Canned NoodlesVery cool use of an M.2 slot. The price from WPG Americas is $212 (here) but as cal5582 stated, there’s no performance numbers so who’s to say if it’s a good value?

1. This is just an announcement that production version is out for our existing customers. In a couple weeks we'll have more volume for direct purchase on WPGA plus a couple other distributors. Also, I'm pretty sure the price will decrease as volume ramps up, since the current inventory is from a small batch.
2. The SDK & documentation will be open-sourced around the same time, and our docs will include benchmarks across a few hundred models. We don't maintain a "zoo" of tuned models, but instead download -> compile -> run out-of-the-box and update the perf. table. If somebody did want to prune & tune the model (retrain), they could do it and get higher performance but it's not necessary.
3. MX3 out-of-the-box FPS is typically faster than Hailo-8's tuned model zoo FPS. But tuned vs. tuned or OOTB vs. OOTB, the MX3 is much faster. Admittedly we're still early in creating tuned versions of models, since our focus is on running models as-is.
4. It's important to note that TOPS is simply the # of MAC units * frequency. True performance comparisons need to consider the utilization of those MAC units when running a real model, not a theoretical max. A 10-TOPS chip at 50% utilization can have higher FPS than a 40-TOPS chip at 10% utilization. Comparing max TOPS alone between different architectures doesn't equate to final performance.
5. It's 10M weights per chip, 40M total for the M.2 at default precision. Half-precision can go up to 80M for the M.2 but that needs accuracy considerations. Being focused on computer vision (think YOLOs) and not LLMs, the # of weights should give reasonably good coverage. In a future SDK release we'll have ways to run bigger models by swapping chunks to/from the host's memory, but it is always better to run on-chip only.

mxdevHey, MemryX engineer here! I can address some of the comments:

1. This is just an announcement that production version is out for our existing customers. In a couple weeks we'll have more volume for direct purchase on WPGA plus a couple other distributors. Also, I'm pretty sure the price will decrease as volume ramps up, since the current inventory is from a small batch.
2. The SDK & documentation will be open-sourced around the same time, and our docs will include benchmarks across a few hundred models. We don't maintain a "zoo" of tuned models, but instead download -> compile -> run out-of-the-box and update the perf. table. If somebody did want to prune & tune the model (retrain), they could do it and get higher performance but it's not necessary.
3. MX3 out-of-the-box FPS is typically faster than Hailo-8's tuned model zoo FPS. But tuned vs. tuned or OOTB vs. OOTB, the MX3 is much faster. Admittedly we're still early in creating tuned versions of models, since our focus is on running models as-is.
4. It's important to note that TOPS is simply the # of MAC units * frequency. True performance comparisons need to consider the utilization of those MAC units when running a real model, not a theoretical max. A 10-TOPS chip at 50% utilization can have higher FPS than a 40-TOPS chip at 10% utilization. Comparing max TOPS alone between different architectures doesn't equate to final performance.
5. It's 10M weights per chip, 40M total for the M.2 at default precision. Half-precision can go up to 80M for the M.2 but that needs accuracy considerations. Being focused on computer vision (think YOLOs) and not LLMs, the # of weights should give reasonably good coverage. In a future SDK release we'll have ways to run bigger models by swapping chunks to/from the host's memory, but it is always better to run on-chip only.

mxdevHey, MemryX engineer here! I can address some of the comments:

1. This is just an announcement that production version is out for our existing customers. In a couple weeks we'll have more volume for direct purchase on WPGA plus a couple other distributors. Also, I'm pretty sure the price will decrease as volume ramps up, since the current inventory is from a small batch.
2. The SDK & documentation will be open-sourced around the same time, and our docs will include benchmarks across a few hundred models. We don't maintain a "zoo" of tuned models, but instead download -> compile -> run out-of-the-box and update the perf. table. If somebody did want to prune & tune the model (retrain), they could do it and get higher performance but it's not necessary.
3. MX3 out-of-the-box FPS is typically faster than Hailo-8's tuned model zoo FPS. But tuned vs. tuned or OOTB vs. OOTB, the MX3 is much faster. Admittedly we're still early in creating tuned versions of models, since our focus is on running models as-is.
4. It's important to note that TOPS is simply the # of MAC units * frequency. True performance comparisons need to consider the utilization of those MAC units when running a real model, not a theoretical max. A 10-TOPS chip at 50% utilization can have higher FPS than a 40-TOPS chip at 10% utilization. Comparing max TOPS alone between different architectures doesn't equate to final performance.
5. It's 10M weights per chip, 40M total for the M.2 at default precision. Half-precision can go up to 80M for the M.2 but that needs accuracy considerations. Being focused on computer vision (think YOLOs) and not LLMs, the # of weights should give reasonably good coverage. In a future SDK release we'll have ways to run bigger models by swapping chunks to/from the host's memory, but it is always better to run on-chip only.

HBSoundI read over the page, and I am trying to figure out if something like this is beneficial. How does this help a system? No software is needed, so how does it integrate into the system?


Thanks for the info.

Do any of your tuned models have anything to do with enhancing for CAD/3D/BIM/Rendering?

HBSoundI read over the page, and I am trying to figure out if something like this is beneficial. How does this help a system? No software is needed, so how does it integrate into the system?


Thanks for the info.

Do any of your tuned models have anything to do with enhancing for CAD/3D/BIM/Rendering?

mxdevAh sorry, by models I meant neural network models, not CAD, etc. If there's a neural net that accelerates/improves these applications though, they might be possible -- integration of our runtime libraries into the applications would likely need us to work with the software vendor though.

The MX3 is designed for real-time computer vision, so applications we could be used in includes AI security cameras, factories (defect inspection, etc.), "smart city" (traffic intersection monitoring, license plate readers, etc.), and robotics (self-driving, depth estimation / navigation, etc.). On desktop PCs we can do stuff like video super-resolution or video call filters, or there was a 3D avatar thing Lenovo did at CES 2024 that used the MX3. Not sure on the category for this one, but I've used an Orange Pi + MX3 + XREAL Air to make some homebrew AR glasses too.

So basically our main areas are secuity, industrial, and robotics.