Jan Gray
@jangray.bsky.social
Kilocore RISC-V FPGA accelerators; former Microsoft dev tools architect; Vice-chair RISC-V SoftCPU SIG & Composable Custom Extensions Task Group; blog: https://fpga.org. Cyclist. Let's try kindness. 🇨🇦-🇺🇸
Pinned
Jan Gray
@jangray.bsky.social
· Nov 12
2GRVI Phalanx at Hot Chips 31 (2019): The First Kilocore RISC-V RV64I with High Bandwidth Memory
This week at Hot Chips 31 (2019) I am presenting a status update poster on the work-in-progress GRVI Phalanx Accelerator Kit: 2GRVI Phalanx: Towards Kilocore RISC-V FPGA Accelerators with HBM2 DRAM…
fpga.org
#FPGA #RISCV
2GRVI Phalanx at Hot Chips 31: The First Kilocore RISC-V RV64I with HBM2 High Bandwidth Memory
fpga.org/2019/08/19/2...
2GRVI Phalanx at Hot Chips 31: The First Kilocore RISC-V RV64I with HBM2 High Bandwidth Memory
fpga.org/2019/08/19/2...
Reposted by Jan Gray
=>
"Numerics and AI", Paulius Micikevicius, GPU Mode, Nov 15, 2025 www.youtube.com/watch?v=ua2N...
github.com/gpu-mode/lec...
Stochastic Rounding, Minisymp on Probabilistic Rounding, Sep 1 bsky.app/profile/ogaw...
Survey, Proc of the IEEE, Jun 26 bsky.app/profile/ogaw...
36 pp, 201 ref
MX+, MICRO25
"Numerics and AI", Paulius Micikevicius, GPU Mode, Nov 15, 2025 www.youtube.com/watch?v=ua2N...
github.com/gpu-mode/lec...
Stochastic Rounding, Minisymp on Probabilistic Rounding, Sep 1 bsky.app/profile/ogaw...
Survey, Proc of the IEEE, Jun 26 bsky.app/profile/ogaw...
36 pp, 201 ref
MX+, MICRO25
November 25, 2025 at 3:33 AM
=>
"Numerics and AI", Paulius Micikevicius, GPU Mode, Nov 15, 2025 www.youtube.com/watch?v=ua2N...
github.com/gpu-mode/lec...
Stochastic Rounding, Minisymp on Probabilistic Rounding, Sep 1 bsky.app/profile/ogaw...
Survey, Proc of the IEEE, Jun 26 bsky.app/profile/ogaw...
36 pp, 201 ref
MX+, MICRO25
"Numerics and AI", Paulius Micikevicius, GPU Mode, Nov 15, 2025 www.youtube.com/watch?v=ua2N...
github.com/gpu-mode/lec...
Stochastic Rounding, Minisymp on Probabilistic Rounding, Sep 1 bsky.app/profile/ogaw...
Survey, Proc of the IEEE, Jun 26 bsky.app/profile/ogaw...
36 pp, 201 ref
MX+, MICRO25
Yes just call it a multimodal utility access thoroughfare or maybe an enalekib
Is it illegal to simply remove car lanes and replace them with nothing?
November 25, 2025 at 2:57 AM
Yes just call it a multimodal utility access thoroughfare or maybe an enalekib
If Ford gov’t doesn’t reopen @onsciencecentre.bsky.social pronto, it will never reopen.
2028, laying off the rest of the Science Centre staff, “I’m as disappointed as anyone,” fibbed the Minister of Infrastructure, reflecting on her 2024 decision to shutter an institution beloved of generations.
2028, laying off the rest of the Science Centre staff, “I’m as disappointed as anyone,” fibbed the Minister of Infrastructure, reflecting on her 2024 decision to shutter an institution beloved of generations.
November 24, 2025 at 5:08 PM
If Ford gov’t doesn’t reopen @onsciencecentre.bsky.social pronto, it will never reopen.
2028, laying off the rest of the Science Centre staff, “I’m as disappointed as anyone,” fibbed the Minister of Infrastructure, reflecting on her 2024 decision to shutter an institution beloved of generations.
2028, laying off the rest of the Science Centre staff, “I’m as disappointed as anyone,” fibbed the Minister of Infrastructure, reflecting on her 2024 decision to shutter an institution beloved of generations.
Is NVLink Fusion a response to Ultra Accelerator Link? Would it have happened sans UAL? 🤪
November 20, 2025 at 4:47 AM
Is NVLink Fusion a response to Ultra Accelerator Link? Would it have happened sans UAL? 🤪
Reposted by Jan Gray
Microsoft Cobalt 200
- TSMC 3 nm
- two chiplets with custom HB interconnect
- 132 cores (Armv9.2), 66 per chiplet
- 3 MB private L2$
- 192 MB L3$
- 12 memory channels
- per core DVFS
www.hardwareluxx.de/index.php/ne...
- TSMC 3 nm
- two chiplets with custom HB interconnect
- 132 cores (Armv9.2), 66 per chiplet
- 3 MB private L2$
- 192 MB L3$
- 12 memory channels
- per core DVFS
www.hardwareluxx.de/index.php/ne...
November 19, 2025 at 8:29 AM
Microsoft Cobalt 200
- TSMC 3 nm
- two chiplets with custom HB interconnect
- 132 cores (Armv9.2), 66 per chiplet
- 3 MB private L2$
- 192 MB L3$
- 12 memory channels
- per core DVFS
www.hardwareluxx.de/index.php/ne...
- TSMC 3 nm
- two chiplets with custom HB interconnect
- 132 cores (Armv9.2), 66 per chiplet
- 3 MB private L2$
- 192 MB L3$
- 12 memory channels
- per core DVFS
www.hardwareluxx.de/index.php/ne...
Hooray for bitblt, a.k.a. rasterop in some backward provinces.
You younguns don’t get it — but if all you knew was 24x80 character displays (*) or worse, bitblt was a long tall drink of water.
Color is fine but 80% of the advance came with the bilevel display.
(*) vector displays notwithstanding
You younguns don’t get it — but if all you knew was 24x80 character displays (*) or worse, bitblt was a long tall drink of water.
Color is fine but 80% of the advance came with the bilevel display.
(*) vector displays notwithstanding
I declare that today, Nov. 19, 2025 is the 50th anniversary of BitBLT, a routine so fundamental to computer graphics that we don't even think about it having an origin. Dan Ingalls came up with a working implementation on the Xerox Alto as part of his work on Smalltalk.
November 19, 2025 at 8:26 AM
Hooray for bitblt, a.k.a. rasterop in some backward provinces.
You younguns don’t get it — but if all you knew was 24x80 character displays (*) or worse, bitblt was a long tall drink of water.
Color is fine but 80% of the advance came with the bilevel display.
(*) vector displays notwithstanding
You younguns don’t get it — but if all you knew was 24x80 character displays (*) or worse, bitblt was a long tall drink of water.
Color is fine but 80% of the advance came with the bilevel display.
(*) vector displays notwithstanding
One thing there is not much of at Supercomputing’25 is FPGAs for HPC accelerators. As far as I have seen so far there is an Achronix booth. There is one Alveo V80 (Versal HBM) card in the AMD booth, unadorned with any demos or collateral. No Altera.
November 19, 2025 at 4:18 AM
One thing there is not much of at Supercomputing’25 is FPGAs for HPC accelerators. As far as I have seen so far there is an Achronix booth. There is one Alveo V80 (Versal HBM) card in the AMD booth, unadorned with any demos or collateral. No Altera.
Reposted by Jan Gray
HPCwire Reveals Winners of the 22nd Annual Readers’ and Editors’ Choice Awards During SC25
#HPCwire #SC25
Link: ow.ly/fC4L50XtflZ
#HPCwire #SC25
Link: ow.ly/fC4L50XtflZ
HPCwire Reveals Winners of the 22nd Annual Readers’ and Editors’ Choice Awards During SC25 - HPCwire
ST. LOUIS, Nov. 17, 2025 — HPCwire, the leading publication for news and information for the high performance computing industry, announced the winners of the 2025 HPCwire Readers’ and Editors’ Choice...
ow.ly
November 18, 2025 at 12:18 AM
HPCwire Reveals Winners of the 22nd Annual Readers’ and Editors’ Choice Awards During SC25
#HPCwire #SC25
Link: ow.ly/fC4L50XtflZ
#HPCwire #SC25
Link: ow.ly/fC4L50XtflZ
November 18, 2025 at 4:42 AM
Reposted by Jan Gray
A wide range of talks so far this morning at the @riscv.org.web.brid.gy #SC25 workshop. Driving analogue systems from #RISCV, coupling with spiking neural networks, and profiling tooling all demonstrates the flexibility the technology gives us to do cool things
November 17, 2025 at 4:36 PM
A wide range of talks so far this morning at the @riscv.org.web.brid.gy #SC25 workshop. Driving analogue systems from #RISCV, coupling with spiking neural networks, and profiling tooling all demonstrates the flexibility the technology gives us to do cool things
November 16, 2025 at 8:08 PM
Reposted by Jan Gray
5400W (!) per GB200 NVL4 board exemplifies the hot Winter of Moore's Law.
The next five years' next 10X @ constant power must somehow shave 90% of today's energy: new algorithms & abstractions, codesign, integration, specialization, accelerators, processing near & in-memory, and new interconnects.
The next five years' next 10X @ constant power must somehow shave 90% of today's energy: new algorithms & abstractions, codesign, integration, specialization, accelerators, processing near & in-memory, and new interconnects.
November 20, 2024 at 2:48 PM
5400W (!) per GB200 NVL4 board exemplifies the hot Winter of Moore's Law.
The next five years' next 10X @ constant power must somehow shave 90% of today's energy: new algorithms & abstractions, codesign, integration, specialization, accelerators, processing near & in-memory, and new interconnects.
The next five years' next 10X @ constant power must somehow shave 90% of today's energy: new algorithms & abstractions, codesign, integration, specialization, accelerators, processing near & in-memory, and new interconnects.
Reposted by Jan Gray
i got suspended for yielding from my coroutine
November 15, 2025 at 6:53 PM
i got suspended for yielding from my coroutine
“Popular demand”
Popular demand is restoring and reopening the real thing, following its abrupt shabby and sordid shuttering and gutting.
Ford, Surma, and (apparently) the Board are manifestly deaf to popular demand.
~100,000 letters to MPPs to @saveosc.bsky.social were ignored.
Popular demand.
Popular demand is restoring and reopening the real thing, following its abrupt shabby and sordid shuttering and gutting.
Ford, Surma, and (apparently) the Board are manifestly deaf to popular demand.
~100,000 letters to MPPs to @saveosc.bsky.social were ignored.
Popular demand.
Back by popular demand, Emett’s Dream Machines have arrived for the holiday season at our CF Sherway Gardens Experience! 🤩 Bring the whole family and experience the one-of-a-kind creations today ☕️🛤️
📍CF Sherway Gardens
📍CF Sherway Gardens
November 14, 2025 at 9:12 PM
“Popular demand”
Popular demand is restoring and reopening the real thing, following its abrupt shabby and sordid shuttering and gutting.
Ford, Surma, and (apparently) the Board are manifestly deaf to popular demand.
~100,000 letters to MPPs to @saveosc.bsky.social were ignored.
Popular demand.
Popular demand is restoring and reopening the real thing, following its abrupt shabby and sordid shuttering and gutting.
Ford, Surma, and (apparently) the Board are manifestly deaf to popular demand.
~100,000 letters to MPPs to @saveosc.bsky.social were ignored.
Popular demand.
ISA extensions evolve. What’s a good programming model for coping with that, providing composition and longevity of old and new code on old and new processors?
github.com/riscv/compos...
github.com/riscv/compos...
November 14, 2025 at 3:08 AM
ISA extensions evolve. What’s a good programming model for coping with that, providing composition and longevity of old and new code on old and new processors?
github.com/riscv/compos...
github.com/riscv/compos...
Took styrofoam to the recycler. There, waiting for me, forsaken, was a 1979 2 volume Compact Ed. of the OED, "Reproduced Micrographically", with magnifying glass.
50 years ago these were an awesome, revered artifact on the teacher's desk.
I had to bring it home, at least to find it a better end.
50 years ago these were an awesome, revered artifact on the teacher's desk.
I had to bring it home, at least to find it a better end.
November 13, 2025 at 10:47 PM
Took styrofoam to the recycler. There, waiting for me, forsaken, was a 1979 2 volume Compact Ed. of the OED, "Reproduced Micrographically", with magnifying glass.
50 years ago these were an awesome, revered artifact on the teacher's desk.
I had to bring it home, at least to find it a better end.
50 years ago these were an awesome, revered artifact on the teacher's desk.
I had to bring it home, at least to find it a better end.
Reposted by Jan Gray
a “Space Force”, if you will
November 13, 2025 at 2:27 PM
a “Space Force”, if you will
Reposted by Jan Gray
=>
Euclyd Unveils CRAFTWERK SiP (system-in-package), Sep 9, 2025 euclyd.ai/euclyd-unvei...
16,384 128-way SIMD processors
8 PF @ FP16
32 PF @ FP4
Ultra Bandwidth Memory (UBM): 1TB, 8,000TB/s
CWS 32 – Exascale AI Factory
32x CRAFTWERK
1 EF @ FP4
7.68M tokens/s, ~125 kW
www.eetimes.com/startup-to-t...
Euclyd Unveils CRAFTWERK SiP (system-in-package), Sep 9, 2025 euclyd.ai/euclyd-unvei...
16,384 128-way SIMD processors
8 PF @ FP16
32 PF @ FP4
Ultra Bandwidth Memory (UBM): 1TB, 8,000TB/s
CWS 32 – Exascale AI Factory
32x CRAFTWERK
1 EF @ FP4
7.68M tokens/s, ~125 kW
www.eetimes.com/startup-to-t...
September 26, 2025 at 12:21 PM
=>
Euclyd Unveils CRAFTWERK SiP (system-in-package), Sep 9, 2025 euclyd.ai/euclyd-unvei...
16,384 128-way SIMD processors
8 PF @ FP16
32 PF @ FP4
Ultra Bandwidth Memory (UBM): 1TB, 8,000TB/s
CWS 32 – Exascale AI Factory
32x CRAFTWERK
1 EF @ FP4
7.68M tokens/s, ~125 kW
www.eetimes.com/startup-to-t...
Euclyd Unveils CRAFTWERK SiP (system-in-package), Sep 9, 2025 euclyd.ai/euclyd-unvei...
16,384 128-way SIMD processors
8 PF @ FP16
32 PF @ FP4
Ultra Bandwidth Memory (UBM): 1TB, 8,000TB/s
CWS 32 – Exascale AI Factory
32x CRAFTWERK
1 EF @ FP4
7.68M tokens/s, ~125 kW
www.eetimes.com/startup-to-t...
Are 2030 AI hyperscalars capital constrained, power constrained, DRAM constrained, flash constrained, compute constrained, software constrained, or :-) demand constrained?
November 11, 2025 at 2:30 PM
Are 2030 AI hyperscalars capital constrained, power constrained, DRAM constrained, flash constrained, compute constrained, software constrained, or :-) demand constrained?
Not sure what’s up here, whether flash, smarts, chiplets, packaging, or otherwise, but completely rethinking memory controllers and functions for compute in/near memory is still a compelling unexplored country.
“DRAM” not in press release.
What’s new with HBF? 128 TB DRAM doesn’t grow on trees.
“DRAM” not in press release.
What’s new with HBF? 128 TB DRAM doesn’t grow on trees.
=>
Majestic Labs Emerges from Stealth with $100M to Redefine AI Infrastructure by Tearing Down the “Memory Wall”, Nov 10, 2025 www.businesswire.com/news/home/20...
Led by Ex-Google & Meta Executives
Custom accelerator
Memory interface chip, up to 128 TB, Fast, Power-Efficient, High-Bandwidth Memory
Majestic Labs Emerges from Stealth with $100M to Redefine AI Infrastructure by Tearing Down the “Memory Wall”, Nov 10, 2025 www.businesswire.com/news/home/20...
Led by Ex-Google & Meta Executives
Custom accelerator
Memory interface chip, up to 128 TB, Fast, Power-Efficient, High-Bandwidth Memory
November 11, 2025 at 2:07 PM
Not sure what’s up here, whether flash, smarts, chiplets, packaging, or otherwise, but completely rethinking memory controllers and functions for compute in/near memory is still a compelling unexplored country.
“DRAM” not in press release.
What’s new with HBF? 128 TB DRAM doesn’t grow on trees.
“DRAM” not in press release.
What’s new with HBF? 128 TB DRAM doesn’t grow on trees.
Reposted by Jan Gray
Initial conditions:
m1=21.4 m2=17.1 m3=52.9 (solar masses)
v1x=2.253 v1y=5.14 v2x=3.9 v2y=-1.408 v3x=-4.494 v3y=-6.724 (km/s)
x1=15.0 y1=1.0 x2=34.0 y2=6.0 x3=35.0 y3=-28.0 (AU from center)
Music: What Did You Do (The Expanse) – Shorter
m1=21.4 m2=17.1 m3=52.9 (solar masses)
v1x=2.253 v1y=5.14 v2x=3.9 v2y=-1.408 v3x=-4.494 v3y=-6.724 (km/s)
x1=15.0 y1=1.0 x2=34.0 y2=6.0 x3=35.0 y3=-28.0 (AU from center)
Music: What Did You Do (The Expanse) – Shorter
November 10, 2025 at 7:20 AM
Initial conditions:
m1=21.4 m2=17.1 m3=52.9 (solar masses)
v1x=2.253 v1y=5.14 v2x=3.9 v2y=-1.408 v3x=-4.494 v3y=-6.724 (km/s)
x1=15.0 y1=1.0 x2=34.0 y2=6.0 x3=35.0 y3=-28.0 (AU from center)
Music: What Did You Do (The Expanse) – Shorter
m1=21.4 m2=17.1 m3=52.9 (solar masses)
v1x=2.253 v1y=5.14 v2x=3.9 v2y=-1.408 v3x=-4.494 v3y=-6.724 (km/s)
x1=15.0 y1=1.0 x2=34.0 y2=6.0 x3=35.0 y3=-28.0 (AU from center)
Music: What Did You Do (The Expanse) – Shorter
Reposted by Jan Gray
Nearly two decades after my book on Cheney, I took a shot at assessing his legacy. It’s more complicated than you might think. www.theatlantic.com/ideas/2025/1...
What I Learned About Dick Cheney
The former vice president was certain he knew better than the citizens he served what was good for them.
www.theatlantic.com
November 9, 2025 at 6:40 PM
Nearly two decades after my book on Cheney, I took a shot at assessing his legacy. It’s more complicated than you might think. www.theatlantic.com/ideas/2025/1...
Reposted by Jan Gray
For onegreatidea @thestar.com
IDEA: reimagine and reopen Ontario Science Centre in Flemingdon Park as the Future of Canada Centre
CC: @mark-carney.bsky.social @onsciencecentre.bsky.social @saveosc.bsky.social @canadianarchitect.bsky.social @normsworld.bsky.social @dontrailtalker.bsky.social
IDEA: reimagine and reopen Ontario Science Centre in Flemingdon Park as the Future of Canada Centre
CC: @mark-carney.bsky.social @onsciencecentre.bsky.social @saveosc.bsky.social @canadianarchitect.bsky.social @normsworld.bsky.social @dontrailtalker.bsky.social
November 9, 2025 at 4:17 AM
For onegreatidea @thestar.com
IDEA: reimagine and reopen Ontario Science Centre in Flemingdon Park as the Future of Canada Centre
CC: @mark-carney.bsky.social @onsciencecentre.bsky.social @saveosc.bsky.social @canadianarchitect.bsky.social @normsworld.bsky.social @dontrailtalker.bsky.social
IDEA: reimagine and reopen Ontario Science Centre in Flemingdon Park as the Future of Canada Centre
CC: @mark-carney.bsky.social @onsciencecentre.bsky.social @saveosc.bsky.social @canadianarchitect.bsky.social @normsworld.bsky.social @dontrailtalker.bsky.social