ADSP-BF607BBCZ-5
Product Specifications
ADSP-BF607BBCZ-5
Analog Devices — Analog & Mixed-Signal ICs
Product Overview
The ADSP-BF607BBCZ-5 from Analog Devices is a high-performance analog / mixed-signal IC providing precision signal acquisition, conversion, and conditioning. Analog Devices (ADI) is a world leader in high-performance analog, mixed-signal, and DSP ICs for instrumentation, healthcare, automotive, and communications.
Superb Automation partners with leading brands including Qorvo, TI, ADI, XILINX, NVIDIA, Intel, ST, MACOM, Renesas, and Coilcraft to provide complete BOM and supply-chain solutions.
Key Features
- High-resolution data conversion with excellent SNR and THD
- Wide dynamic range and high linearity across operating conditions
- Integrated voltage reference with low temperature drift
- Flexible digital interface — SPI, I2C, or parallel options
- Industrial/automotive temperature range with guaranteed specs
- Low power consumption with power-down modes
Key Specifications
| Resolution | bits |
| Channels | — |
| Sample Rate | kSPS |
| Interface | — |
| Supply Voltage | V |
| Operating Temp | °C |
| Package | — |
| INL | LSB |
| SNR | dB |
| THD | dB |
Applications
ADSP-BF607BBCZ-5: two 500 MHz Blackfin cores, 2000 MMACs and 552 KB on chip
The ADSP-BF607BBCZ-5 is an Analog Devices dual core Blackfin processor in a 19 mm by 19 mm chip scale BGA, with each core running to 500 MHz and 1000 MMACs for a total of 2000 MMACs. On chip memory is 552 KB: 148 KB of level one SRAM per core with parity, plus a 256 KB shared level two block with error correction. The peripheral set is wide for an embedded processor: three parallel peripheral interfaces up to 24 bits wide with ITU-R BT.656 support, three synchronous serial ports, USB 2.0 high speed on the go, a removable storage interface, a CAN controller, two 10/100 Ethernet MACs with IEEE 1588, four link ports, two SPI ports, two UARTs, two two-wire interfaces and eight timer counters. A CRC engine, a fault management unit and parity and error correction on the internal memories are provided for safety related work, and both commercial and industrial temperature grades are offered. It is on the analog and mixed-signal IC line card.
Two thousand MMACs means two multiply accumulates per cycle in each core
Dividing 2000 MMACs by the 1 GHz of aggregate clock gives two multiply accumulate operations per cycle per core, or 4 GFLOP in the usual counting. That number is what separates this part from a microcontroller with a single cycle multiplier: a 1024 point complex FFT needs on the order of five times N log2 N multiply accumulates, about 51,200, which at 2000 MMACs is 26 us before memory movement is counted. Whether that is enough depends on how much of the data is in the 552 KB of on chip memory, which is why the memory map matters more than the headline rate.
| Parameter | Vendor figure | Derived figure |
|---|---|---|
| Cores | 2 Blackfin at 500 MHz | 2000 MMACs total, 2 per cycle per core |
| Arithmetic rate | 1000 MMACs per core | about 4 GFLOP in floating point counting |
| On chip memory | 148 KB L1 per core plus 256 KB shared L2 | 552 KB in total, 2.16 times the shared block alone |
| Memory protection | parity on L1, error correction on L2 | a single bit error is corrected, not just detected |
| Video interfaces | 3 parallel ports up to 24 bits | BT.656 and direct TFT connection |
| Serial ports | 3 synchronous, I2S and TDM capable | six audio channels per port at 48 kHz |
| Connectivity | USB high speed, 2 Ethernet MACs, CAN, 4 link ports | 1588 timestamping on both MACs |
| Safety features | CRC engine, fault management unit | the memory and bus are checked in hardware |
| Package | 19 mm by 19 mm CSP BGA | 361 mm2 of board area for the whole processor |
Five hundred and fifty two kilobytes is the number that decides the partition
A level one block of 148 KB per core holds a 32 bit sample buffer of about 37,000 words, which at 48 kHz is 0.77 seconds of mono audio, while the shared 256 KB is 65,536 words and holds the tables and the shared queues. A build that wants each core to run its own filter chain puts the coefficients and the delay lines in its own level one memory and passes results through the shared block. A build that wants one core to do the physical layer and the other to do the protocol normally does the opposite, and the split is what the memory map allows rather than what the algorithm would prefer.
Selection notes
- Both MACs carry IEEE 1588 timestamping, which is what makes the part usable as a node in a synchronised industrial network.
- The parallel ports connect image sensors and TFT panels directly, removing a video decoder from the bill of materials.
- Parity and error correction are on the internal memories and not on an external bus, so a safety case has to cover the external memory separately.
- The industrial temperature grade is a separate order code rather than a suffix, so the part number has to be fixed before the qualification run.
Questions buyers ask
Is one core enough for a 1024 point complex FFT at audio rates? Yes by a wide margin, and the second core is usually spent on the protocol or on a second channel rather than on the same transform.
What tools are needed? A JTAG emulator and the vendor's cross development tools, plus the boot stream written through the same port or from flash.
Is the part still in production? Yes, the vendor lists it as production, and it is a long lived industrial part rather than a consumer device, but a socket that lives for a decade still needs a stock plan.
Digital parts are sourced alongside their memory and interface devices on BOM shortage recovery, and the lead time question is handled on component lead time solutions.
552 Kilobytes and Two Frames: What the Memory Budget Actually Holds
The ADSP-BF607BBCZ-5 carries 552 KB of on chip memory, 148 KB of level one SRAM per core with parity plus a 256 KB shared level two block with error correction. That sounds generous until a frame size is put against it, and the arithmetic of what fits decides whether the board needs external memory and, if it does, whether the two 10/100 Ethernet MACs and the parallel ports can keep up with it.
One VGA frame at 16 bits does not fit
A 640 by 480 frame at 16 bits per pixel is 614,400 bytes, 600 KB. That is more than the whole 552 KB of on chip memory, so a single frame cannot be held even if both cores give up their level one. A 320 by 240 frame at 16 bits is 153,600 bytes, 150 KB, which fits in one core's level one with 2 KB to spare, and a 320 by 240 frame at 8 bits is 75 KB, which leaves half of the level one free for a second buffer.
The shared block is 256 KB and it is the one both cores can see
The 256 KB level two block is the only memory both cores address directly, and it is what a producer consumer arrangement uses. Two 320 by 240 8-bit frames are 150 KB and fit with 106 KB left for the coefficients and the queues. A pipeline that needs three frames for a rolling average is 225 KB and leaves 31 KB, which is tight for a queue but workable. The level one blocks are per core and private, so a shared buffer that is copied into level one costs a copy per frame and that copy is measurable at 500 MHz with a 16 bit bus.
Two thousand MMACs is four multiply accumulates per cycle
2000 MMACs across two 500 MHz cores is two multiply accumulates per core per cycle, four in total. A 1024 point complex fast Fourier transform needs five times N times log2 N, which is 51,200 multiplies, so it takes 25.6 us at the aggregate rate or 51.2 us on one core. A 4096 point transform is 245,760 multiplies, 123 us at the aggregate rate. A frame rate of 30 per second gives 33 ms per frame, so a 1024 point transform per frame uses 0.08 percent of the machine and a 4096 point transform uses 0.4 percent. The processor is not short of arithmetic for a transform, it is short of memory for the frame.
| Workload | Multiplies | Time at 2000 MMACs | Time on one core |
| 256 point complex FFT | 10,240 | 5.1 us | 10.2 us |
| 1024 point complex FFT | 51,200 | 25.6 us | 51.2 us |
| 4096 point complex FFT | 245,760 | 123 us | 246 us |
| 8 by 8 block match, one frame | about 4.7 million | 2.4 ms | 4.7 ms |
The peripherals decide the external memory question
Three parallel peripheral interfaces up to 24 bits wide, three synchronous serial ports, USB 2.0 high speed on the go, two 10/100 Ethernet MACs with IEEE 1588, a CAN controller, four link ports, two SPI ports, two UARTs, two two-wire interfaces and eight timer counters. A camera on one parallel port, an FPGA on a link port and a host on Ethernet is a normal arrangement for this part, and in that arrangement the external memory bus is what carries the frames while the on chip memory carries the working set. The 19 mm by 19 mm chip scale BGA has the balls for a wide memory bus, which is why the package and not the core count is often the reason this part is chosen over a smaller one.
Parity and error correction are a procurement reason
Parity on the level one memories, error correction on the level two block, a CRC engine and a fault management unit are the features that put this part on a rail or industrial board rather than a consumer one. They come at no extra cost in cycles and they change what a safety case can claim about undetected memory faults. A board that has already qualified a part without them cannot simply claim the same coverage after a substitution, which is the point to note when a second source is offered on the basis of the core count and the clock rate alone.
The processor is listed with the other parts on the analog and mixed-signal IC page. Reel dates are on the component lead time page, and a build that is short of one device is handled through the BOM shortage recovery route.
Request Quote — In Stock, 100% New & Original, Full Warranty