I just hope this lives up to the hype. I built my last machine using Phenom II X6 1090T but I always regretted it after; as the 6 core advantage that I thought future games will use, never materialized and per core performance of that was not as good as the Intel CPUs of the same generation. Most of the benchmark leaks have focussed on top of the line Ryzens, I would be interested in seeing how the mid level Ryzens are comparing to Intel I5s.
There are already reports of Intel CPUs getting price cuts, so this looks good for now at least.
Yes and no: here is a short article about it from Brad Wardell the CEO from Stardock [0]. Vulkan/Mantle/DirectX12 allow the game developer to really take advantage of multicore CPU's. Which is why we see the most gains with these new low level API's for AMD's FX chips because they are better at multi-threaded workloads compared to single threaded workloads like with previous DirectX versions.
A quote from the linked article:
To to summarize:
DirectX 11: Your CPU communicates to the GPU 1 core to 1 core at a time. It is still a big boost over DirectX 9 where only 1 dedicated thread was allowed to talk to the GPU but it’s still only scratching the surface.
DirectX 12: Every core can talk to the GPU at the same time and, depending on the driver, I could theoretically start taking control and talking to all those cores.
That’s basically the difference. Oversimplified to be sure but it’s why everyone is so excited about this.
Indeed, they are designed to reduce CPU load. Even so, if the game can't use multiple cores and previously was running game logic & OpenGL driver on the same core, now it's only the game logic. [1]
--
[1] nVidia actually implements Vulkan on top of their OpenGL driver, but nVidia's drivers have been relatively well optimized in terms of CPU usage already. AMD is the bigger winner here.
The OpenGL threading model is completely screwed up. You essentially have a global lock per GL context, which means you're restricted to 1 thread for issuing rendering. And in GL a context is for everything, including shaders, texture states, you can't even reasonable do uploads of new scene data in a separate thread.
Vulkan fixes this big time, by allowing apps to construct GPU workloads for a single GPU in parallel. Only the final submission step (which is supposed to be very low overhead if the driver design is decent) is single-threaded per GPU context. And even for that Vulkan is better: It allows you to allocate different contexts for separate engines (e.g. rendering vs. compute vs. copy engine for data up/download to/from the GPU vram).
The lower CPU overhead is just the icing on the cake, the real deal is that Vulkan fixed the threading/locking model.
The phenom lineup was very competitive (even slightly better than Intel's) for most of its lifetime. It wasn't really until a couple of generation into the Core 2 Duos that Intel started taking the lead.
Competitive on price, sure. But it never held the performance crown. AMD hasn't held that since the Athlon 64 era.
The first gen Phenoms were outpaced by the Core 2 lineup, and the second gen Phenoms that became quite popular were handily beaten by the first gen i7's. The Phenom II was a great bang-for-the-buck chip though, no doubt.
I've owned all the relative players here: the Phenom 9500, Core2 Quad Q6600, Phenom II 955BE, and a core i7-920 (which is still in my main work machine that I'm typing this from right now).
The Q6600 and 955 were comparable in performance (955 much better at stock, slightly better when both chips OC'd), but the i7-920 was out at the 955's release, essentially leaving AMD a generation behind in performance. The 9500 was a dog and well known for an errata that hampered performance after a microcode update, and the i7 was leaps and bounds ahead of anything else at the time, and is still pretty usable today. My little brother has the 955 and it's showing its age in games (KotK being the worst offender).
The Phenom X6 had great performance both in terms of price and actual performance for developer / thread-heavy workloads. Although this is now ~seven years ago.
That only matters for applications that are fully vectorized however. Most real world HPC applications depend much more on memory bandwidth and/or multicore performance and how well the compilers can optimize for the vector units (typically not very well). If I can choose two cores with 128 bit vectors for the same money as one core with 256bit, I'll choose the two core system any day of the week.
I certainly agree that vector performance is overrated. But it seems to dominate HPC spending. At least on the top500. But maybe AMD can pick up some HPC spending for smaller clusters designed to run less specialized code? I don't know how big that market is.
I do think it's big. What you see on top500 is the high end of the market, the big bulk is in small to medium sized clusters or in cloud infrastructure where people rather care for benchmarks like requests per second of some specific servers. There again memory throughput, non-vectorized performance and cores per dollar are probably the most relevant characteristics of a processor. And those who chase to have the highest flop numbers have switched to accelerators and custom chips anyways.
I think this is a bit too simplistic for HPC. Yes, you can go really low cost if all you care is theoretical performance / $. In practice, the cost efficiency of the system should be seen as 1 / (time to solution for typical applications) / (cost of ownership per year).
Even 20% higher memory bandwidth or computational performance per socket matters for the time to solution, since a higher number of nodes is uncertain to replace that due to the communication overhead. As does the performance per watt (costs of cooling).
Furthermore, the biggest news for me regarding Ryzen is what they're doing with APUs. Affordable APUs with unified memory supporting HBM (stacked memory) could be a game changer. HBM means an order of magnitude higher memory bandwidth than what we're used to from CPUs, matching what the high end Nvidia Pascal cards are currently offering. By coupling this to a capable multicore X86 CPU with fat cores (as opposed to Knights landing with a higher number of slow cores), could mean tremendous speedups without any programming work for bandwidth bound applications - which includes pretty much any stencil application, e.g. atmospheric models, ocean models, earthquake prediction etc. By 'tremendous' I mean 5-10x per Socket, which is a big deal in these fields.
> Yes, you can go really low cost if all you care is theoretical performance / $.
Yes that's a fact, and if you're trying to evaluate a solution based on performance/$ then the comparison is made with regards to this factor, and not others.
> In practice, the cost efficiency of the system should be seen as 1 / (time to solution for typical applications) / (cost of ownership per year).
If you believe that criteria is relevant then you should know that Ryzen's advertised TDP is around 95W while some Opteron 61XX processors have a TDP of 85W.
> Furthermore, the biggest news for me regarding Ryzen is what they're doing with APUs. Affordable APUs with unified memory supporting HBM (stacked memory) could be a game changer.
For 40$ you can buy 4 opterons with 12 cores each. This means that for less than 1000$ you can put together a 48-core system that supports up to 512 GB of RAM. For around 500$, anyone can put together a 24-core system that also supports a couple of GPUs.
Anyone that cares about HPC on a budget knows quite well that used Opterons is where the optimal price/performance ratio can be found, particularly as Opterons already out-perform Xeons in BLAS-based work.
There are already reports of Intel CPUs getting price cuts, so this looks good for now at least.