I bought two of these and from day one it was a journey of disappointment as I realized I'd been scammed by the Leather Jacket, particularly due to NVFP4 performance (lack thereof). But my impression has gradually improved and I've learned to take advantage of what they're actually good at. Compared to Strix Halo:

1) You can very easily link two together to roughly double performance and VRAM capacity to get it in a much more usable and interesting range. It's not exactly double but considerably better than single-node. Tensor parallelism in this mode works extremely well and is perfectly stable. This opens the door to ~400B models or even more with a modest REAP. You can also go up to four or even eight nodes but the scaling isn't linear and you need a fancy network box thing that's pretty expensive.

2) Despite still lacking NVFP4 the FP8 performance is quite good and with a pair of them together there's a wide variety of models you can run that way. FP8 runs at almost exactly the same speed as INT4. Strix Halo doesn't support FP8 acceleration.

3) A simple symlink from one spark to the other (and vice versa) effectively allows you to double the available storage by storing models on only one Spark and loading them on-demand across the super fast 200 mbps direct connection. This means even the 1TB models are perfectly adequate, that's tons of storage for models considering you'll probably only need to have a handful at the ready.

4) I found a cheap source of the aforementioned 200G cables from China. Since you actually need two to get the full advertised bandwidth and you can get two such Chinese cables for cheaper than one "official" cable and this was a nice surprise since they're impossible to get your hands on here. And alternatively you can get a pair of much cheaper 100G cables very easily from Amazon for a pretty fair price.

5) There are community attempts to get NVFP4 working which are making progress, although in all honesty at this point I've gotten used to ignoring that function, so I'll need to see some KLD and perplexity benchmarks before I'll believe it's worth the bother. I do believe that NVFP4 can potentially be a game-changer but it seems like it requires a lot of legwork on the part of the person who does the quantization to "get it right" and I'd be hesitant to trust randos on huggingface to do that.

6) Minimax 2.7 is expected very soon and I'll be able to run that at INT4 (if not NVFP4) with absolute assloads of room left over for full FP16 200K context with 2-4x concurrency. Strix Halo will need to quantize to Q3 and even then may not have enough room for full 200K for single-user let alone concurrency. Considering how exciting that model seems to be, this alone is a huge thing in Spark's favor.