New GPU is 10 times faster than the RTX 5090 | Hacker News Reader

New GPU is 10 times faster than the RTX 5090

xda-developers.com

6 pointsqwertox5 comments

Source article: https://www.servethehome.com/bolt-graphics-zeus-the-new-gpu-...

Marking fluff, no hardware out, not designed to compete with a 5090 as it is a data center GPU.

But it doesn't have to be. It's RISC-V on chiplets combined with IO-chiplets. Skimming the stuff, one of those would suffice to run rings around current iGPUs of APUs at around 120Watts. While having some LPDDR5X soldered on as VRAM, and two SO-DIMM Slots for further expansion of that. Basically at least twice as fast as anything else, while only using half of the Watts.

Now the 'envisioned' price would be interesting, for the different SKUs.

Oh, and giving some of them to hotz/geohot, too!

LPDDR5X is slower than GDDR7 present on the 5090.

This card lacks hardware-based rasterization and other tech that makes a graphics card a useful graphics card for gaming.

This is an HPC card at most, when it isn't vaporware.

Then why has the (assumed) rendering of the board with the low-end single chiplet Zeus 1c26-032

(which btw. won't have 2.25 TB VRAM and/or 800GbE, that would be the Zeus 4c26-256)

HDMI 2.1b and Displayport 2.1a?

The article is sloppy, in that it tries to compare a whole range of options to one single architecture and implementation.

For instance, it isn't obvious how the memory-bandwith of 32GB LPDDR5x + up to 128GB DDR5 via S0-DIMM is calculated for the lowest, single chiplet model 1c26-032.

Says 363.3GB/s

Then the 'mid-range' with two chiolets at 250Watts, Zeus 2c26-064/128 with either 64MB or 128MB LPDDR5x per chiplet and up to 320/386GB via 4x SO-DIMM DDR5. (still got HDMI + DP)

Says 725GB/s

Their 'high-end' Zeus 4c256 at 500Watts with 4 chiplets, 256GB LPDDR5x + up to 2TB via 8x SO-DIMM DDR5. (no more HDMI or DP, massive networking instead)

Says 1.45TB/s

All the while performing FP64 vector flops rather well. FP32/FP16 not so much. Int16/32 matrix flops look ok.

And clusterable by many means.

Wake me up around 26Q4/27Q1, according to their projected timeline :)

When the prices, software support and first real benchmarks are known :)

Edit: The mysterious 'on-chip-cache' in varying amounts and its relation to the LPPDDR5x are unclear to me :(

It's vaporware. This product doesn't exist thus the manufacture makes up claims. LPDDR5X is slower than GDDR7, no questions. 5090 has >1TB/s memory bandwidth, though only 32GiB of course.

That said, FP64 isn't the precision required for gaming. Again, this card at best is for HPC-type environments, not for the gaming desktop. It may be faster than the 5090 should it actually exist, but the 5090 and this card don't compete in the same space. Just the fact it has SODIMM slots should be telling enough that this isn't a high performance rasterization card.

Reply on news.ycombinator.com