AMD Versal HBM Adaptive SoCs Enter Production

[phoronix]

Phoronix: AMD Versal HBM Adaptive SoCs Enter Production

AMD announced today that their Versal HBM adaptive SoCs have entered production for these newest wares with high bandwidth memory...

https://www.phoronix.com/news/AMD-Versal-HBM-SoCs-Production

[timofonic]

Only $14,995! Limited offer!

Seriously, this is absurdly expensive...

Does it come from the future or what?

[unwind-protect]

Will be interesting to see how much faster this compiles the Linux kernel.

[stormcrow]

Only $14,995! Limited offer!

Seriously, this is absurdly expensive...

Does it come from the future or what?

Not really that surprising for a developer's kit with hardware. Board development kits aren't cheap even for traditional evaluation boards, on the order of several thousand dollars each. But when you start looking at kits for what's essentially HPC systems... yeah, not totally surprising considering high end workstations are often still over 10k USD. These aren't cheap Raspberry Pi equivalents with a compute hat. LLM compute and neural processing requires a crap ton of customized hardware otherwise you're burning several times more electricity for modest gain. This is why an experimental project I was part of back in the 90s abandoned the prospect of using neural nets to pattern match camera photos (pointed at the sky from a balloon) up to visible star patterns to accurately determine detector angles. The hardware didn't exist at the time to do it efficiently even on the ground let alone onboard a balloon gondola floating over Antarctica. It exists today, but it still requires customized hardware to do it efficiently - which is why newer computers are getting neural co-processors even for the client side of neural processing such as the M-class Macs and likely x64 PCs in the near future.

Edit to add: What a lot of people forget (and me too till I sat and thought a moment) is that devel kits like this also include support contracts. Skilled labor isn't cheap, and engineers especially so. A significant chunk of that money is to pay the engineers on call to iron out problems or answer questions. Something you just don't get with cheap hobby boards. Nor if you buy a random bare AMD or Intel CPU for that matter and only in a very limited fashion if you buy a packaged CPU kit.

[kdemello1980]

Not really that surprising for a developer's kit with hardware. Board development kits aren't cheap even for traditional evaluation boards, on the order of several thousand dollars each. But when you start looking at kits for what's essentially HPC systems... yeah, not totally surprising considering high end workstations are often still over 10k USD. These aren't cheap Raspberry Pi equivalents with a compute hat. LLM compute and neural processing requires a crap ton of customized hardware otherwise you're burning several times more electricity for modest gain. This is why an experimental project I was part of back in the 90s abandoned the prospect of using neural nets to pattern match camera photos (pointed at the sky from a balloon) up to visible star patterns to accurately determine detector angles. The hardware didn't exist at the time to do it efficiently even on the ground let alone onboard a balloon gondola floating over Antarctica. It exists today, but it still requires customized hardware to do it efficiently - which is why newer computers are getting neural co-processors even for the client side of neural processing such as the M-class Macs and likely x64 PCs in the near future.

Edit to add: What a lot of people forget (and me too till I sat and thought a moment) is that devel kits like this also include support contracts. Skilled labor isn't cheap, and engineers especially so. A significant chunk of that money is to pay the engineers on call to iron out problems or answer questions. Something you just don't get with cheap hobby boards. Nor if you buy a random bare AMD or Intel CPU for that matter and only in a very limited fashion if you buy a packaged CPU kit.

There's that, and this thing is a beast, hardware-wise.

Consisting of 58G/112 Gbps PAM4 and 32 Gbps NRZ transceivers, the Versal HBM adaptive SoC has very scalable transceivers that deliver up to 5.6 Tbps of serial I/O bandwidth. The 112G PAM4 transceivers allow the industry to roll out 800G and single-lane 100G capable infrastructure now. For 400G ramp and deployment, 58G PAM4 transceivers enable the latest generation interfaces for maximum bandwidth density. For mainstream power-optimized 100G interfaces, 32 Gbps NRZ transceivers are ideal.

It looks like it's got 4 qsfp and 2 qsfp-dd ports, so they're expecting some serious stuff to plug into this thing.