Intel graphics cards: Arc B580 and B570 with "Battlemage" Xe2 architecture

Intel's new "Battlemage" graphics chip architecture made its debut back in September. Back then, it was still – compared to plug-in cards – weak integrated graphics in the Core Ultra 200V Lunar Lake mobile processors.

The first graphics cards with the revised architecture for the PCI Express slot of desktop PCs will go on sale from December 13. These cards should be around 10 percent faster than Nvidia's GeForce RTX 4060 in WQHD resolution with 2560 × 1440 pixels and also outperform their predecessor, the Arc A750, by around a quarter. The 12 GB graphics memory, which is 50 percent larger, also plays an important role, especially in resolutions above Full HD.

As with its predecessor, the Alchemist (Arc A series), Intel is launching a "Limited Edition" version of the B580 with a recommended retail price of 249 US dollars. Intel was unable to give us a euro price 10 days before the sales launch, but including taxes and minus the exchange rate, it could amount to 269 euros in Germany. In Germany, cards from Acer, Asrock and Sparkle will also be available, while Gunnir, Maxsun and Onix will also be involved internationally.

The somewhat weaker B570 with 10 GB of graphics memory, which will follow from January 16, will be available from 219 US dollars (approx. 239 euros with tax) only from partners. Intel has not yet provided its own performance assessment. Based on the key data, it should still manage to exceed the level of the previous generation in the form of the Arc A750, but will have a hard time beating the RTX 4060.

The Arc B580 has a thermal design power of 190 watts, while the smaller B570 manages with 150 watts. Nevertheless, both are equipped with a classic eight-pin power connector to increase the maximum 75 watts from the PCIe 4.0 x8 connection and to absorb load peaks. When asked by c't, Intel stated that it had also worked on one of the biggest criticisms of the previous generation, the power consumption in idle mode. In many cases, they are now competitive in this respect, but still lag behind in certain combinations of high-resolution display and high refresh rate.

Performance show

So far, only the performance values from Intel's preliminary information are available. Our own measured values from c't will follow at the start of sales in the magazine and on heise online.

Intel is mainly pitting the B580 with 12 GB against its predecessor Arc A750, Nvidia's GeForce RTX 4060 and AMD's Radeon RX 7600 – all with 8 instead of just 12 GB of graphics memory. The higher the resolution, the greater the memory and the higher the memory transfer rate in some cases. So it's no wonder that Intel has set the WQHD resolution of 2560 × 1440 pixels as the target.

In comparison with older, more common graphics cards such as the GeForce GTX 1060 or 1660 Super, the newcomer is of course clearly ahead. Intel states a factor of 2.79 compared to the former and just under 2 compared to the 1660 Super.

Compared to the Arc A750, the new B580 is said to be 24 percent ahead on average. The frame rates quoted by Intel range from 48 frames per second (fps) in Dragons Dogma 2 and The Last of Us Part I to several hundred fps in Monster Hunter Rise and League of Legends. In some of the 47 games, however, you have to resort to the help of the AI-supported upscaler XeSS to achieve smooth frame rates, such as the 60 fps in the shooter Robocop Rogue City or the 81 fps in Starfield. In percentage terms, the range extends from a tie in the CPU-limited League of Legends and DOTA 2 to 78 percent in Fortnite –, the latter again only with the help of XeSS.

Compared to a GeForce RTX 4060, Cyberpunk 2077 is up to 43 percent faster, Intel stated in the preliminary information. Other games such as Control, Diablo IV or F1 24 were between a quarter and a third faster – where it is questionable, but without specifying whether ray tracing was active or switched off.

With ray tracing, things are likely to be much tighter, as Intel sees the self-calculated performance-per-dollar advantage over the 4060 shrink from 32 to 25 percent.

Xe2 – more memory, more AI

The chip developed under the code name BMG-G21 is manufactured by Taiwanese chip manufacturer TSMC using the proven N5 process and, according to Intel, measures 272 mm². It contains 19.6 billion transistors –. Despite the improvements, both parameters are significantly lower than those of its predecessor, the ACM-G10. This is because the predecessor, the Arc A770, had up to 32 Xe cores, while the B580 GPU only has 20 active Xe cores (18 in the B570) and is therefore one number lower. However, it is still unclear when a corresponding BMG-G10 will be available. Intel also left it unclear whether the GPU with Xe cores is the full version or whether there could possibly be a sixth render slice. However, the published schematics are patient render images that may or may not correspond to the actual unit distribution.

With a new arrangement of the execution units, a switch to a more efficient SIMD format with less control logic in favor of higher throughput and other improvements, Intel claims to have achieved 70 percent more performance per Xe core and increased the performance per watt by 50 percent.

In addition to improvements to the architecture of the XE cores – compared to the Arc 140V in the Core Ultra 200V, other changes have also been made – The Battlemage GPUs now also support additional data formats in the XMX units, which are primarily intended to handle matrix-matrix multiplications for AI applications. Everything that could be useful is included, right down to 2-bit integer numbers (INT2) and the TF32 tensor format widely used by Nvidia.

Intel has also shown its own AI benchmarks that measure the number of tokens per second in four different large-language models. Using the ONNX runtime with DirectML, they claim to be almost 40 to a good 50 percent faster than the Nvidia cards. However, whether these were also measured with ONNX or under Nvidia's in-house and often faster CUDA environment remains to be seen.

Intel Arc B580/B570 in comparison
	Arc B570	Arc B580	Arc A 750	Arc A 770 LE
Chip	BMG-G21	BMG-G21	ACM-G10	ACM-G10
Manufacturing process	TSMC N5 (272 mm², 19,6 billion transistors)	TSMC N5 (272 mm², 19,6 billion transistors)	TSMC N6 (406 mm², 21,7 billion transistors)	TSMC N6 (406 mm², 21,7 Mrd. Transistoren)
Render-Slices	5	5	7	8
Xe-/RT-Cores	18/18	20/20	28/28	32/32
Shader-Cores	2304	2560	3584	4096
AI units (XMX)	144	160	448	256
Throughput (KI-TOPS, INT8)	203	233	229	262
Throughput (TFLOPS, FP32)	11,5	13,7	14,7	17,2
Chip clock	2500 MHz	2670 MHz	2050	2100 MHz
Graphics memory	10 GByte	12 GByte	8 GByte	16 GByte
Storage interface	160 Bit	192 Bit	256 Bit	256 Bit
Memory transfer rate	380 GByte/s	456 GByte/s	512 GByte/s	560 GByte/s
Total Board Power	150 watts	190 watts	225 watts	225 watts
PCIe	4.0 x8	4.0 x8	4.0 x16	4.0 x16
Displayengine	3 × DP2.1 (UHBR13.5), 1 × HDMI 2.1	3 × DP2.1 (UHBR13.5), 1 × HDMI 2.1	3 × DP2.0 (UHBR10), 1 × HDMI 2.0	3 × DP2.0 (UHBR10), 1 × HDMI 2.0
Videoengine	u.a. AV1, HEVC, VP9, XAVC-H	u.a. AV1, HEVC, VP9, XAVC-H	u.a. AV1, HEVC, VP9	u.a. AV1, HEVC, VP9
Available from	16.01.25	13.12.24	12.10.22	12.10.22
RRP (plus VAT)	219 US-$	249 US-$	k.A.	k.A.

(csp)