- Joined
- Mar 7, 2008
The kit:
i5-5675C 4C/4T 4MB L3 cache, 128MB eDRAM, 3.1 GHz base, 3.6 GHz turbo
Crucial Ballistix 1600 DDR3L 2x4GB
Some cheap 92mm fan Coolermaster hsf, can't remember exact model but bigger wouldn't fit in the case
Asus Z97M-plus
AMD R7 260X (so the integrated GPU doesn't steal resources)
This is intended as a bang for buck pure cruncher, so I actually raided an existing i3 Haswell system for parts apart from the CPU and mobo.
Broadwell is something I wanted to test for a while. My interest is in Prime number finding, and for that you want two things: fast floating point performance (AVX2), and not to be limited by ram performance. Upgrading from Sandy Bridge to Haswell saw a 50% IPC boost where not ram limited. Skylake gave about another 15% IPC boost over Haswell. But when tasks get big, ram limits and performance drops significantly. Faster cores need faster ram to feed it. Enter Broadwell. Instead of looking for ever higher performance ram, would the eDRAM negate that need? 128MB is plenty big enough for existing tasks I'm participating in.
With the system assembled and running, the first test was to see what the CPU does when it is not limited. I ran small tasks and found Broadwell IPC was 6% down from Haswell. Can this be right? Thinking more, could the smaller L3 cache at 4MB be hindering it vs. a Haswell i5 at 6MB? The tasks being run had a data set of 1MB, so with 4 cores active, this is right on the limit of Broadwell i5. Ok, I ran both systems with only 3 tasks instead of 4. The difference remained. For whatever reason, it seems like Broadwell is reduced in IPC compared to Haswell.
But that isn't why I got Broadwell, and put bigger tests on it. I used the Prime95 built in benchmark to check out core scaling, as more cores are active, more demands are made of ram. Sure enough, I saw better core scaling. That is, running 4 tasks the throughput peaked about 3.8x that of a single task, whereas my fastest Skylake with high speed DDR4 was only scaling to about 3.3x. However the Skylake is also much faster in peak core performance so it would be more demanding.
I saw I could overclock the eDRAM. I turned it up from default 1800 to 2200 and... it made no significant difference. I was rather hoping that if bandwidth was the limit, this would result in a gain. I don't know if there is some other bottleneck working here.
What about different ram configurations. In short, for combined tasks which fit within the eDRAM, the ram performance didn't matter. Only once that was exceeded, did you see the ram performance affect throughput.
And last but not least, how about overclocking the core? The cooler isn't the best but the mobo by default did some kind of enhanced turbo and run 4 cores at 3.5 GHz. Fixing it at default voltage, I went to 4 GHz which was bench stable. Temps were into the 80's C so I didn't want to push harder or higher voltage. The bench results were promising, coming in only 5% down compared to my 4.2 GHz Skylake with high speed ram. The drawback was, when I tried running real work the system wasn't stable and I had crashes. So I've returned to 3.5 GHz and haven't tried tinkering again since then. I will eventually revisit this and see if I can find a balance between temps and clock+voltage.
I've left it like that for now. With better cooling and a tad more voltage I think it would give Skylake a run for its money in memory demanding tasks, at lower cost. Although the respective overclockable i5 CPUs are similar in price, the Broadwell mobos tend to be cheaper, plus you don't need to get high speed DDR4 ram. Any cheap DDR3 would suffice. Right now I have a single stick running in it, as I stole the other stick to temporarily resurrect the i3 I had removed to build this with.
i5-5675C 4C/4T 4MB L3 cache, 128MB eDRAM, 3.1 GHz base, 3.6 GHz turbo
Crucial Ballistix 1600 DDR3L 2x4GB
Some cheap 92mm fan Coolermaster hsf, can't remember exact model but bigger wouldn't fit in the case
Asus Z97M-plus
AMD R7 260X (so the integrated GPU doesn't steal resources)
This is intended as a bang for buck pure cruncher, so I actually raided an existing i3 Haswell system for parts apart from the CPU and mobo.
Broadwell is something I wanted to test for a while. My interest is in Prime number finding, and for that you want two things: fast floating point performance (AVX2), and not to be limited by ram performance. Upgrading from Sandy Bridge to Haswell saw a 50% IPC boost where not ram limited. Skylake gave about another 15% IPC boost over Haswell. But when tasks get big, ram limits and performance drops significantly. Faster cores need faster ram to feed it. Enter Broadwell. Instead of looking for ever higher performance ram, would the eDRAM negate that need? 128MB is plenty big enough for existing tasks I'm participating in.
With the system assembled and running, the first test was to see what the CPU does when it is not limited. I ran small tasks and found Broadwell IPC was 6% down from Haswell. Can this be right? Thinking more, could the smaller L3 cache at 4MB be hindering it vs. a Haswell i5 at 6MB? The tasks being run had a data set of 1MB, so with 4 cores active, this is right on the limit of Broadwell i5. Ok, I ran both systems with only 3 tasks instead of 4. The difference remained. For whatever reason, it seems like Broadwell is reduced in IPC compared to Haswell.
But that isn't why I got Broadwell, and put bigger tests on it. I used the Prime95 built in benchmark to check out core scaling, as more cores are active, more demands are made of ram. Sure enough, I saw better core scaling. That is, running 4 tasks the throughput peaked about 3.8x that of a single task, whereas my fastest Skylake with high speed DDR4 was only scaling to about 3.3x. However the Skylake is also much faster in peak core performance so it would be more demanding.
I saw I could overclock the eDRAM. I turned it up from default 1800 to 2200 and... it made no significant difference. I was rather hoping that if bandwidth was the limit, this would result in a gain. I don't know if there is some other bottleneck working here.
What about different ram configurations. In short, for combined tasks which fit within the eDRAM, the ram performance didn't matter. Only once that was exceeded, did you see the ram performance affect throughput.
And last but not least, how about overclocking the core? The cooler isn't the best but the mobo by default did some kind of enhanced turbo and run 4 cores at 3.5 GHz. Fixing it at default voltage, I went to 4 GHz which was bench stable. Temps were into the 80's C so I didn't want to push harder or higher voltage. The bench results were promising, coming in only 5% down compared to my 4.2 GHz Skylake with high speed ram. The drawback was, when I tried running real work the system wasn't stable and I had crashes. So I've returned to 3.5 GHz and haven't tried tinkering again since then. I will eventually revisit this and see if I can find a balance between temps and clock+voltage.
I've left it like that for now. With better cooling and a tad more voltage I think it would give Skylake a run for its money in memory demanding tasks, at lower cost. Although the respective overclockable i5 CPUs are similar in price, the Broadwell mobos tend to be cheaper, plus you don't need to get high speed DDR4 ram. Any cheap DDR3 would suffice. Right now I have a single stick running in it, as I stole the other stick to temporarily resurrect the i3 I had removed to build this with.
