Fig. 1 Memory mapTop view, not to scale
AD10212 x GDDR648GB total864 GB/s peak bandwidth
Ada Lovelace, PCIe dual-slotHover to light the stacks
Datacenter, Ada Lovelace
48 GB

L40S 48GB

The universal datacenter card. 48 GB, FP8, video encode, runs in any 2U with 350 W to spare.

$8,900

48 GB VRAM $185/GB VRAM

In stock. Lead time 72 hours.

Manufacturer warranty, DHL Express insured

Volume pricing

UnitsDiscountUnit price
1 unitList$8,900
2 to 7 units3% off$8,633
8 to 31 units6% off$8,366
32+ units10% off$8,010

Unit price at 1: $8,900

Request a bulk quote

Specifications

The full sheet.

Figures from the manufacturer datasheet. Price is an indicative street price, shown as from.

VRAM
48 GB
Memory
GDDR6
Bandwidth
864 GB/s
FP8
1.466 PFLOPS
FP4
Not supported
Board power
350 W
Form factor
PCIe dual-slot
NVLink
No
Architecture
Ada Lovelace
GPUs
1
Per GB of VRAM
$185.42
Availability
In stock, ships 72h
Sale
Buy online
SKU
L40S
Overview

What it is for.

Workloads, sizing and what ships in the box.

48 GB, FP8 and video encode in any 2U.

The general-purpose datacenter card. 48 GB of GDDR6, 1.466 PFLOPS of FP8, hardware video encode and decode, 350 W. It fits a standard 2U with dual-slot clearance and needs no NVLink planning. In stock, ships in 72 hours. $8,900, about $185 per GB of VRAM. Bulk tiers start at 2 units: 3% off, then 6% at 8 and 10% at 32.

Good for

  • Llama 3.1 8B at FP16 (19 GB with headroom) at high batch.
  • Mistral Small 24B at FP8 (29 GB) on one card.
  • Llama 3.1 70B at 4-bit (42 GB) on one card, short context.

Fit rule: FP16 weights take 2 bytes per parameter, FP8 1 byte, 4-bit 0.5 bytes, plus 20% headroom for KV cache at short context.

Pairs with

What fits

How much VRAM per box.

Card count is capped by slots and by the chassis power budget.

2U 4x PCIe
4 x 48 GB, 1,400 W of 3,200 W
Configure this box
Power44% of budget
192 GBper box
4U 8x PCIe
8 x 48 GB, 2,800 W of 6,400 W
Configure this box
Power44% of budget
384 GBper box
Pairs with

Parts bought alongside.

Same workload, a different trade between VRAM, bandwidth and price.