Instinct MI210 Price & Used Prices: Which Local LLMs Can It Run?

• Memory64 GB · HBM2e
• Bandwidth1,638 GB/s
• FP16 / BF16181 TFLOPS
• INT8181 TOPS

Bus width

4,096-bit

Architecture

CDNA 2 (Aldebaran)

Cores

6,656

Tensor cores

416

Ecosystem

ROCm

Interface

PCIe 4.0 x16

Power

300 W

Power connector

8-pin EPS

Slots

2

Length

267 mm

Type

Data center

Launch date

Mar 22, 2022

Model	Size	Cards	Total price	Total VRAM	Speed t/s	Total power
Qwen3.8 Flash Next176B (6B active)	111.3 GB	2 cards	$9,988	128 GB	124.5	600 W
DeepSeek V4 Flash284B (13B active)	155.1 GB	4 cards	$19,976	256 GB	95.9	1,200 W
Hy3295B (21B active)	182.2 GB	4 cards	$19,976	256 GB	52.1	1,200 W
GLM 5.3 Flash320B (18B active)	199.7 GB	4 cards	$19,976	256 GB	72.5	1,200 W
GLM 5.3744B (40B active)	467.3 GB	8 cards	$39,952	512 GB	38.7	2,400 W
Hy4 preview770B (49B active)	467.3 GB	8 cards	$39,952	512 GB	33.9	2,400 W
DeepSeek V4.1 Flash763B (8 / 16B active)	510.3 GB	8 cardsEngram in RAM	$39,952	512 GB	64	2,400 W

Rows marked “Engram in RAM” need about 203 GB of extra system RAM. Full calculation rules are on the Methodology page.

1. 1

   Power is EPS 8-pin

   The MI210 has a single 8-pin power connector, but it uses the EPS pinout for CPU power, not the PCIe 8-pin pinout graphics cards normally use, and no power cable comes with the card. With a desktop power supply, you can use a spare CPU 8-pin cable or an adapter clearly labeled “PCIe to EPS”. Check the pinout before plugging in, and don't force a graphics card's PCIe 8-pin cable into it.

2. 2

   Designed only for server airflow

   The MI210 has no fan and is designed for strong front-to-back server airflow. One or two can make do in a tower case with a shroud and fans; if you plan to run four with Infinity Fabric bridges, we recommend going straight to a GPU server that supports passive cards.

3. 3

   The most complete software support among used Instinct cards

   Compared with the MI50 and MI100, the MI210 has the fullest software support: it's on the ROCm 10.0 support list (Linux only: Ubuntu, RHEL, SLES), vLLM's official documentation lists the MI200 series, and AMD's own llama.cpp containers list the MI210 as a validated card. Two caveats: it's not on the list of chips validated for vLLM in the ROCm 10.0 compatibility table, so go by vLLM's own documentation on whether it works; and AWQ and GPTQ don't work on AMD cards, so pick original BF16 weights or GGUF.

Spec	Instinct MI210	Instinct MI100	A100 40GB PCIe	CMP 170HX 64GB
Memory	64 GB	32 GB50% less	40 GB38% less	64 GBSame
Bandwidth	1,638 GB/s	1,229 GB/s25% lower	1,555 GB/s5% lower	1,493 GB/s9% lower
BF16 / FP16 compute	181 TFLOPS	92 TFLOPS49% lower	312 TFLOPS72% higher	202 TFLOPS12% higher
INT8 compute	181 TOPS	92 TOPS49% lower	624 TOPS245% higher	404 TOPS123% higher
Qwen3.8 27B speed (2-bit)	106.2 t/s	82.8 t/s22% slower	101.6 t/s4% slower	98.1 t/s8% slower
Largest dense model (4-bit)	107B	51B52% less	65B39% less	107BSame
Power	300 W	300 WSame	250 W17% lower	250 W17% lower
GPU interconnect	—	—	NVLink 600 GB/s	Not supported
Interface	PCIe 4.0 x16	PCIe 4.0 x16	PCIe 4.0 x16	PCIe 2.0 x4
Released	2022-03	2020-11	2020-06	2021-06
eBay median	$4,994Buy Instinct MI210	$1,503Buy Instinct MI100	$7,750Buy A100 40GB PCIe	$3,000Buy CMP 170HX 64GB

As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.