512GB unified memory is absolutely wild for AI stuff! Compared to how many NVIDIA GPUs you would need, the pricing looks almost reasonable.
(sorry, should have specified that the NPU and GPU cores need to access that ram and have reasonable performance). I specified it above, but people didn't read that :-)
CUDA has had managed memory for a long time now. You absolutely can address the entire host memory from your GPU. It will fetch it, if it's needed. Not fast, but addressable.
Also, 32GB DDR5 RDIMMS are ~200, so that’s 5K for 24 right there. Then you need 2x CPUs, at ~1K for the cheapest, and you need 2, and then a motherboard that’s another 1K. So for 8K (more, given you need a case, power supply, and cooling!), you get a system with about half the memory bandwidth, much higher power consumption, and very large.
You do not need 2 CPUs. If however you use 2 CPUs, then the memory bandwidth doubles, to 1152 GB/s, exceeding Apple by 40% in memory bandwidth. The cost of the memory would be about the same, by using 16 GB modules, but the MB would be more expensive and the second CPU would add to the price.
The memory bandwidth does not double, I believe. See this random issue for a graph that has single/dual socket measurements, there is essentially no difference: https://github.com/abetlen/llama-cpp-python/issues/1098
Perhaps this is incorrect now, but I also know with 2x 4090s you don’t get higher tokens per second than 1x 4090 with llama.cpp, just more memory capacity.
(All if this only applies to llama.cpp, I have no experience with other software and how memory bandwidth may scale across sockets)
With a badly organized program, the performance can be limited not by the memory bandwidth, which is always exactly double for a dual-socket system, but by the transfers on the inter-socket links.
Moreover, your link is about older Intel Xeon Sapphire Rapids CPUs, with inferior memory interfaces and with more quirks in memory optimization.
But where is your data? For llama.cpp? For whatever dual socket CPU system you want. That’s all I am claiming.
https://github.com/ggml-org/llama.cpp/discussions/11733
about the scaling of llama.cpp and DeepSeek on some dual-socket AMD systems.
While it was rather tricky, after many experiments they have obtained an almost double speed on two sockets, especially on AMD Turin.
However, if you look at the actual benchmark data, that must be much lower than what is really possible, because their test AMD Turin system (named there P1) had only two thirds of the memory channels populated, i.e. performance limited by memory bandwidth could be increased by 50%, and they had 16-core CPUs, so performance limited by computation could be increased around 10 times.
A single 192 core Epyc is 11k by itself, so I’d probably go for the simpler integrated M3 ultra solution…
Time to first token, context length, and tokens/s are significantly inferior on CPUs when dealing with larger models even if the bandwidth is the same.
When used for ML/AI applications, a consumer GPU has much better performance per dollar.
Nevertheless, when it is desired to use much more memory than in a desktop GPU, a dual-socket server can have higher memory bandwidth than most desktop GPUs, i.e. more than an RTX 4090, and a computational capability that for FP32 could exceed an RTX 4080, but it would be slower for low-precision data where the NVIDIA tensor cores can be used.
INT8, INT4, FP8 and soon FP4
Both CPUs (with the BF16 instructions and with the VNNI instructions for INT8 inference) and the GPUs have a higher throughput for lower precision data types than for FP32, but the exact acceleration factors are hard to find.
The Intel server CPUs have the advantage vs. AMD that they also have the AMX matrix instructions, which are intended to compete for inference applications with the NVIDIA tensor cores, but the Intel CPUs are much more expensive for a number of cores big enough to be competitive with GPUs.
Thinking about it you can get a decent 256gb on consumer platforms now too, but the speed will be a bit crap and would need to make sure the platform ully supports ECC UDIMMs
Like another poster said, 768 GB of ECC RDIMM DDR5-6000 costs around $5000.
Any program whose performance is limited by memory bandwidth, as it can be frequently the case for inference, will run significantly faster in such an EPYC server than in the Apple system, even when running on the CPU.
Even for computationally-limited programs, the difference between server CPUs and consumer GPUs is not great. One Epyc CPU may have about the same number of FP32 execution units as an RTX 4070, while running at a higher clock frequency (but it lacks the tensor units of an NVIDIA GPU, which can greatly accelerate the execution, where applicable).
Any program whose performance is limited by memory bandwidth, as it can be frequently the case for inference, will run significantly faster in such an EPYC server than in the Apple system, even when running on the CPU.
However Apple does not say anything about the GPU clock frequency, which I assume that it is significantly less than that of NVIDIA.
In comparison, a dual-socket AMD Turin can have up to 12288 FP32 execution units, i.e. 20% more than an Apple GPU.
Moreover, the clock frequency of the AMD CPU must be much higher than that of the Apple GPU, so it is likely that the AMD system may be at least twice faster for computing some graphic application than the Apple M3 Ultra GPU.
I do not know what facilities exist in the Apple GPU for accelerating the computations with low-precision data types, like the tensor cores of NVIDIA GPUs.
While for graphic applications big server CPUs are actually less compute constrained than almost all consumer GPUs (except RTX 4090/5090), the GPUs can be faster for ML/AI applications that use low-precision data types, but this is not at all certain for the Apple GPU.
Even if the Apple GPU happens to be faster for some low-precision data type, the difference cannot be great.
However a server that would beat the Apple M3 Ultra GPU computationally would cost much more than $10k, because it would need CPUs with many cores.
If the goal is only to have a system with 50% more memory and 40% more memory bandwidth than the Apple system, that can be done at a $10k price.
While such a system would become compute constrained more often than an Apple GPU, it would still beat it every time when the memory would be the bottleneck.
I have just compared the FP32 computational capabilities, i.e. what is used for graphics, between the Apple M3 Ultra GPU and AMD server CPUs, because these numbers are easily available and they demonstrate the size relationships between them.
Both GPUs and server CPUs have greater throughputs for lower precision data (CPUs have instructions for BF16 and INT8 inference), but the exact acceleration factors are hard to find and it is more difficult to estimate the speeds without access to such systems for running benchmarks.
I'd like to see some proper benchmarking on this though, but it looks like the Apple systems might just be extremely good value if you want to run the large DeepSeek model.
says who? NVIDIA has essentially entrenched themselves thanks to CUDA
Don't tell my boss! I still get paid.