No) — this cost-optimized entry-tier card uses plain GDDR5 without the ECC that MI8's HBM and MI25/MI50's HBM2 carry. targetId (gfx803) is shared with MI8 (both Polaris10 and Fiji use the same gfx803 LLVM/ROCm target ISA, per LLVM's AMDGPU backend docs) but doesn't cross-link to any `rocm` entry in this collection — every ROCm version tracked here (oldest: 5.0.0) already dropped gfx803 support, so the empty ROCm-compatibility section is accurate, matching this round's K80/MI25 precedent. No MxGPU/multi-instance field — not listed for this generation (introduced starting with MI25/MI50). No transistor count, process node, or MSRP — AMD doesn't publish Instinct list prices (see this collection's other Instinct entries). 2026-09-15 — added computeUnitCount (36 Compute Units). AMD's own specifications tool leaves its Compute Units column empty for this card (noted above), but 36 is not a guess: GCN fixes a Compute Unit at 64 shader cores, so the 2,304 shaderCoreCount AMD does publish gives 2,304 / 64 = 36. It is the same CU x 64 relationship this collection's MI210 entry already documents, applied in the opposite direction. matrixCoreCount stays empty — Polaris has no Matrix Cores at all.

← All GPUs · Compare GPUs →