I'm particularly interested in the benefits for web browsing, given that pretty much all modern browsers make good use of multiple processor cores, though whether additional processor threads are significantly beneficial I really have no idea. I've only got an i7-920 here with hyperthreading so I can't easily test this for myself in any meaningful way, and I assume that hyperthreading implementations plus architecture improvements would render any testing with such an older processor a bit of a moot point.
I found this: "It needs to be mentioned that every single benefit Ryzen 7 has without SMT was mirrored by the i7-6900K with its own Hyper-Threading disabled. One interesting result of this testing was Ryzen 7’s SMT scaling beating the Broadwell-E; turning on Simultaneous Multithreading typically led to larger gains than Hyper-Threading. It looks like AMD’s version is simply more efficient than Intel's older architecture. How does it fare against the improvements baked into Skylake and Kaby Lake? That's a question best left to another article."
I found this: "It needs to be mentioned that every single benefit Ryzen 7 has without SMT was mirrored by the i7-6900K with its own Hyper-Threading disabled. One interesting result of this testing was Ryzen 7’s SMT scaling beating the Broadwell-E; turning on Simultaneous Multithreading typically led to larger gains than Hyper-Threading. It looks like AMD’s version is simply more efficient than Intel's older architecture. How does it fare against the improvements baked into Skylake and Kaby Lake? That's a question best left to another article."
SMT benefits for browser workloads depend on the number of physical cores available. Dual cores will show the best results, quad cores will still benefit, six cores an up will likely see no perceivable gains.
I disagree. Intel is better at extracting the most IPC from a single thread meaning the second or HT-thread will get less time to run and hence less performance increase from HT vs SMT. The more "load" your core has from the first thread the less you can gain with HT. So getting large gains from HT/SMT can also be interpreted as a negative especially since we know that intels single-thread performance is still a lot better than AMDs. I would rather like to see AMD catching up to intel in ST performance but getting less gains from SMT.
It's also a design decision. The more hardware (ports, execution units,...) you put at the problem, the higher the HT/SMT benefit can be.
So IMHO it is not possible to say HT or SMT is better.
I disagree. Intel is better at extracting the most IPC from a single thread meaning the second or HT-thread will get less time to run and hence less performance increase from HT vs SMT. The more "load" your core has from the first thread the less you can gain with HT. So getting large gains from HT/SMT can also be interpreted as a negative especially since we know that intels single-thread performance is still a lot better than AMDs. I would rather like to see AMD catching up to intel in ST performance but getting less gains from SMT.
It's also a design decision. The more hardware (ports, execution units,...) you put at the problem, the higher the HT/SMT benefit can be.
So IMHO it is not possible to say HT or SMT is better.
SMT doesn't just give you free performance. It allows tasks to utilize resources that were otherwise going to waste with the current workload. A higher yield from SMT implies poor utilization of resources. It doesn't mean that the SMT implementation in Ryzen is better than Intel, it means Intel's architecture can better use it's own resources, leaving less on the table to gain with a second thread.
We can't isolate SMT implementation from the actual hardware it's running on and say it's better or worse, but we can judge throughtput of the cores when SMT is utilized and differentiate based on that. Zen cores shine under throughput oriented loads - thread count close to logical core count, Lake cores look a lot better when thread usage hovers around the physical core count.
This way you are both right at the same time
BTW, this way of evaluating SMT performance is also useful when choosing a CPU for browsing and other daily activities. My rule of thumb would look something like this:
for maximum value aim for a logical core count close to what modern browsers can utilize. Based on the last article I read on the matter, that would be around 6 threads, so an 4c/8t CPU would be nice.
for maximum performance aim for a physical core count close to the magical number - in this case 6. That includes everything going up from 6c/6t.
The rule above applies when clocks are similar, a 20% difference in clocks would turn everything upside down.
SMT doesn't just give you free performance. It allows tasks to utilize resources that were otherwise going to waste with the current workload. A higher yield from SMT implies poor utilization of resources. It doesn't mean that the SMT implementation in Ryzen is better than Intel, it means Intel's architecture can better use it's own resources.
We can only assume that based on the fact that we benchmarked both cores and found Intel's cores are faster in ST loads. That being said, while a high SMT yield may indicate poor utilization of resources in ST loads, a poor SMT yield does not necessarily indicate good utilization of resources. Also, it's a bit misleading to say high SMT yield indicates poor utilization of resources, since those "extra" resources were likely put in place with the exact purpose of being leveraged through SMT.
Hardware resources and SMT are interlinked and should be judged together.
SMT doesn't just give you free performance. It allows tasks to utilize resources that were otherwise going to waste with the current workload. A higher yield from SMT implies poor utilization of resources. It doesn't mean that the SMT implementation in Ryzen is better than Intel, it means Intel's architecture can better use it's own resources, leaving less on the table to gain with a second thread.
Not quite. Back in the day when it was launched (P4) HT was more like the SMT we see from AMD. But programs were so badly threaded there was often a large hit with HT turned on. Surprisingly this applied more to Server programs than most others. So between tweaks to HT, patches in windows, and Intel's further developments. HT is what it is now. A low priority, labeled as not a core, cycle borrower. That isn't how SMT is on Ryzen. To describe how SMT and HT (Intel SMT) works its easy to think of it as two sets of instructions being squeezed out the same pipeline so they fit in the gaps to get their work done. But there is a lot more going on there than that specially with AMD's implementation. AMD getting a larger benefit from SMT doesn't mean the CPU isn't doing what it should be doing on a single thread. That isn't to say Ryzen is up to Intel's performance on single threaded work. It's obvious it's not. But what you are describing isn't an apples to apples comparison. I can't think of a good analogy, I mean maybe a car one, but the point is that there are enough differences in CPU arch that trying to state that SMT runs better on Ryzen because it sucks at Single thread is absolutely baseless.
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.
Facebook
Twitter