Alder Lake-S micro-architecture & DRAM Latencies
In my previous article I included a chart that showed the stock CPU and DRAM frequencies with various overclocks to the Efficient Cores and the DRAM. Shortly after that article was posted I was able to lower my latencies across the entire platform even further through various DRAM timings. I did not update or write another article regarding my performance increases since I quickly shifted to overclocking the Alder Lake processor. So I decided to run all of the latency benchmarks again with my CPU overclocks in the chart below. DRAM is a very important factor in determining and increasing performance for me personally. While trying to find a balance I like to start with DRAM first then tackle other areas on the processor to overclock. I have overclocked the Performance Cores (5.1GHz) and Efficient Cores (4.1GHz) while tightening the extended DRAM timings. I have also overclocked the Ring Bus to 4.0GHz. So while the chart will show that “ (40-39-39-76: 2T) “ for the DRAM, the extended timings have been tweaked and trust me there are many timings to tweak. So now let’s take a look at the micro-architecture latencies.
Starting with the DRAM Bandwidth I have increased my bandwidth from 75.7 GB\s (DDR5-4800Mhz) to 89.5GB \s (DDR5-5600Mhz + tight extended timings) which is an 18.22% increase. In my previous article you can see that I hit 89.4 GB\s and now I have managed to increase that to 89.5 GB\s this time around. Moving down to the “DRAM Latency (Small dataset)” column we see the latency decrease from 66.43 ns to 53.56 ns. That is a 21% difference in latency. I was able to lower that latency from my previous DDR5-5600 overclock in the last article (57.07 ns) to 53.56 ns with tighter extended timings. This is roughly 6% to 7% performance increase and remember when it comes to latency lower is better. Now looking at the larger datasets in DRAM we see a very nice decrease from the stock results (75.40 ns – 4800MHz) down to only 63.56 ns on average. From stock results to my current DDR5-5600Mhz results that is an extremely nice decrease of 11.84 ns. That is a 19% difference in latency. Continuing with the larger datasets (Large DRAM dataset); comparing my latest results to my previous DDR5-5600Mhz (E-Core @4.0GHz) results we see a decrease from 68.45 ns to only 63.56 ns. That is an additional decrease of nearly 5.0 ns over my previous article which is incredible for large datasets in memory. I was not expecting this big of a performance increase by overclocked the DDR5-4800Mhz DRAM especially since its First Gen DDR5. I have combined the average of all the P-Cores and E-Cores in their respective columns instead of listing each core individually this time around. The Performance Cores shows a decrease of 13% when compared to the stock settings. The Efficient Cores shows a decrease of 12%. The Efficient Cores have always shown great low-latency to DRAM and could lead to some great performance results in future micro-architectures from Intel. Now that we have seen how much performance I have gained over the past few weeks let’s see how well my overclocks perform in various synthetic benchmarks.