Slower than Windsor? Brisbane in Synthetic Benchmarks

Same-frequency CPUs on the Brisbane and Windsor cores shouldn’t differ much in performance as they incorporate the same number of transistors and, as AMD claims, don’t differ at all in their micro-architecture. However, there are publications on the Web that argue that point. Let’s check it out.

First, we compared the speed of the computing units of processors on the Brisbane and Windsor cores (the latter had a total of 1MB of L2 cache). For the comparison to be correct, we set the clock rate of both CPUs at 2.4GHz. We tested them in the CPU benchmarks from the SiSoftware Sandra XI suite.

	Brisbane 2.4GHz	Windsor 2.4GHz
Sandra XI, Arithmetic ALU	17489	17480
Sandra XI, Arithmetic SSE3	14786	14788
Sandra XI, Multi-Media Integer MMX/SSE	44863	44897
Sandra XI, Multi-Media Floating-Point SSE2	49339	49335

The main computing units of the two CPUs indeed provide the same performance. The difference fits within the measurement error range.

The CPU benchmarks from SiSoftware Sandra XI do not depend on the memory subsystem speed. They are indicative of the “pure?performance of a CPU. But in real-life applications the speed at which the CPU is receiving data from memory has an effect on performance, too. So, we measured the bandwidth and latency of system memory as well as the latency of the L2 cache. We installed dual-channel DDR2-800 memory with timings of 4-4-4-12-1T for this test.

	Brisbane 2.4GHz	Windsor 2.4GHz
Sandra XI, L2 Cache Latency, clk	22.7	17.7
Sandra XI, Memory Bandwidth, MB/s	8351	8675
Sandra XI, Memory Latency, ns	107	92

Here’s a strange surprise to you. The L2 cache of CPUs on the new 65nm Brisbane core has higher latency and this increases the latency of the memory subsystem in general. And this ultimately reduces the overall memory subsystem bandwidth.

The discouraging results produced by SiSoftware Sandra XI are confirmed by other synthetic benchmarks of the memory subsystem.

	Brisbane 2.4GHz	Windsor 2.4GHz
CPU-Z, L2 Cache Latency, clk	20	12
CPU-Z, Memoy Latency, clk	115	108
ScienceMark 2.0, L2 Cache Latency, clk	20	13
ScienceMark 2.0, Memory Latency, clk	114	106
ScienceMark 2.0, Memory Bandwidth, MB/s	7619	8202
EVEREST 2006, Memory Read, MB/s	7816	8044
EVEREST 2006, Memory Write, MB/s	6833	6932
EVEREST 2006, Memory Copy, MB/s	7914	8152
EVEREST 2006, Memory Latency, ns	51	48.7

So, there can’t be any doubt: the L2 cache has become slower in the new core for Athlon 64 X2 CPUs. As a result, the new CPUs work slower with data in memory than their 90nm predecessors, which may show up in real-life applications, too. But the organization of the L2 cache hasn’t changed: 16-way associativity with a line length of 64 bytes.

So, we should seek elsewhere for the root of the problem. AMD has commented that the latency of the L2 cache has increased as a consequence of the engineers having left a reserve for enlarging the cache in the future. This doesn’t sound convincing to us, however. First, AMD’s plans don’t contain any information about enlarging the L2 cache even on the transition to the K8L micro-architecture. Second, Windsor-core CPUs with a 2x1MB L2 cache do not differ in cache performance from their counterparts that are equipped with a 2x512KB L2 cache. So, it is not yet clear to us why the speed parameters of the cache have changed.

The increased latency of the cache memory is not the only problem that can have a negative effect on performance of Brisbane-core CPUs. Another problem is about the fractional CPU frequency multipliers ?the real memory frequency has been reduced in some modes because the default CPU frequencies now change with a step of 100MHz. In CPUs with the K8 micro-architecture the memory frequency is actually based on the CPU frequency and an integer divider. We’ve met this problem before, but it has grown worse with the new CPUs. To illustrate our point, here is a table that shows the real memory frequency in the different modes of the memory controller integrated into the CPU.

	Processor frequency, MHz
	2000	2100	2200	2300	2400	2500	2600	2800
DDR2-800	800	700	733	767	800	714	743	800
DDR2-667	667	600	629	657	600	625	650	622
DDR2-533	500	525	489	511	533	500	520	509

That’s not a catastrophe, of course. There are very poor memory modes with CPUs that have integer frequency multipliers, too. Yet you should be aware of this thing because the real memory frequency is often much lower than the expected one with CPUs that have fractional multipliers. This has a negative effect on the overall system performance, of course.