New GPU contender? and its a startup . >100Gb memory ?

[Simmed]

i didnt know this was around
apparently they have been at it for 6? years ?
and it is a very interesting board
sodimm slots, potentially over 100Gb memory ?


00:00 - Bolt Graphics GPU
02:52 - Dreams of a Gaming GPU
03:40 -- attack the content creators first
05:08 - A Peculiar Design
10:54 - Bolt Zeus GPU Architecture
12:14 --  he says the controller cpu onboard is licensed stuff and he cant say who it is xilinx ?
18:55 - Raster, Ray Tracing, and Path Tracing
20:43 - Negative RT Sentiment and Strategy
22:54 - The Price Dream
23:41 - Cache and Memory Limitations
27:12 - Memory Impact on Bolt
28:09 - Functional Demo




https://www.tomshardware.com/tech-industry/semiconductors/bolt-graphics-tapes-out-its-first-zeus-gpu-test-chip-on-tsmc-12nm

The test chip was designed into TSMC's 12nm FFC process and will be used for customer benchmarking ahead of a planned production ramp in Q4 2027. Bolt says it’s targeting companies in the HPC, electromagnetic simulation, and graphics rendering markets, which it values at over $55 billion combined, that "can’t afford the level of compute needed for their applications."

https://www.tomshardware.com/pc-components/gpus/startup-claims-its-zeus-gpu-is-10x-faster-than-nvidias-rtx-5090-bolts-first-gpu-coming-in-2026
last year

There is one catch: Zeus can only beat the RTX 5090 GPU in path tracing and FP64 compute workloads. It's not clear how well it will handle traditional rendering techniques, as that was less of a focus. In speaking with Bolt Graphics, the card does support rasterization, but there was less emphasis on that aspect of the GPU, and it may struggle to compete with the best graphics cards when it comes to gaming. And when it comes to data center options like Nvidia's Blackwell B200, it's an entirely different matter.

even if the compute power does not compete with 5090
just the 128Gb memory alone will stick
but since it is AI season ... *cough* ... how much will a 128Gb costs ?

currently if i get a strix 128Gb
the 128Gb option which is chinese LPDDR is extra USD1.5k
so if they use that + board cost ... i think ball park price for this could be USD3k++ ?

[tom66]

If it's good at FP64, it probably won't be that good at FP16, FP8, INT8, INT4... that AI needs.  But maybe its architecture allows for that, hard to say.  They so far have only focused on FP64 performance which suggests they are not going for AI workloads but HPC.

Edit: just seen the spec sheet in your attachment, so it's roughly 1/3rd the INT8 of an RTX5090.  Puts it about halfway between 5070 and 5080 by my maths.

[Simmed]

If it's good at FP64, it probably won't be that good at FP16, FP8, INT8, INT4... that AI needs.  But maybe its architecture allows for that, hard to say.  They so far have only focused on FP64 performance which suggests they are not going for AI workloads but HPC.

Edit: just seen the spec sheet in your attachment, so it's roughly 1/3rd the INT8 of an RTX5090.  Puts it about halfway between 5070 and 5080 by my maths.

i see
i didnt know about the specializations at certain FPxx
well they say the production is around end 2027
by then, i think apple will have come around with maybe M5 + another unified memory of *me thinks* 1Tb memory