Perfomrance comparison of GPU boards from AMD as of Oct 2009

We have tested GPU borads from AMD with our test program implementing a simple N² force evaluation algorithm. A nominal performacne of each board is shown in the following table.

board	arch	clock	memory	No.SC	SPmul-add	DPadd	DPmul	BW
HD4850	RV770	625 MHz	DDR3 662 MHz 256bit	800	1040 GFLOPS	208 GFLOPS	104 GFLOPS	63.6 GB/sec
HD4870	RV770	750 MHz	DDR5 900 MHz 256bit	800	1200 GFLOPS	240 GFLOPS	120 GFLOPS	115.2 GB/sec
HD4770	RV740	750 MHz	DDR5 800 MHz 128bit	640	960 GFLOPS	192 GFLOPS	96 GFLOPS	51.2 GB/sec
HD5870	RV870	850 MHz	DDR5 1.2 GHz 256bit	1600	2720 GFLOPS	544 GFLOPS	272 GFLOPS	153.6 GB/sec

The program is basically same as our demo program posted here. The demo program should work with 5870 but I did not test it.

Result

Note

• We count one-force-interaction as 38 FP operations. This is a traditional flop-count in this field.
• At large N, 5870 shows 2.2x better performance than 4870. It reaches ~ 2.2 Tflops.
• 4850 and 4770 show identical performance. Memory BW is not critical at all in this test program.
• I add a new result "5870_opt". It is obtained with a new optimized kernel. It boosts the performance ~ 18 %.

Analysis of VLIW instructions

Our kernel is directly written in IL. After compilation into a machine code, I analyze the generated VLIW instructions. Results are obtained with fglrx 8.65.4 (Catalyst 9.9) and CAL 1.4beta.

Both RV770 and RV870 archtecture has 5-way VLIW units at its heart. Depnding on detailed computations, it is supposed that the device driver (or internal compiler?) tries to fill 5-slots as much as possible. The first column shows a number of occupied slots and the second column shows a number of corresponding VLIW instructions appeared in the kernel. The third column indicates a fraction of VLIW instructions with a given number of occupied slots. The last row presents a total numberf VLIW instructions in the kernel.

5870_opt kernel

#slots	#inst	fraction
1	1	2 %
2	2	3 %
3	4	6 %
4	16	24 %
5	43	65 %
total	66

old kernel

#slots	#inst	fraction
1	12	15 %
2	6	7 %
3	4	5 %
4	16	20 %
5	43	53 %
total	81

Almost 90% of the time, "5870_opt kernel" runs at maximum efficiency (here, I mean more than 4 slots are occupied by some operations). It is compared to 73% in the case of our old kernel.

Comparison with 5870_opt kernel

Benchmark system

CPU	Core2 E8400 3.0 GHz
MB	Asus P5EWS
Memory	DDR2 800 1 GB x 4
Power unit	Schythe CorePower3 600W
OS	Ubuntu 8.04 x86_64 2.6.24-23-generic
Catalyst	9.9
CAL version	1.4beta

This system looks old but PCIe bus on P5EWS is faster than any other systems we have so far.