The short version
- ECC on: roughly 28 GB usable, with memory error correction.
- ECC off: roughly 32 GB usable, without that correction.
- The extra capacity can prevent OOM, CPU offload, or a smaller context. That can be a large practical speed win.
- Intel still specifies the B70 at 608 GB/s. We have no controlled evidence that ECC-off increases raw bandwidth, so we do not claim it does.
What the switch changes
The B70 has 32 GB of physical GDDR6 and supports ECC. With ECC enabled, part of that memory is used for error correction. Intel's own support article describes the visible result as 28 GB instead of 32 GB. Intel documented
| Choose | Best fit | What you give up |
|---|---|---|
| ECC off | Local inference, model development, or benchmarking where capacity is the hard limit | VRAM error correction |
| ECC on | Long unattended jobs, reliability-sensitive work, or diagnosing possible memory errors | About 4 GB of usable capacity |
If your workload already fits fully in VRAM, turning ECC off may not move tok/s in a meaningful way. If the extra capacity prevents CPU/RAM offload, paging, an out-of-memory failure, a smaller KV cache, or reduced concurrency, the practical improvement can be enormous. That is a capacity effect—not evidence of higher raw memory bandwidth.
Check ECC on Linux
Use Intel XPU-SMI. Discover the IDs first; do not assume that the B70 you want is always device 0.
xpu-smi discovery
xpu-smi config --device 0
xpu-smi config --device 1Only query IDs that discovery actually lists. Under Memory ECC, look for both:
Current: disabled
Pending: disabledCurrent is active now. Pending is what should become active after the required reset or reboot. If they differ, the pending state is not yet valid benchmark metadata.
Disable ECC on Linux
Stop inference servers, containers, desktop applications, and other processes using the selected GPU. The change may require a reboot. Save your work before proceeding.
Change each explicitly identified B70, one at a time:
sudo xpu-smi config --device 0 --memoryecc 0
sudo xpu-smi config --device 1 --memoryecc 0
sudo rebootRun the device-1 command only if device 1 is another B70 you intend to change. After reboot, repeat discovery and configuration checks. Require both Current and Pending to say disabled; do not infer the state only from a larger memory number.
Re-enable ECC on Linux
Use 1 instead of 0, then reboot and verify:
sudo xpu-smi config --device 0 --memoryecc 1
sudo xpu-smi config --device 1 --memoryecc 1
sudo rebootBoth fields should say enabled afterward. Re-check your model, KV-cache, and concurrency budget before relaunching: the usable capacity will shrink again.
Windows
Intel documents the setting in Intel Graphics Software → Graphics → ECC. Stop GPU workloads, choose the desired state, reboot if requested, and verify both the setting and reported VRAM afterward. Driver interfaces move over time, so use Intel's current support article rather than treating an old screenshot as authoritative.
Pros, cons, and side effects
- More usable VRAM with ECC off. Intel documents approximately 28 GB versus 32 GB. Exact MiB/GiB displays vary with units and driver reservations.
- Less fault protection with ECC off. An uncorrected memory fault can affect any computation. ECC on is the conservative choice for reliability-sensitive work.
- No published bus change. The card remains a 256-bit GDDR6 design rated at 608 GB/s. Intel has not published a matched B70 bandwidth A/B for these two states.
- No intentional model-quality change. Weights and arithmetic are the same, but memory corruption is always capable of changing a result.
- The state persists. Treat it as part of the machine's benchmark identity, not a per-process flag.
What our lab is using
On August 30, 2026, both B70s on our current two-card host reported ECC Current: disabled and Pending: disabled. Each card reported 32,656 MiB total memory. Lab-checked Exact observation
Recent runs made on this host therefore used ECC disabled unless the run artifact says otherwise. Older manifests did not always record ECC state, so we will not retroactively label every historical result as ECC-off. Going forward, ECC-sensitive results should retain the device IDs and both ECC fields in their hardware receipt.
If the command is missing or unsupported
- If
xpu-smiis missing, install a current Intel XPU Manager/XPU-SMI package supported by your driver stack. - If
--memoryeccis absent fromxpu-smi config --help, do not invent anxemodule option or copy an undocumented sysfs command. Update the supported management stack or check with Intel or the system vendor. - If Pending changes but Current does not, reboot. Intel says a device reboot may be required.
- If a model stops fitting after ECC is enabled, reduce context, cache allocation, concurrency, or model size before treating it as a runtime bug.