Feels likely to be a pretty direct test of memory bandwidth. Strix Halo (the AI Max+ 395) has quad ddr5-8000, which is both twice the width and more mhz. This is a great opportunity to see what workloads will make use of the 3d-vcache too!
I'm personally pretty excited for a new round of Mobile on Desktop. Usually a pretty good % of desktop speed, but tuned for much lower power, and the price is usually a little better. I have the Minisforum 795s7 with the 7945hx, which is a delightful zen4 that's very affordable.
https://www.tomshardware.com/pc-components/cpus/amd-unleashe...
Indeed yes it seems like the dual channel 6000MHz wasn't the limiting factor in many*
tests. But also the lower power 395 has a couple massive wins, and I suspect it was almost entirely the memory bandwidth. For example the LLM test on page 2: it was that memory bandwidth making that colossal difference.
There were also tests like OpenVINO where memory bandwidth was the factor, but the x3d's cache was evidently big enough, with it easily scoring first place, the Strix Halo second, then the regular 9950 last place!
In a lot of tests, it seems like the extra 100w the 9950's were getting probably helped a lot!
It would have been interesting to see the graphics performance and AI performance of the 395 compared to discrete gpus. None of the benchmarks in this article really seem useful for real world buying guidance.
ie how does it stack up against using real graphics cards
Terrible benchmarks. They should probably compare the 395 GPU against the 9950 CPU or against a 7600/9060. If you're just going to use the CPU obviously the 395 is not what you want.
M.L. was benchmarking various standard Phoronix tests on the same CPU core in three different 16-core configurations: one that was asymmetric with fast cache, one with the highest per core TDP, and one with lower peak clocks but a fast "inter-CCD" (TBF, I'm not sure they qualify as CCDs in the 395) connect and high memory bandwidth.
While I've always disliked the way M.L. does their overall "geometric mean", each of the three configurations performed favorably on a subset of the tests, which implies that more than just TDP was tested.
I hate the naming so much. I cannot tell at all which is meant to be newer or superior. It's also an odd benchmark considering the LLM should be running on the GPU with shared memory rather than CPU
Intel and AMD have pretty much never made it straightforward to compare laptop part numbers against desktop part numbers and know what's faster or newer. That's not a new problem.
If you compare the Ryzen AI Max+ 395 against the other AMD mobile parts in their current lineup, it's pretty clear where it ranks. And for however long they're using a new and different naming scheme for their mobile parts while still using the old naming scheme for their desktop parts, there's the distinct and rare advantage that the model numbers are not misleading about how a high-end mobile part ranks against desktop parts.
Feels likely to be a pretty direct test of memory bandwidth. Strix Halo (the AI Max+ 395) has quad ddr5-8000, which is both twice the width and more mhz. This is a great opportunity to see what workloads will make use of the 3d-vcache too!
I'm personally pretty excited for a new round of Mobile on Desktop. Usually a pretty good % of desktop speed, but tuned for much lower power, and the price is usually a little better. I have the Minisforum 795s7 with the 7945hx, which is a delightful zen4 that's very affordable. https://www.tomshardware.com/pc-components/cpus/amd-unleashe...
I don't know if the CCDs can actually use that much bandwidth. It's really for the GPU.
There were sometimes quite notable differences between 32-core ThreadRipper at quad channel ddr5 6400mhz vs 4800mhz. https://www.phoronix.com/review/threadripper-9000-ddr5-6400-...
Indeed yes it seems like the dual channel 6000MHz wasn't the limiting factor in many* tests. But also the lower power 395 has a couple massive wins, and I suspect it was almost entirely the memory bandwidth. For example the LLM test on page 2: it was that memory bandwidth making that colossal difference.
There were also tests like OpenVINO where memory bandwidth was the factor, but the x3d's cache was evidently big enough, with it easily scoring first place, the Strix Halo second, then the regular 9950 last place!
In a lot of tests, it seems like the extra 100w the 9950's were getting probably helped a lot!
It would have been interesting to see the graphics performance and AI performance of the 395 compared to discrete gpus. None of the benchmarks in this article really seem useful for real world buying guidance.
ie how does it stack up against using real graphics cards
Terrible benchmarks. They should probably compare the 395 GPU against the 9950 CPU or against a 7600/9060. If you're just going to use the CPU obviously the 395 is not what you want.
I don't think that was the point of these benchmarks. Instead it is interesting to see how memory bandwidth effects CPU performance.
These benchmarks are measuring TDP not memory bandwidth.
M.L. was benchmarking various standard Phoronix tests on the same CPU core in three different 16-core configurations: one that was asymmetric with fast cache, one with the highest per core TDP, and one with lower peak clocks but a fast "inter-CCD" (TBF, I'm not sure they qualify as CCDs in the 395) connect and high memory bandwidth.
While I've always disliked the way M.L. does their overall "geometric mean", each of the three configurations performed favorably on a subset of the tests, which implies that more than just TDP was tested.
I hate the naming so much. I cannot tell at all which is meant to be newer or superior. It's also an odd benchmark considering the LLM should be running on the GPU with shared memory rather than CPU
Intel and AMD have pretty much never made it straightforward to compare laptop part numbers against desktop part numbers and know what's faster or newer. That's not a new problem.
If you compare the Ryzen AI Max+ 395 against the other AMD mobile parts in their current lineup, it's pretty clear where it ranks. And for however long they're using a new and different naming scheme for their mobile parts while still using the old naming scheme for their desktop parts, there's the distinct and rare advantage that the model numbers are not misleading about how a high-end mobile part ranks against desktop parts.