I am afraid not anymore. AMD has become as greedy as Nvidia or Intel. Their server CPUs are not any cheaper than Intel's and the Ryzen series enjoys a solid reputation but I do see advantages of why to go with an Intel CPU. One is the better support for Intel's vector extensions. AMD has not yet implemented AVX-512 (which they are currently working on) and that makes Intel offer a solid performance boost for vectorized computations. If you ever compared AMD and Intel CPUs when training and inferencing deep neural network workloads you would know what I mean. While some work can be outsourced to be computed on GPUs, the CPU still plays a huge role. Let me give you an example. Tensorflow offers an entire data pipeline architecture in its Keras namespace. The data pipeline is the part between your data source and the DNN layers of your network. When you cache and prefetch the data then the pipeline ensures that the GPU is constantly utilized and provided with batches. Keras does that by spinning up several threads on CPUs in its thread pool. Many of the data pipeline operations are operating on NumPy arrays and the transformations are mostly vectorized. Doing so with a modern Intel CPU vs AMD makes a huge difference, I saw 50-100% performance improvements with several I9 CPUs I used vs 1st and 2nd gen Threadrippers and EPYCs.I can't speak for the newer retail Ryzens but from a technical perspective, there is no reason why they should perform better. In this space it's not about raw GHz but also about how performant the CPU is, operating on vectors and matrixes. Of course, this is a moot point for anyone who does not work with applications that benefit from vectorizations.
yep, a workstation for beggars and the homeless. They also give those away at the local Salvation Army.
Good information and I've read the same but also that the current situation has changed. I only have second-hand info and superficial knowledge so I cannot definitely claim anything. Not a big fan of the instruction sets, what I saw on some current i7 CPUs was that using AVX-512 caused the units to overheat within a short period of time, making it unusable for the most part.
I don't share the experience about overheating, I use the vector extensions every day. You can see at respected hardware sites that the CPUs usually throttle down when running the vector units. That keeps the temp within a reasonable range. Raw GHz cycles matter less when operating on matrices hence I think it was a smart choice of Intel to operate the extensions at a lower tact rate. I am not a hardware expert either just share my experience. Ryzen looks very promising and we should all embrace the competition it brings to the cpu market. I guess I am just a burnt sheep, I configured a 1st gen threadripper platform and was sorely disappointed how much it underperformed my 7900x when it came to running back tests and training DNN models. Perhaps the ryzen CPUs perform better now. I can't say. I could not find credible sources that run performance comparisons when it comes to vector operations.
Just picked this up for ~$300: Dell Precision 5810, Xeon E5-1603v3 2.8GHz, 32GB RAM, 256GB SSD + 2TB HDD, Win 10 Pro, AMD FirePro W2100. Did I do well? The amount of RAM on this setup should keep it usable for a long time.
You might be disappointed with that E5-1603, but try it and see. You can always upgrade the CPU for cheap. Otherwise, you did GREAT!
It's incredibly hard to find many reliable tests for MKL/BLAS on Intel/AMD. This is interesting but now dated and very limited in tests - Puget Systems When mentioning the temperatures and throttling, this is the review I was talking about (granted it's not a Xeon) - Anandtech It seems disturbing to it hitting thermal limits and throttling performance, this is quite alarming IMO. It's much worse if your ambient temps are higher and plan to run it long. It's also sad that the reviews state that 11700K is just a repeat of the previous model with higher clocks. Hopefully Intel on 10nm will do better.
My old rig had C2Q Q6600 and it got enough speed to handle anything that I threw at it. The only bottleneck was the lack of RAM when the mobo maxed out at 8GB. So 32GB RAM in the new rig should give it enough headroom to stay usable in the foreseeable future.