Fenvi Riser Works, Dual GPU Limited

Fenvi Riser Works, Dual GPU Limited | Generated by AI

Home 2026.07

Question: Can the Fenvi PCIe 5.0 x16 riser cable work with MSI MAG B760M Mortar and allow two RTX GPUs?

Answer:

Short answer: yes, the PCIe 5.0 x16 riser cable is compatible with MSI MAG B760M Mortar, but it does NOT solve the “two RTX GPU” problem by itself. The limitation is the motherboard PCIe layout, not the riser cable.

Your setup:

Motherboard: MSI MAG B760M Mortar
CPU: likely Intel 12th/13th/14th gen
GPU: RTX 4070 / RTX 3090 etc.
Riser: Fenvi PCIe 5.0 x16 20cm

1. Will PCIe 5.0 riser work?

Yes.

PCIe is backward compatible:

PCIe 5.0 riser
        |
        v
PCIe 4.0 motherboard slot
        |
        v
RTX 4070 / RTX 3090

It will negotiate down:

PCIe 5.0 x16 cable
        ↓
PCIe 4.0 x16 link
        ↓
GPU runs normally

RTX 4070 is PCIe 4.0 x16. RTX 3090 is PCIe 4.0 x16.

No performance loss from a good 20cm PCIe 5.0 riser.

2. Can MSI B760M Mortar run two RTX cards?

The answer is: physically yes, practically limited.

The board usually has:

CPU PCIe lanes:

16 lanes -> top GPU slot
4 lanes  -> M.2 SSD

Chipset PCIe lanes:

x4/x1 slots

Typical layout:

CPU
 |
 +---- PCIe x16 slot
 |        |
 |        RTX GPU #1
 |
 +---- M.2


Chipset
 |
 +---- PCIe x16 physical slot (x4 electrical)
          |
          RTX GPU #2

So:

GPU #1:

PCIe 4.0 x16

GPU #2:

PCIe 4.0 x4

The second GPU works, but bandwidth is much lower.

3. For AI workloads, is x4 enough?

Depends.

For inference:

Example:

RTX 4070 12GB
+
RTX 3090 24GB

You can run:

vLLM multi-GPU
llama.cpp split mode
PyTorch model parallel
CUDA distributed

The second GPU being x4 is often acceptable.

Reason:

LLM inference is mostly:

GPU compute
+
VRAM capacity

not PCIe bandwidth.

Example:

70B model

GPU0:
35GB

GPU1:
35GB

PCIe traffic:
only activation synchronization

Not constantly moving the whole model.

For training:

Much worse.

Example:

DDP training

GPU0 <---- gradients ----> GPU1

You want:

PCIe x16
NVLink
InfiniBand

x4 becomes a bottleneck.

4. Physical problem: GPU size

The bigger issue with B760M Mortar:

Two RTX cards normally cannot fit.

Example:

RTX 3090:

3-slot width

RTX 4070:

2-slot width

Together:

3090 ███
4070 ██

needs 5 slots

mATX board:

[ GPU slot ]
[ blocked ]
[ small slot ]

Usually impossible.

A riser helps:

GPU #1
vertical outside case


GPU #2
normal slot

Example:

        RTX3090
          |
          |
     PCIe riser
          |
----------------
 MSI B760M
----------------
      RTX4070

5. Recommended configuration

For your AI experiments:

Option A: Keep B760M

Good:

RTX3090
+
RTX4070

RTX3090 -> main x16 slot
RTX4070 -> riser / second slot x4

Works for:

local LLM inference
agents
vLLM testing
LoRA experiments

Option B: Serious multi-GPU

Move to workstation platform:

Threadripper
EPYC
Xeon W

Example:

PCIe lanes:

GPU1 x16
GPU2 x16
GPU3 x16
GPU4 x16

Then multiple RTX 3090/4090 makes sense.

6. About Fenvi PCIe 5.0 riser

The product you mention:

Fenvi PCIe5.0 X16
20cm
90 degree

is actually a good choice.

Why:

PCIe 5.0 signal quality margin
backward compatible
short cable (20cm is better than 50cm)
x16 wiring

For AI rigs, I would prefer:

PCIe 5.0 x16 riser
20cm
shielded

over cheap PCIe 3.0 risers.

Final recommendation for your B760M:

MSI B760M Mortar
        |
        +-- RTX3090
        |     PCIe x16
        |
        +-- RTX4070
              PCIe x4 via Fenvi riser

This is a reasonable AI inference / agent workstation.

But if your goal is:

4x RTX3090
5x RTX3060
multi-GPU training

then don’t spend money on risers for B760M. The CPU PCIe lane count is the real wall.

Back Donate