The Intel Arc B580 kicks off the next-generation GPU releases, arriving a month or more ahead of the competition — or alternatively, a year and a half behind the competition. But it tackles the budget to mainstream audiences with a $249 price tag, upgraded Battlemage architecture, and much-improved drivers compared to the first round of Arc GPUs. How does Arc B580 fare against the best graphics cards? Let’s find out.
We covered many of the details of the Battlemage architecture with the Arc B580 and B570 announcement earlier this month. Intel provided an early tease of performance, claiming 10% higher gaming framerates than Nvidia’s competing RTX 4060, with a $50 lower price tag. Now it’s time to peel off the wrapping and see what the second round of dedicated Intel Arc graphics cards delivers.
Swipe to scroll horizontally
Graphics Card | Arc B580 | Arc B570 | Arc A770 16GB | Arc A750 | Arc A580 |
---|---|---|---|---|---|
Architecture | BMG-G21 | BMG-G21 | ACM-G10 | ACM-G10 | ACM-G10 |
Process Technology | TSMC N5 | TSMC N5 | TSMC N6 | TSMC N6 | TSMC N6 |
Transistors (Billion) | 19.6 | 19.6 | 21.7 | 21.7 | 21.7 |
Die size (mm^2) | 272 | 272 | 406 | 406 | 406 |
Xe-Cores | 20 | 18 | 32 | 28 | 24 |
GPU Cores (Shaders) | 2560 | 2304 | 4096 | 3584 | 3072 |
XMX Cores | 160 | 144 | 512 | 448 | 384 |
Ray Tracing Cores | 20 | 18 | 32 | 28 | 24 |
Boost Clock (MHz) | 2850 | 2750 | 2400 | 2400 | 1700 |
VRAM Speed (Gbps) | 19 | 19 | 17.5 | 16 | 16 |
VRAM (GB) | 12 | 10 | 16 | 8 | 8 |
VRAM Bus Width | 192 | 160 | 256 | 256 | 256 |
L2 Cache | 18 | 18 | 16 | 16 | 16 |
Render Output Units | 80 | 80 | 128 | 128 | 128 |
Texture Mapping Units | 160 | 144 | 256 | 224 | 192 |
TFLOPS FP32 (Boost) | 14.6 | 12.7 | 19.7 | 17.2 | 10.4 |
Peak TFLOPS FP16 (INT8 TOPS) | 117 (233) | 101 (203) | 157 (315) | 138 (275) | 84 (167) |
Bandwidth (GB/s) | 456 | 380 | 560 | 512 | 512 |
TBP (watts) | 190 | 150 | 225 | 225 | 185 |
PCIe Interface | x8 PCIe 4.0 | x8 PCIe 4.0 | x16 PCIe 4.0 | x16 PCIe 4.0 | x16 PCIe 4.0 |
Launch Date | Dec 2024 | Jan 2025 | Oct 2022 | Oct 2022 | Oct 2023 |
Launch Price | $249 | $219 | $349 | $289 | $179 |
Online Price | $250 | $220 | $230 | $200 | $170 |
Looking at the old and new Arc GPUs doesn’t tell the whole story. On paper, the B580 looks decidedly less powerful than the existing A750, never mind the full-fat A770 16GB card. (We’ve omitted the A770 8GB as it never really gained any traction and hasn’t been on sale anywhere we could find for a year or more.)
The B580 ‘only’ has 20 Xe-cores, compared to 28 on the A750 and 32 on the A770 — even the lowly A580 has 24 Xe-cores. You’d be wrong to assume that the only benefit from Battlemage will be the higher boost clocks. And while we’re on the subject of clock speeds, please note that Intel lists official “Graphics Clocks” that are very conservative, but our testing usually has the GPUs running at or very near the maximum boost clock, so we’ve used those numbers for the tables.
Battlemage has some major architectural design changes. Intel shared this slide showing low-level benchmarks that target specific graphics workloads, and you can see some of the largest improvements.
On the far left are compute and draw “execute indirect” bars. Alchemist didn’t support XI in hardware and had to rely on software (driver) workarounds. Battlemage gets a huge improvement by adding native hardware support for the feature. Mesh and vertex throughput also see large improvements of 2X or more, as does sampler feedback. Ray tracing also sees a 1.5X to 2.1X improvement per Xe-core.
Another big change with Battlemage is the move to native SIMD16 (Single Instruction Multiple Data, 16-wide) instructions, compared to Alchemist’s SIMD32 units. That will improve the GPU utilization, as there are a variety of workloads where it’s more difficult to find 32 pieces of data that all need the same instruction. Overall, Intel says Battlemage delivers 70% more performance per Xe-core.
If you do the math, a 20 Xe-core B580 should behave roughly on the level of a 34 Xe-core Alchemist chip (which doesn’t exist). Except we do need to factor in clock speeds as well, and the B580 looks to clock on average around 20% higher than the A770. That works out to around 28% more performance in the real world, give or take.
That’s why the TFLOPS and TOPS figures don’t really matter much. You can compare within an architecture, and the numbers are more meaningful, but when you go to different architectures, all bets are off. We’ve seen that for over a decade with Nvidia and AMD GPUs, so this is nothing new. It’s just something to keep in mind. The B580, with a theoretical 14.6 TFLOPS FP32 and 233 TOPS INT8, will generally beat the A770 with its theoretical 19.7 TFLOPS FP32 and 315 TOPS INT8 of compute.
There are other changes in specifications. Alchemist had up to 16MB of L2 cache, and that’s bumped to 18MB for Battlemage — so there’s not a “huge” L2 or L3 as we’ve seen with the current generation Nvidia and AMD GPUs. The L1 cache for Battlemage is 50% larger than on Alchemist, however.
There’s also the memory interface and capacity. The A770 is most commonly shipped with 16GB of GDDR6 memory on a 256-bit interface, delivering 560 GB/s of bandwidth. But the A750 and A580 cut the capacity in half with 8GB, and slightly lower clocks resulted in 512 GB/s of bandwidth. Battlemage mixes things up with a 192-bit interface and 12GB of capacity, alongside 456 GB/s of bandwidth, thanks to using higher-clocked GDDR6 memory. That should be enough for most use cases, as well as gaming at 1440p and lower resolutions.
Last and perhaps least, the Arc B580 and B570 come with a PCIe 4.0 x8 interface. That’s half of what Intel provided with the Alchemist GPUs (except for the A380 and A310), and PCIe Resizeable BAR support is still required for optimal performance. As for the interface width, it shouldn’t be a problem as data gets sent to the GPU and then stays there for computationally intensive workloads.
- MORE: Best Graphics Cards
- MORE: GPU Benchmarks and Hierarchy
- MORE: All Graphics Content