A100 80GB PCIe Price, VRAM & Local LLMs

• Memory80 GB · HBM2e
• Bandwidth1,935 GB/s
• FP16 / BF16312 TFLOPS
• INT8624 TOPS

Bus width

5,120-bit

Architecture

Ampere (GA100)

Cores

6,912

Tensor cores

432

Ecosystem

CUDA

Interface

PCIe 4.0 x16

NVLink

600 GB/s

Power

300 W

Power connector

8-pin EPS

Slots

2

Length

267 mm

Type

Data center

Launch date

Jun 28, 2021

Model	Size	Cards	Total price	Total VRAM	Speed t/s	Total power
Qwen3.8 Flash Next176B (6B active)	111.3 GB	2 cards	$37,998	160 GB	132.9	600 W
DeepSeek V4 Flash284B (13B active)	155.1 GB	4 cards	$75,996	320 GB	104.7	1,200 W
Hy3295B (21B active)	182.2 GB	4 cards	$75,996	320 GB	58.9	1,200 W
GLM 5.3 Flash320B (18B active)	199.7 GB	4 cards	$75,996	320 GB	80.7	1,200 W
GLM 5.3744B (40B active)	467.3 GB	8 cards	$151,992	640 GB	44.2	2,400 W
Hy4 preview770B (49B active)	467.3 GB	8 cards	$151,992	640 GB	38.9	2,400 W
Muse Spark Open816B (42B active)	505.5 GB	8 cards	$151,992	640 GB	43.3	2,400 W
DeepSeek V4.1 Flash763B (8 / 16B active)	510.3 GB	8 cardsEngram in RAM	$151,992	640 GB	71.7	2,400 W
Kimi K2.61,000B (32B active)	583.7 GB	8 cards	$151,992	640 GB	55.4	2,400 W
Kimi K2.7 Code1,000B (32B active)	583.7 GB	8 cards	$151,992	640 GB	55.4	2,400 W
Mistral Large 41,000B (52B active)	604.8 GB	8 cards	$151,992	640 GB	37.5	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

   Tell the factory PCIe version from SXM4 conversions

   On the used market, some sellers fit server SXM4 modules onto adapter boards and sell them as PCIe cards. Listing titles often say “SXM4 converted”, and even the sellers admit the firmware still identifies the card as the SXM version. Converted cards work, but cooling, power delivery and stability depend on whoever made the adapter board, and there is no factory warranty. Run nvidia-smi once the card arrives: the factory PCIe version shows “A100 80GB PCIe”, while a converted card shows “A100-SXM4-80GB” (lspci device IDs 20B5 and 20B2 respectively).

2. 2

   All 300W leaves through airflow

   The card has no fan, so all 300W of heat has to be carried off by the heatsink and case airflow, 50W more than the 40GB version. In a workstation, either add an air shroud with a high-static-pressure fan or switch to a water block. Power comes in through one 8-pin CPU (EPS) connector, so a regular power supply needs a PCIe-to-CPU 8-pin adapter cable.

3. 3

   Splitting it among users with MIG

   MIG splits one card into several mini GPUs that don't interfere with each other, and every slice on the 80GB version is twice the size of the 40GB version's. Splitting all the way to seven slices gives 10GB each, enough for Gemma 4 12B at 4-bit; splitting into two 40GB slices lets each run Qwen3.8 27B at 8-bit (29 GB).

4. 4

   FP8 models run, but no faster

   Ampere doesn't support FP8 compute. FP8 models still load in vLLM: the weights are stored in FP8 and converted to FP16/BF16 for computation, so you save VRAM but gain no speed. For throughput, look for INT4 versions such as AWQ or GPTQ; if you have enough VRAM, you can also just use the original BF16 weights.

Spec	A100 80GB PCIe	H100 PCIe	RTX PRO 6000
Memory	80 GB	80 GBSame	96 GB20% more
Bandwidth	1,935 GB/s	2,000 GB/s3% higher	1,792 GB/s7% lower
BF16 / FP16 compute	312 TFLOPS	756 TFLOPS142% higher	500 TFLOPS60% higher
INT8 compute	624 TOPS	1,513 TOPS142% higher	1,000 TOPS60% higher
Qwen3.8 27B speed (2-bit)	122.2 t/s	125.6 t/sSame	114.6 t/s6% slower
Largest dense model (4-bit)	136B	136BSame	165B21% more
Power	300 W	350 W17% higher	600 W100% higher
GPU interconnect	NVLink 600 GB/s	NVLink 600 GB/s	Not supported
Interface	PCIe 4.0 x16	PCIe 5.0 x16	PCIe 5.0 x16
Released	2021-06	2022-03	2025-03
eBay median	$18,999Buy A100 80GB PCIe	View on eBayView on eBay · H100 PCIe	$18,550Buy RTX PRO 6000

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.