To use autocomplete as a metaphor for LLMs is representative but woefully incomplete. There’s something extraordinary to be said at how we’re representing every connection mathematically. What foot is to walk, what wind is to air is all in a dataset interpreted using math. Are we building for the everything algorithm?
0
0
0
41
Rob Imbeault retweeted
We beat DeepInfer head to head, more to come, but hope @SemiAnalysis_ notices @BackboardIO soon, even though we’re a Canadian lab without hundreds of millions of dollars and thousands of contributors. A win is a win.
0
3
6
542
Rob Imbeault retweeted
Backboard’s latest internal benchmarks (Nov 2026 research and blogs) claim solid gains over both engines on Qwen3.8/GLM models: up to 2.2× aggregate output throughput and far lower median TTFT vs vLLM, plus 4–26% throughput and large interactivity wins vs SGLang (sometimes 10×+ TTFT). They emphasize full BF16 precision on single or multi-GPU setups.
These are self-reported, not always equal-precision or identical hardware, and focus more on sustained capacity/latency than pure no-cache prefill. No public DeepSeek V4.1 Flash numbers on DGX Sparks exist yet, so direct ranking against that specific SGLang result isn’t possible.
These are self-reported, not always equal-precision or identical hardware, and focus more on sustained capacity/latency than pure no-cache prefill. No public DeepSeek V4.1 Flash numbers on DGX Sparks exist yet, so direct ranking against that specific SGLang result isn’t possible.
1
1
1
35
Rob Imbeault retweeted
More useful work from the same GPUs.
Using Qwen3.8-Flash-Next on eight H100 GPUs. Backboard managed up to 2.2× higher output throughput per GPU than the SGLang results on SemiAnalysis AgentX.
This was done at BF16 versus SGLang at FP8. Better throughput without sacrificing model precision.
Using Qwen3.8-Flash-Next on eight H100 GPUs. Backboard managed up to 2.2× higher output throughput per GPU than the SGLang results on SemiAnalysis AgentX.
This was done at BF16 versus SGLang at FP8. Better throughput without sacrificing model precision.
0
2
3
133
For an inference provider running 8×H100s, these results could translate into ~$236,000 in ADDITIONAL annual revenue per server.
Same GPUs. No additional hardware. Just more sellable compute.
Same GPUs. No additional hardware. Just more sellable compute.
Agentic AI needs more than fast tokens. It needs responsive tokens.
On SemiAnalysis AgentX, Backboard delivered up to 12.8× higher p90 interactivity than SGLang on Qwen3.8-Flash-Next across eight H100 GPUs. That’s quicker starts, smoother answers, and a more consistent ran BF16. The SGLang reference ran FP8.
1
0
5
223
Rob Imbeault retweeted
We're so bad at comms, but we're getting better. dev.to/jon_at_backboa…
0
2
5
81
Rob Imbeault retweeted
More concurrent requests shouldn't mean giving up throughput.
On Qwen3.8-27B, Backboard maintained up to a 1.26× output-throughput advantage over the fastest results from vLLM 0.33 and SGLang, with up to 8K inputs and concurrency up to 64 users.
All on a single RTX PRO 6000 Blackwell GPU.
On Qwen3.8-27B, Backboard maintained up to a 1.26× output-throughput advantage over the fastest results from vLLM 0.33 and SGLang, with up to 8K inputs and concurrency up to 64 users.
All on a single RTX PRO 6000 Blackwell GPU.
0
1
3
88
I trained at Gold's a few years ago and Arnold would show up every morning. It took everything not to chase after him asking for a photo every day for a month. I didn't disturb him.
One day while one the last few reps of a deep squat he came over and cheered me on, patted me on the back and said "good finish." I could barely stand.
He did this with so many people whenever I was there. I was already a fan but this made me a super fan.
One day while one the last few reps of a deep squat he came over and cheered me on, patted me on the back and said "good finish." I could barely stand.
He did this with so many people whenever I was there. I was already a fan but this made me a super fan.
I have to say I’m so proud of you. The lessons you took away for our workout are exactly the ones I hoped you’d take away. All out every set. Focus. Full range of motion. Keep it simple, ignore all the social media noise and just train. It’s what I tell people every day in the Pump Club.
usatoday.com/story/life/hea…
0
0
2
77
Rob Imbeault retweeted
Your inference engine should be consistently fast, not just in a few conditions.
Backboard won 102 of 104 measured InferenceX metrics against SGLang on Qwen3.8-27B using a lone RTX PRO 6000 Blackwell GPU. That includes throughput, time to first token, and generation speed.
This nds up to scrutiny.
Backboard won 102 of 104 measured InferenceX metrics against SGLang on Qwen3.8-27B using a lone RTX PRO 6000 Blackwell GPU. That includes throughput, time to first token, and generation speed.
This nds up to scrutiny.
0
2
3
113
Rob Imbeault retweeted
You want your AI inference to win virtually every time. Ours does.
Backboard won every measured InferenceX metric against vLLM 0.33 on Qwen3.8-27B, across test configurations covering up to 8K inputs and 64-user concurrency.
That’s with one RTX PRO 6000 Blackwell GPU, BF16, and tion.
Backboard won every measured InferenceX metric against vLLM 0.33 on Qwen3.8-27B, across test configurations covering up to 8K inputs and 64-user concurrency.
That’s with one RTX PRO 6000 Blackwell GPU, BF16, and tion.
0
2
3
99
Be so good they can't ignore you.
At ~1.4s P90 time to first token:
Backboard: $6.62/GPU-hour
SGLang: ~$2.75/GPU-hour
~2.4× modeled token revenue. Same 8× H100, same Qwen3.8 Flash Next model.
Makin H100's Great Again. (<-- can we still say that?)
@SemiAnalysis_ @dylan522p
@RobImbeault @jon_curio
1
1
8
229
Fresh out of the oven. We were told to use the AgentX by @SemiAnalysis_ benchmark so we ran it against the current best score, SGLang.
We beat it by every measure and we didn't reconfigure per concurrency. The @BackboardIO team cooks.
PS. Backboard is on full BF16 precision ot full model).
We beat it by every measure and we didn't reconfigure per concurrency. The @BackboardIO team cooks.
PS. Backboard is on full BF16 precision ot full model).
0
3
5
115
This is huge for the next generation of Canadian builders!
Budget 2026 is a budget for builders. 🍁
Budget 2026 announced the Canada Growth Incentive: up to $15M in tax-free capital gains per person, per business, on shares of eligible 🇨🇦 companies bought from today.
An Ontario founder selling for $15M now keeps 96%, up from 71%. 🏗️
0
0
1
99
Rob Imbeault retweeted
More inference performance doesn't have to mean more GPUs.
On a single NVIDIA RTX PRO 6000 Blackwell GPU driving Qwen3.6, Backboard delivered 1.57×–1.95× the aggregate output throughput of vLLM across concurrency levels from C8 to C256.
More output from the same hardware.
On a single NVIDIA RTX PRO 6000 Blackwell GPU driving Qwen3.6, Backboard delivered 1.57×–1.95× the aggregate output throughput of vLLM across concurrency levels from C8 to C256.
More output from the same hardware.
0
1
4
153
This is fascinating and so good to see another US open weight model-model(?)
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron re: ricursive.com/ripple
0
0
2
194
Compared against vLLM 0.33 (released yesterday): Backboard serves up to 95% more output while serving up to 3.6X FASTER!
The heavier the load, the better we performed.
More to come.
The heavier the load, the better we performed.
More to come.
1
3
5
151